forked from Blendman974/faceclaw
Multi-mic support and replay mode for Transcribe app
This commit is contained in:
1 parent
c162c8e633
commit
9372b2742c
45 files changed
+3926
-288
No files matched your search
@@ -69,3 +69,41 @@ actual fun fileLength(path: String): Long = java.io.File(path).let { if (it.exis
|
||||
actual fun deleteFile(path: String): Boolean = java.io.File(path).delete()
|
||||
|
||||
actual fun renameFile(from: String, to: String): Boolean = java.io.File(from).renameTo(java.io.File(to))
|
||||
|
||||
actual fun openFileOutput(path: String): SeekableOutput = object : SeekableOutput {
|
||||
private val file = java.io.RandomAccessFile(path, "rw").also { it.setLength(0) }
|
||||
override var size = 0L
|
||||
private set
|
||||
|
||||
override fun write(bytes: ByteArray, offset: Int, length: Int) {
|
||||
file.write(bytes, offset, length)
|
||||
size += length
|
||||
}
|
||||
|
||||
override fun writeAt(position: Long, bytes: ByteArray) {
|
||||
require(position >= 0 && position + bytes.size <= size)
|
||||
file.seek(position)
|
||||
file.write(bytes)
|
||||
file.seek(size)
|
||||
}
|
||||
|
||||
override fun close() = file.close()
|
||||
}
|
||||
|
||||
actual fun openFileInput(path: String): SeekableInput = object : SeekableInput {
|
||||
private val file = java.io.RandomAccessFile(path, "r")
|
||||
override val length = file.length()
|
||||
|
||||
override fun readAt(position: Long, buffer: ByteArray, offset: Int, length: Int): Int {
|
||||
file.seek(position)
|
||||
var total = 0
|
||||
while (total < length) {
|
||||
val read = file.read(buffer, offset + total, length - total)
|
||||
if (read <= 0) break
|
||||
total += read
|
||||
}
|
||||
return total
|
||||
}
|
||||
|
||||
override fun close() = file.close()
|
||||
}
|
||||
@@ -20,3 +20,33 @@ expect fun fileLength(path: String): Long
|
||||
expect fun deleteFile(path: String): Boolean
|
||||
|
||||
expect fun renameFile(from: String, to: String): Boolean
|
||||
|
||||
/** A file being written front to back whose earlier bytes can still be patched (container headers). */
|
||||
interface SeekableOutput {
|
||||
/** Bytes appended so far. */
|
||||
val size: Long
|
||||
|
||||
/** Appends [length] bytes of [bytes] from [offset]. */
|
||||
fun write(bytes: ByteArray, offset: Int, length: Int)
|
||||
|
||||
/** Overwrites bytes already written, at [position]; later appends still go to the end. */
|
||||
fun writeAt(position: Long, bytes: ByteArray)
|
||||
|
||||
fun close()
|
||||
}
|
||||
|
||||
/** A file read at arbitrary positions without loading it whole. */
|
||||
interface SeekableInput {
|
||||
val length: Long
|
||||
|
||||
/** Reads up to [length] bytes at [position] into [buffer]; returns the count read (short only at the end). */
|
||||
fun readAt(position: Long, buffer: ByteArray, offset: Int, length: Int): Int
|
||||
|
||||
fun close()
|
||||
}
|
||||
|
||||
/** Creates (or truncates) [path] for writing; parent directories must exist. Throws when it cannot. */
|
||||
expect fun openFileOutput(path: String): SeekableOutput
|
||||
|
||||
/** Opens [path] for reading. Throws when it cannot. */
|
||||
expect fun openFileInput(path: String): SeekableInput
|
||||
@@ -0,0 +1,764 @@
|
||||
package com.faceclaw.app
|
||||
|
||||
import kotlin.jvm.JvmField
|
||||
|
||||
/**
|
||||
* A small Matroska (EBML) muxer and demuxer for audio captures (.mka): PCM16 audio tracks that
|
||||
* each have their own sample rate and channel count, UTF-8 text tracks for timed metadata, and
|
||||
* file attachments. Every block carries its own timestamp, so tracks on independent clocks and
|
||||
* gaps in a track survive in the file. See notes/mic-recording-format.md for how captures use it.
|
||||
*
|
||||
* The writer produces files standard players open (VLC, mpv, ffmpeg): known-size clusters of
|
||||
* about a second each, cues and the segment size written when the recording finishes. A
|
||||
* recording cut short by a crash is still a valid live-style stream of everything up to its last
|
||||
* complete cluster, and the reader takes it as such.
|
||||
*/
|
||||
internal object MatroskaIds {
|
||||
const val EBML = 0x1A45DFA3
|
||||
const val EBML_VERSION = 0x4286
|
||||
const val EBML_READ_VERSION = 0x42F7
|
||||
const val EBML_MAX_ID_LENGTH = 0x42F2
|
||||
const val EBML_MAX_SIZE_LENGTH = 0x42F3
|
||||
const val DOC_TYPE = 0x4282
|
||||
const val DOC_TYPE_VERSION = 0x4287
|
||||
const val DOC_TYPE_READ_VERSION = 0x4285
|
||||
const val VOID = 0xEC
|
||||
const val SEGMENT = 0x18538067
|
||||
const val SEEK_HEAD = 0x114D9B74
|
||||
const val SEEK = 0x4DBB
|
||||
const val SEEK_ID = 0x53AB
|
||||
const val SEEK_POSITION = 0x53AC
|
||||
const val INFO = 0x1549A966
|
||||
const val TIMESTAMP_SCALE = 0x2AD7B1
|
||||
const val DURATION = 0x4489
|
||||
const val DATE_UTC = 0x4461
|
||||
const val TITLE = 0x7BA9
|
||||
const val MUXING_APP = 0x4D80
|
||||
const val WRITING_APP = 0x5741
|
||||
const val TRACKS = 0x1654AE6B
|
||||
const val TRACK_ENTRY = 0xAE
|
||||
const val TRACK_NUMBER = 0xD7
|
||||
const val TRACK_UID = 0x73C5
|
||||
const val TRACK_TYPE = 0x83
|
||||
const val FLAG_DEFAULT = 0x88
|
||||
const val FLAG_LACING = 0x9C
|
||||
const val NAME = 0x536E
|
||||
const val LANGUAGE = 0x22B59C
|
||||
const val CODEC_ID = 0x86
|
||||
const val AUDIO = 0xE1
|
||||
const val SAMPLING_FREQUENCY = 0xB5
|
||||
const val CHANNELS = 0x9F
|
||||
const val BIT_DEPTH = 0x6264
|
||||
const val ATTACHMENTS = 0x1941A469
|
||||
const val ATTACHED_FILE = 0x61A7
|
||||
const val FILE_NAME = 0x466E
|
||||
const val FILE_MIME_TYPE = 0x4660
|
||||
const val FILE_DATA = 0x465C
|
||||
const val FILE_UID = 0x46AE
|
||||
const val CLUSTER = 0x1F43B675
|
||||
const val TIMESTAMP = 0xE7
|
||||
const val SIMPLE_BLOCK = 0xA3
|
||||
const val BLOCK_GROUP = 0xA0
|
||||
const val BLOCK = 0xA1
|
||||
const val BLOCK_DURATION = 0x9B
|
||||
const val CUES = 0x1C53BB6B
|
||||
const val CUE_POINT = 0xBB
|
||||
const val CUE_TIME = 0xB3
|
||||
const val CUE_TRACK_POSITIONS = 0xB7
|
||||
const val CUE_TRACK = 0xF7
|
||||
const val CUE_CLUSTER_POSITION = 0xF1
|
||||
}
|
||||
|
||||
/** Builds EBML elements into a byte buffer. */
|
||||
internal class EbmlWriter(capacity: Int = 256) {
|
||||
private val sink = ByteSink(capacity)
|
||||
|
||||
fun size(): Int = sink.size()
|
||||
|
||||
fun bytes(): ByteArray = sink.toByteArray()
|
||||
|
||||
/** An element ID, written as given: its length marker is part of the value. */
|
||||
fun id(id: Int) {
|
||||
if (id > 0xFFFFFF) sink.write(id ushr 24)
|
||||
if (id > 0xFFFF) sink.write(id ushr 16)
|
||||
if (id > 0xFF) sink.write(id ushr 8)
|
||||
sink.write(id)
|
||||
}
|
||||
|
||||
/** A data size in [width] bytes, or in the fewest that hold it when [width] is 0. */
|
||||
fun dataSize(size: Long, width: Int = 0) {
|
||||
var bytes = width
|
||||
if (bytes == 0) {
|
||||
bytes = 1
|
||||
// The all-ones value of each width means "unknown size".
|
||||
while (bytes < 8 && size >= (1L shl (7 * bytes)) - 1) bytes++
|
||||
}
|
||||
val marked = size or (1L shl (7 * bytes))
|
||||
for (index in bytes - 1 downTo 0) sink.write((marked ushr (8 * index)).toInt())
|
||||
}
|
||||
|
||||
fun uint(id: Int, value: Long) {
|
||||
var bytes = 1
|
||||
while (bytes < 8 && (value ushr (8 * bytes)) != 0L) bytes++
|
||||
id(id)
|
||||
dataSize(bytes.toLong())
|
||||
for (index in bytes - 1 downTo 0) sink.write((value ushr (8 * index)).toInt())
|
||||
}
|
||||
|
||||
/** A fixed eight-byte integer (dates). */
|
||||
fun int64(id: Int, value: Long) {
|
||||
id(id)
|
||||
dataSize(8)
|
||||
for (index in 7 downTo 0) sink.write((value ushr (8 * index)).toInt())
|
||||
}
|
||||
|
||||
fun float64(id: Int, value: Double) = int64(id, value.toRawBits())
|
||||
|
||||
fun string(id: Int, text: String) {
|
||||
val bytes = text.encodeToByteArray()
|
||||
binary(id, bytes, 0, bytes.size)
|
||||
}
|
||||
|
||||
fun binary(id: Int, bytes: ByteArray, offset: Int, length: Int) {
|
||||
id(id)
|
||||
dataSize(length.toLong())
|
||||
sink.write(bytes, offset, length)
|
||||
}
|
||||
|
||||
fun raw(bytes: ByteArray) = sink.write(bytes)
|
||||
|
||||
fun raw(value: Int) = sink.write(value)
|
||||
|
||||
fun master(id: Int, build: EbmlWriter.() -> Unit) {
|
||||
val child = EbmlWriter()
|
||||
child.build()
|
||||
id(id)
|
||||
dataSize(child.size().toLong())
|
||||
sink.write(child.bytes())
|
||||
}
|
||||
|
||||
companion object {
|
||||
/** [element] followed by a Void element filling the rest of [reserved] bytes, or null when it doesn't fit. */
|
||||
fun padded(element: ByteArray, reserved: Int): ByteArray? {
|
||||
val filler = reserved - element.size
|
||||
// A Void element is at least two bytes.
|
||||
if (filler < 0 || filler == 1) return null
|
||||
val out = element.copyOf(reserved)
|
||||
if (filler == 0) return out
|
||||
val at = element.size
|
||||
out[at] = MatroskaIds.VOID.toByte()
|
||||
if (filler < 9) {
|
||||
out[at + 1] = (0x80 or (filler - 2)).toByte()
|
||||
} else {
|
||||
val size = (filler - 9).toLong()
|
||||
out[at + 1] = 0x01
|
||||
for (index in 0 until 7) out[at + 2 + index] = (size ushr (8 * (6 - index))).toByte()
|
||||
}
|
||||
return out
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
class MkaAttachment(@JvmField val name: String, @JvmField val mimeType: String, @JvmField val data: ByteArray)
|
||||
|
||||
/**
|
||||
* Writes one Matroska audio file. Tracks can be added at any time before the first block that
|
||||
* uses them: the track list sits in a reserved region that is rewritten in place. Timestamps are
|
||||
* milliseconds from the start of the file. Not thread-safe.
|
||||
*/
|
||||
class MkaWriter(
|
||||
private val output: SeekableOutput,
|
||||
title: String?,
|
||||
/** When the recording began (Unix epoch ms), for the file's date. */
|
||||
dateUtcMs: Long,
|
||||
attachments: List<MkaAttachment>,
|
||||
) {
|
||||
companion object {
|
||||
const val CODEC_PCM = "A_PCM/INT/LIT"
|
||||
const val CODEC_TEXT = "S_TEXT/UTF8"
|
||||
const val TRACK_AUDIO = 2
|
||||
const val TRACK_SUBTITLE = 0x11
|
||||
const val APP_NAME = "Faceclaw"
|
||||
private const val SEEK_RESERVED = 128
|
||||
private const val TRACKS_RESERVED = 4096
|
||||
private const val CLUSTER_SPAN_MS = 1000L
|
||||
private const val CLUSTER_MAX_BYTES = 1 shl 20
|
||||
// Block timestamps are signed 16-bit offsets from their cluster's.
|
||||
private const val CLUSTER_MAX_BACKSTEP_MS = 30_000L
|
||||
private const val DURATION_PATCH_CLUSTERS = 5
|
||||
// Matroska dates count nanoseconds from 2001-01-01T00:00:00 UTC.
|
||||
private const val DATE_EPOCH_MS = 978_307_200_000L
|
||||
}
|
||||
|
||||
private class Track(
|
||||
val number: Int,
|
||||
val type: Int,
|
||||
val name: String,
|
||||
val sampleRate: Int,
|
||||
val channels: Int,
|
||||
val preferred: Boolean,
|
||||
)
|
||||
|
||||
private class Cue(val clusterPosition: Long, val timeMs: Long)
|
||||
|
||||
private val tracks = ArrayList<Track>()
|
||||
private val cues = ArrayList<Cue>()
|
||||
private val segmentSizePosition: Long
|
||||
private val segmentStart: Long
|
||||
private val seekPosition: Long
|
||||
private val infoPosition: Long
|
||||
private val durationPosition: Long
|
||||
private val tracksPosition: Long
|
||||
private val attachmentsPosition: Long
|
||||
private var cluster: EbmlWriter? = null
|
||||
private var clusterBaseMs = 0L
|
||||
private var clustersSincePatch = 0
|
||||
private var finished = false
|
||||
private val hasAttachments = attachments.isNotEmpty()
|
||||
|
||||
/** The end of the latest audio or text written, in ms from the start of the file. */
|
||||
var durationMs = 0L
|
||||
private set
|
||||
|
||||
init {
|
||||
val head = EbmlWriter()
|
||||
head.master(MatroskaIds.EBML) {
|
||||
uint(MatroskaIds.EBML_VERSION, 1)
|
||||
uint(MatroskaIds.EBML_READ_VERSION, 1)
|
||||
uint(MatroskaIds.EBML_MAX_ID_LENGTH, 4)
|
||||
uint(MatroskaIds.EBML_MAX_SIZE_LENGTH, 8)
|
||||
string(MatroskaIds.DOC_TYPE, "matroska")
|
||||
uint(MatroskaIds.DOC_TYPE_VERSION, 4)
|
||||
uint(MatroskaIds.DOC_TYPE_READ_VERSION, 2)
|
||||
}
|
||||
head.id(MatroskaIds.SEGMENT)
|
||||
segmentSizePosition = head.size().toLong()
|
||||
// Unknown size (a live stream) until finish() knows it.
|
||||
head.raw(0x01)
|
||||
repeat(7) { head.raw(0xFF) }
|
||||
segmentStart = head.size().toLong()
|
||||
append(head.bytes())
|
||||
|
||||
seekPosition = output.size
|
||||
infoPosition = seekPosition + SEEK_RESERVED
|
||||
val info = EbmlWriter()
|
||||
info.master(MatroskaIds.INFO) {
|
||||
uint(MatroskaIds.TIMESTAMP_SCALE, 1_000_000)
|
||||
string(MatroskaIds.MUXING_APP, APP_NAME)
|
||||
string(MatroskaIds.WRITING_APP, APP_NAME)
|
||||
int64(MatroskaIds.DATE_UTC, (dateUtcMs - DATE_EPOCH_MS) * 1_000_000)
|
||||
if (!title.isNullOrEmpty()) string(MatroskaIds.TITLE, title)
|
||||
// Last, so its eight bytes are the element's final ones.
|
||||
float64(MatroskaIds.DURATION, 0.0)
|
||||
}
|
||||
durationPosition = infoPosition + info.size() - 8
|
||||
tracksPosition = infoPosition + info.size()
|
||||
attachmentsPosition = tracksPosition + TRACKS_RESERVED
|
||||
|
||||
append(seekHead(-1))
|
||||
append(info.bytes())
|
||||
append(EbmlWriter.padded(ByteArray(0), TRACKS_RESERVED)!!)
|
||||
if (attachments.isNotEmpty()) {
|
||||
val files = EbmlWriter()
|
||||
files.master(MatroskaIds.ATTACHMENTS) {
|
||||
attachments.forEachIndexed { index, file ->
|
||||
master(MatroskaIds.ATTACHED_FILE) {
|
||||
string(MatroskaIds.FILE_NAME, file.name)
|
||||
string(MatroskaIds.FILE_MIME_TYPE, file.mimeType)
|
||||
binary(MatroskaIds.FILE_DATA, file.data, 0, file.data.size)
|
||||
uint(MatroskaIds.FILE_UID, index + 1L)
|
||||
}
|
||||
}
|
||||
}
|
||||
append(files.bytes())
|
||||
}
|
||||
}
|
||||
|
||||
private fun append(bytes: ByteArray) = output.write(bytes, 0, bytes.size)
|
||||
|
||||
private fun seekHead(cuesPosition: Long): ByteArray {
|
||||
val head = EbmlWriter()
|
||||
head.master(MatroskaIds.SEEK_HEAD) {
|
||||
fun entry(id: Int, position: Long) = master(MatroskaIds.SEEK) {
|
||||
val idBytes = EbmlWriter()
|
||||
idBytes.id(id)
|
||||
binary(MatroskaIds.SEEK_ID, idBytes.bytes(), 0, idBytes.size())
|
||||
uint(MatroskaIds.SEEK_POSITION, position - segmentStart)
|
||||
}
|
||||
entry(MatroskaIds.INFO, infoPosition)
|
||||
entry(MatroskaIds.TRACKS, tracksPosition)
|
||||
if (hasAttachments) entry(MatroskaIds.ATTACHMENTS, attachmentsPosition)
|
||||
if (cuesPosition >= 0) entry(MatroskaIds.CUES, cuesPosition)
|
||||
}
|
||||
return EbmlWriter.padded(head.bytes(), SEEK_RESERVED) ?: error("seek head exceeds its reserved space")
|
||||
}
|
||||
|
||||
/**
|
||||
* Adds a PCM16 little-endian audio track and returns its number, or -1 when the track list
|
||||
* is full. A [preferred] track is listed ahead of the others, and the first listed track is
|
||||
* the one players pick by default.
|
||||
*/
|
||||
fun addAudioTrack(name: String, sampleRate: Int, channels: Int, preferred: Boolean): Int =
|
||||
addTrack(Track(tracks.size + 1, TRACK_AUDIO, name, sampleRate, channels, preferred))
|
||||
|
||||
/** Adds a UTF-8 text track (one string per block) and returns its number, or -1 when the track list is full. */
|
||||
fun addTextTrack(name: String): Int = addTrack(Track(tracks.size + 1, TRACK_SUBTITLE, name, 0, 0, false))
|
||||
|
||||
private fun addTrack(track: Track): Int {
|
||||
check(!finished) { "recording already finished" }
|
||||
// A block's track number is written in one byte.
|
||||
if (track.number > 126) return -1
|
||||
tracks.add(track)
|
||||
val region = EbmlWriter.padded(tracksElement(), TRACKS_RESERVED)
|
||||
if (region == null) {
|
||||
tracks.remove(track)
|
||||
return -1
|
||||
}
|
||||
output.writeAt(tracksPosition, region)
|
||||
return track.number
|
||||
}
|
||||
|
||||
private fun listedTracks(): List<Track> =
|
||||
tracks.filter { it.type == TRACK_AUDIO && it.preferred } +
|
||||
tracks.filter { it.type == TRACK_AUDIO && !it.preferred } +
|
||||
tracks.filter { it.type != TRACK_AUDIO }
|
||||
|
||||
private fun tracksElement(): ByteArray {
|
||||
val listed = listedTracks()
|
||||
val defaultTrack = listed.firstOrNull { it.type == TRACK_AUDIO }
|
||||
val out = EbmlWriter(1024)
|
||||
out.master(MatroskaIds.TRACKS) {
|
||||
for (track in listed) {
|
||||
master(MatroskaIds.TRACK_ENTRY) {
|
||||
uint(MatroskaIds.TRACK_NUMBER, track.number.toLong())
|
||||
uint(MatroskaIds.TRACK_UID, track.number.toLong())
|
||||
uint(MatroskaIds.TRACK_TYPE, track.type.toLong())
|
||||
uint(MatroskaIds.FLAG_DEFAULT, if (track === defaultTrack) 1 else 0)
|
||||
uint(MatroskaIds.FLAG_LACING, 0)
|
||||
string(MatroskaIds.NAME, track.name)
|
||||
string(MatroskaIds.LANGUAGE, "und")
|
||||
if (track.type == TRACK_AUDIO) {
|
||||
string(MatroskaIds.CODEC_ID, CODEC_PCM)
|
||||
master(MatroskaIds.AUDIO) {
|
||||
float64(MatroskaIds.SAMPLING_FREQUENCY, track.sampleRate.toDouble())
|
||||
uint(MatroskaIds.CHANNELS, track.channels.toLong())
|
||||
uint(MatroskaIds.BIT_DEPTH, 16)
|
||||
}
|
||||
} else {
|
||||
string(MatroskaIds.CODEC_ID, CODEC_TEXT)
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
return out.bytes()
|
||||
}
|
||||
|
||||
private fun track(number: Int): Track = tracks.getOrNull(number - 1) ?: error("no track $number")
|
||||
|
||||
/** Appends interleaved PCM16 little-endian samples to an audio track, starting at [timeMs]. */
|
||||
fun writeAudio(trackNumber: Int, timeMs: Long, pcm16le: ByteArray, offset: Int, length: Int) {
|
||||
val track = track(trackNumber)
|
||||
require(track.type == TRACK_AUDIO) { "track $trackNumber is not audio" }
|
||||
if (length <= 0) return
|
||||
val block = openCluster(timeMs)
|
||||
block.id(MatroskaIds.SIMPLE_BLOCK)
|
||||
block.dataSize(4L + length)
|
||||
blockHeader(block, trackNumber, timeMs, 0x80)
|
||||
block.raw(if (offset == 0 && length == pcm16le.size) pcm16le else pcm16le.copyOfRange(offset, offset + length))
|
||||
val frames = length / (2 * track.channels)
|
||||
noteEnd(timeMs + frames * 1000L / track.sampleRate)
|
||||
}
|
||||
|
||||
/** Appends one string to a text track, shown from [timeMs] for [durationMs]. */
|
||||
fun writeText(trackNumber: Int, timeMs: Long, durationMs: Long, text: String) {
|
||||
require(track(trackNumber).type == TRACK_SUBTITLE) { "track $trackNumber is not text" }
|
||||
val bytes = text.encodeToByteArray()
|
||||
val block = openCluster(timeMs)
|
||||
block.master(MatroskaIds.BLOCK_GROUP) {
|
||||
id(MatroskaIds.BLOCK)
|
||||
dataSize(4L + bytes.size)
|
||||
blockHeader(this, trackNumber, timeMs, 0)
|
||||
raw(bytes)
|
||||
uint(MatroskaIds.BLOCK_DURATION, maxOf(1L, durationMs))
|
||||
}
|
||||
noteEnd(timeMs + maxOf(1L, durationMs))
|
||||
}
|
||||
|
||||
private fun blockHeader(block: EbmlWriter, trackNumber: Int, timeMs: Long, flags: Int) {
|
||||
val relative = (maxOf(0L, timeMs) - clusterBaseMs).toInt()
|
||||
block.raw(0x80 or trackNumber)
|
||||
block.raw(relative shr 8)
|
||||
block.raw(relative)
|
||||
block.raw(flags)
|
||||
}
|
||||
|
||||
private fun noteEnd(endMs: Long) {
|
||||
if (endMs > durationMs) durationMs = endMs
|
||||
}
|
||||
|
||||
/** The cluster a block at [timeMs] goes in: the open one, or a new one once that is full or too far away in time. */
|
||||
private fun openCluster(timeMs: Long): EbmlWriter {
|
||||
check(!finished) { "recording already finished" }
|
||||
val time = maxOf(0L, timeMs)
|
||||
val open = cluster
|
||||
if (open != null) {
|
||||
val relative = time - clusterBaseMs
|
||||
if (relative < CLUSTER_SPAN_MS && relative > -CLUSTER_MAX_BACKSTEP_MS && open.size() < CLUSTER_MAX_BYTES) return open
|
||||
flushCluster()
|
||||
}
|
||||
clusterBaseMs = time
|
||||
return EbmlWriter(64 * 1024).also { cluster = it }
|
||||
}
|
||||
|
||||
private fun flushCluster() {
|
||||
val body = cluster ?: return
|
||||
cluster = null
|
||||
if (body.size() == 0) return
|
||||
val timestamp = EbmlWriter(16)
|
||||
timestamp.uint(MatroskaIds.TIMESTAMP, clusterBaseMs)
|
||||
val header = EbmlWriter(32)
|
||||
header.id(MatroskaIds.CLUSTER)
|
||||
header.dataSize(timestamp.size().toLong() + body.size())
|
||||
header.raw(timestamp.bytes())
|
||||
cues.add(Cue(output.size - segmentStart, clusterBaseMs))
|
||||
append(header.bytes())
|
||||
append(body.bytes())
|
||||
if (++clustersSincePatch >= DURATION_PATCH_CLUSTERS) patchDuration()
|
||||
}
|
||||
|
||||
private fun patchDuration() {
|
||||
clustersSincePatch = 0
|
||||
val value = durationMs.toDouble().toRawBits()
|
||||
output.writeAt(durationPosition, ByteArray(8) { (value ushr (8 * (7 - it))).toByte() })
|
||||
}
|
||||
|
||||
/** Writes what is buffered, then the cues, duration and segment size, and closes the file. */
|
||||
fun finish() {
|
||||
if (finished) return
|
||||
try {
|
||||
flushCluster()
|
||||
finished = true
|
||||
val cuesPosition = output.size
|
||||
val cueTrack = listedTracks().firstOrNull()?.number ?: 1
|
||||
if (cues.isNotEmpty()) {
|
||||
val index = EbmlWriter(cues.size * 16 + 16)
|
||||
index.master(MatroskaIds.CUES) {
|
||||
for (cue in cues) {
|
||||
master(MatroskaIds.CUE_POINT) {
|
||||
uint(MatroskaIds.CUE_TIME, cue.timeMs)
|
||||
master(MatroskaIds.CUE_TRACK_POSITIONS) {
|
||||
uint(MatroskaIds.CUE_TRACK, cueTrack.toLong())
|
||||
uint(MatroskaIds.CUE_CLUSTER_POSITION, cue.clusterPosition)
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
append(index.bytes())
|
||||
output.writeAt(seekPosition, seekHead(cuesPosition))
|
||||
}
|
||||
patchDuration()
|
||||
val size = EbmlWriter(8)
|
||||
size.dataSize(output.size - segmentStart, 8)
|
||||
output.writeAt(segmentSizePosition, size.bytes())
|
||||
} finally {
|
||||
finished = true
|
||||
output.close()
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
class MkaTrackInfo(
|
||||
@JvmField val number: Int,
|
||||
/** [MkaWriter.TRACK_AUDIO], [MkaWriter.TRACK_SUBTITLE] or another Matroska track type. */
|
||||
@JvmField val type: Int,
|
||||
@JvmField val codecId: String,
|
||||
@JvmField val name: String,
|
||||
@JvmField val sampleRate: Int,
|
||||
@JvmField val channels: Int,
|
||||
@JvmField val bitDepth: Int,
|
||||
) {
|
||||
val isPcm16: Boolean get() = type == MkaWriter.TRACK_AUDIO && codecId == MkaWriter.CODEC_PCM && bitDepth == 16
|
||||
}
|
||||
|
||||
class MkaBlock(
|
||||
@JvmField val track: Int,
|
||||
@JvmField val timeMs: Long,
|
||||
@JvmField val data: ByteArray,
|
||||
/** From the block's group, or -1 when the file doesn't say (audio blocks). */
|
||||
@JvmField val durationMs: Long,
|
||||
)
|
||||
|
||||
/**
|
||||
* Reads the blocks of a Matroska file in file order, a cluster at a time. Handles what
|
||||
* [MkaWriter] and ffmpeg write for PCM audio: known-size clusters, unlaced blocks. Throws
|
||||
* IllegalArgumentException for anything that is not such a file.
|
||||
*/
|
||||
class MkaReader(private val input: SeekableInput) {
|
||||
private class ClusterRef(val position: Long, val size: Int, val rawTime: Long)
|
||||
|
||||
private val clusters = ArrayList<ClusterRef>()
|
||||
private val trackList = ArrayList<MkaTrackInfo>()
|
||||
private val attachmentList = ArrayList<MkaAttachment>()
|
||||
private var timestampScale = 1_000_000L
|
||||
private var infoDurationMs = 0L
|
||||
private var nextCluster = 0
|
||||
private var blocks: List<MkaBlock> = emptyList()
|
||||
private var blockIndex = 0
|
||||
private var skipBeforeMs = 0L
|
||||
|
||||
val tracks: List<MkaTrackInfo> get() = trackList
|
||||
val attachments: List<MkaAttachment> get() = attachmentList
|
||||
|
||||
/** The end of the last audio or text in the file, in ms. */
|
||||
val durationMs: Long
|
||||
|
||||
init {
|
||||
scan()
|
||||
var end = infoDurationMs
|
||||
if (clusters.isNotEmpty()) {
|
||||
for (block in readCluster(clusters.last())) end = maxOf(end, block.timeMs + blockSpanMs(block))
|
||||
}
|
||||
durationMs = end
|
||||
}
|
||||
|
||||
fun track(number: Int): MkaTrackInfo? = trackList.firstOrNull { it.number == number }
|
||||
|
||||
/** Continue from the first block at or after [timeMs]. */
|
||||
fun seekTo(timeMs: Long) {
|
||||
val index = clusters.indexOfLast { toMs(it.rawTime) <= timeMs }
|
||||
nextCluster = maxOf(0, index)
|
||||
blocks = emptyList()
|
||||
blockIndex = 0
|
||||
skipBeforeMs = timeMs
|
||||
}
|
||||
|
||||
/** The next block, or null at the end of the file. */
|
||||
fun next(): MkaBlock? {
|
||||
while (true) {
|
||||
while (blockIndex >= blocks.size) {
|
||||
if (nextCluster >= clusters.size) return null
|
||||
blocks = readCluster(clusters[nextCluster++])
|
||||
blockIndex = 0
|
||||
}
|
||||
val block = blocks[blockIndex++]
|
||||
if (block.timeMs >= skipBeforeMs) return block
|
||||
}
|
||||
}
|
||||
|
||||
fun close() = input.close()
|
||||
|
||||
private fun toMs(raw: Long): Long = raw * timestampScale / 1_000_000L
|
||||
|
||||
private fun blockSpanMs(block: MkaBlock): Long {
|
||||
if (block.durationMs >= 0) return block.durationMs
|
||||
val track = track(block.track) ?: return 0
|
||||
if (!track.isPcm16 || track.channels <= 0 || track.sampleRate <= 0) return 0
|
||||
return block.data.size / (2L * track.channels) * 1000L / track.sampleRate
|
||||
}
|
||||
|
||||
/** An element header at [at] in [bytes]: its ID, payload size (-1 unknown) and header length; null when cut off. */
|
||||
private class Header(val id: Int, val size: Long, val length: Int)
|
||||
|
||||
private fun header(bytes: ByteArray, at: Int, end: Int): Header? {
|
||||
if (at >= end) return null
|
||||
val first = bytes[at].toInt() and 0xff
|
||||
val idLength = first.countLeadingZeroBits() - 24 + 1
|
||||
if (idLength > 4 || at + idLength >= end) return null
|
||||
var id = 0
|
||||
for (index in 0 until idLength) id = (id shl 8) or (bytes[at + index].toInt() and 0xff)
|
||||
val sizeAt = at + idLength
|
||||
val sizeFirst = bytes[sizeAt].toInt() and 0xff
|
||||
val sizeLength = sizeFirst.countLeadingZeroBits() - 24 + 1
|
||||
if (sizeLength > 8 || sizeAt + sizeLength > end) return null
|
||||
var size = (sizeFirst and (0xff shr sizeLength)).toLong()
|
||||
var allOnes = size == (0xff shr sizeLength).toLong()
|
||||
for (index in 1 until sizeLength) {
|
||||
val byte = bytes[sizeAt + index].toInt() and 0xff
|
||||
if (byte != 0xff) allOnes = false
|
||||
size = (size shl 8) or byte.toLong()
|
||||
}
|
||||
return Header(id, if (allOnes) -1 else size, idLength + sizeLength)
|
||||
}
|
||||
|
||||
private fun readAt(position: Long, length: Int): ByteArray {
|
||||
val buffer = ByteArray(length)
|
||||
val read = input.readAt(position, buffer, 0, length)
|
||||
return if (read == length) buffer else buffer.copyOf(maxOf(0, read))
|
||||
}
|
||||
|
||||
private fun uint(bytes: ByteArray, at: Int, size: Int): Long {
|
||||
var value = 0L
|
||||
for (index in 0 until size) value = (value shl 8) or (bytes[at + index].toLong() and 0xff)
|
||||
return value
|
||||
}
|
||||
|
||||
private fun float(bytes: ByteArray, at: Int, size: Int): Double = when (size) {
|
||||
4 -> Float.fromBits(uint(bytes, at, 4).toInt()).toDouble()
|
||||
8 -> Double.fromBits(uint(bytes, at, 8))
|
||||
else -> 0.0
|
||||
}
|
||||
|
||||
/** Calls [visit] with the ID, payload offset and payload size of each complete child element. */
|
||||
private inline fun children(bytes: ByteArray, start: Int, end: Int, visit: (id: Int, at: Int, size: Int) -> Unit) {
|
||||
var at = start
|
||||
while (true) {
|
||||
val child = header(bytes, at, end) ?: return
|
||||
val payload = at + child.length
|
||||
if (child.size < 0 || child.size > end - payload) return
|
||||
visit(child.id, payload, child.size.toInt())
|
||||
at = payload + child.size.toInt()
|
||||
}
|
||||
}
|
||||
|
||||
private fun scan() {
|
||||
val length = input.length
|
||||
val start = readAt(0, 64)
|
||||
val ebml = header(start, 0, start.size)
|
||||
require(ebml != null && ebml.id == MatroskaIds.EBML && ebml.size >= 0) { "not a Matroska file" }
|
||||
var position = ebml.length + ebml.size
|
||||
val segment = readAt(position, 12).let { header(it, 0, it.size) }
|
||||
require(segment != null && segment.id == MatroskaIds.SEGMENT) { "not a Matroska file" }
|
||||
position += segment.length
|
||||
val end = if (segment.size < 0) length else minOf(length, position + segment.size)
|
||||
while (position < end) {
|
||||
val head = readAt(position, 12)
|
||||
val element = header(head, 0, head.size) ?: break
|
||||
val payload = position + element.length
|
||||
if (element.id == MatroskaIds.CLUSTER) {
|
||||
require(element.size >= 0) { "unsupported Matroska file (clusters of unknown size)" }
|
||||
// A recording cut short ends in a partial cluster; its complete blocks still count.
|
||||
val available = minOf(element.size, end - payload)
|
||||
require(available <= Int.MAX_VALUE) { "unsupported Matroska file (oversized cluster)" }
|
||||
clusters.add(ClusterRef(payload, available.toInt(), clusterTime(payload, available.toInt())))
|
||||
} else if (element.size < 0) {
|
||||
break
|
||||
} else if (element.size <= end - payload && element.size <= 64L * 1024 * 1024) {
|
||||
when (element.id) {
|
||||
MatroskaIds.INFO -> parseInfo(readAt(payload, element.size.toInt()))
|
||||
MatroskaIds.TRACKS -> parseTracks(readAt(payload, element.size.toInt()))
|
||||
MatroskaIds.ATTACHMENTS -> parseAttachments(readAt(payload, element.size.toInt()))
|
||||
}
|
||||
}
|
||||
if (element.size < 0) break
|
||||
position = payload + element.size
|
||||
}
|
||||
}
|
||||
|
||||
private fun clusterTime(position: Long, size: Int): Long {
|
||||
val head = readAt(position, minOf(size, 16))
|
||||
val first = header(head, 0, head.size)
|
||||
if (first != null && first.id == MatroskaIds.TIMESTAMP && first.size in 0..8 && first.length + first.size <= head.size) {
|
||||
return uint(head, first.length, first.size.toInt())
|
||||
}
|
||||
val whole = readAt(position, size)
|
||||
var time = 0L
|
||||
children(whole, 0, whole.size) { id, at, length -> if (id == MatroskaIds.TIMESTAMP) time = uint(whole, at, length) }
|
||||
return time
|
||||
}
|
||||
|
||||
private fun parseInfo(bytes: ByteArray) {
|
||||
var duration = 0.0
|
||||
children(bytes, 0, bytes.size) { id, at, size ->
|
||||
when (id) {
|
||||
MatroskaIds.TIMESTAMP_SCALE -> timestampScale = maxOf(1L, uint(bytes, at, size))
|
||||
MatroskaIds.DURATION -> duration = float(bytes, at, size)
|
||||
}
|
||||
}
|
||||
infoDurationMs = (duration * timestampScale / 1_000_000.0).toLong()
|
||||
}
|
||||
|
||||
private fun parseTracks(bytes: ByteArray) {
|
||||
trackList.clear()
|
||||
children(bytes, 0, bytes.size) { id, at, size ->
|
||||
if (id == MatroskaIds.TRACK_ENTRY) {
|
||||
var number = 0
|
||||
var type = 0
|
||||
var codec = ""
|
||||
var name = ""
|
||||
var rate = 8000.0
|
||||
var channels = 1
|
||||
var depth = 0
|
||||
children(bytes, at, at + size) { field, fieldAt, fieldSize ->
|
||||
when (field) {
|
||||
MatroskaIds.TRACK_NUMBER -> number = uint(bytes, fieldAt, fieldSize).toInt()
|
||||
MatroskaIds.TRACK_TYPE -> type = uint(bytes, fieldAt, fieldSize).toInt()
|
||||
MatroskaIds.CODEC_ID -> codec = bytes.decodeToString(fieldAt, fieldAt + fieldSize).trimEnd('\u0000')
|
||||
MatroskaIds.NAME -> name = bytes.decodeToString(fieldAt, fieldAt + fieldSize).trimEnd('\u0000')
|
||||
MatroskaIds.AUDIO -> children(bytes, fieldAt, fieldAt + fieldSize) { audio, audioAt, audioSize ->
|
||||
when (audio) {
|
||||
MatroskaIds.SAMPLING_FREQUENCY -> rate = float(bytes, audioAt, audioSize)
|
||||
MatroskaIds.CHANNELS -> channels = uint(bytes, audioAt, audioSize).toInt()
|
||||
MatroskaIds.BIT_DEPTH -> depth = uint(bytes, audioAt, audioSize).toInt()
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
val audio = type == MkaWriter.TRACK_AUDIO
|
||||
trackList.add(
|
||||
MkaTrackInfo(number, type, codec, name, if (audio) rate.toInt() else 0, if (audio) channels else 0, if (audio) depth else 0),
|
||||
)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
private fun parseAttachments(bytes: ByteArray) {
|
||||
children(bytes, 0, bytes.size) { id, at, size ->
|
||||
if (id == MatroskaIds.ATTACHED_FILE) {
|
||||
var name = ""
|
||||
var mime = ""
|
||||
var data = ByteArray(0)
|
||||
children(bytes, at, at + size) { field, fieldAt, fieldSize ->
|
||||
when (field) {
|
||||
MatroskaIds.FILE_NAME -> name = bytes.decodeToString(fieldAt, fieldAt + fieldSize)
|
||||
MatroskaIds.FILE_MIME_TYPE -> mime = bytes.decodeToString(fieldAt, fieldAt + fieldSize)
|
||||
MatroskaIds.FILE_DATA -> data = bytes.copyOfRange(fieldAt, fieldAt + fieldSize)
|
||||
}
|
||||
}
|
||||
attachmentList.add(MkaAttachment(name, mime, data))
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
private fun readCluster(cluster: ClusterRef): List<MkaBlock> {
|
||||
val bytes = readAt(cluster.position, cluster.size)
|
||||
val out = ArrayList<MkaBlock>()
|
||||
var clusterRaw = cluster.rawTime
|
||||
children(bytes, 0, bytes.size) { id, at, size ->
|
||||
when (id) {
|
||||
MatroskaIds.TIMESTAMP -> clusterRaw = uint(bytes, at, size)
|
||||
MatroskaIds.SIMPLE_BLOCK -> block(bytes, at, size, clusterRaw, -1)?.let { out.add(it) }
|
||||
MatroskaIds.BLOCK_GROUP -> {
|
||||
var blockAt = -1
|
||||
var blockSize = 0
|
||||
var duration = -1L
|
||||
children(bytes, at, at + size) { field, fieldAt, fieldSize ->
|
||||
when (field) {
|
||||
MatroskaIds.BLOCK -> {
|
||||
blockAt = fieldAt
|
||||
blockSize = fieldSize
|
||||
}
|
||||
MatroskaIds.BLOCK_DURATION -> duration = toMs(uint(bytes, fieldAt, fieldSize))
|
||||
}
|
||||
}
|
||||
if (blockAt >= 0) block(bytes, blockAt, blockSize, clusterRaw, duration)?.let { out.add(it) }
|
||||
}
|
||||
}
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
/** A block's payload: track number, 16-bit timestamp offset, flags, then the frame. Laced blocks are skipped. */
|
||||
private fun block(bytes: ByteArray, at: Int, size: Int, clusterRaw: Long, durationMs: Long): MkaBlock? {
|
||||
if (size < 4) return null
|
||||
val first = bytes[at].toInt() and 0xff
|
||||
val trackLength = first.countLeadingZeroBits() - 24 + 1
|
||||
if (trackLength > 2 || size < trackLength + 3) return null
|
||||
var track = first and (0xff shr trackLength)
|
||||
if (trackLength == 2) track = (track shl 8) or (bytes[at + 1].toInt() and 0xff)
|
||||
val head = at + trackLength
|
||||
val relative = ((bytes[head].toInt() shl 8) or (bytes[head + 1].toInt() and 0xff)).toShort().toLong()
|
||||
val flags = bytes[head + 2].toInt() and 0xff
|
||||
if (flags and 0x06 != 0) return null
|
||||
return MkaBlock(track, toMs(clusterRaw + relative), bytes.copyOfRange(head + 3, at + size), durationMs)
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,93 @@
|
||||
package com.faceclaw.app
|
||||
|
||||
import kotlin.jvm.JvmField
|
||||
import kotlin.jvm.JvmStatic
|
||||
|
||||
/**
|
||||
* The parts of the CFW mic_control wire format the capture session speaks itself: the 'SM'
|
||||
* multi-channel stream frame each temple sends, and the fixed STOP/RENEW control records. The
|
||||
* CONFIGURE record is built by the caller (app/apps/microphones/mic-protocol.ts encodes it; the
|
||||
* contract is the header of g2flash patches/mic_control.c).
|
||||
*/
|
||||
object MicArrayProtocol {
|
||||
const val VERSION = 1
|
||||
const val HEADER_BYTES = 21
|
||||
const val OP_CONFIGURE = 1
|
||||
const val OP_STOP = 3
|
||||
const val OP_RENEW = 4
|
||||
/** How long an armed temple streams without a CONFIGURE or RENEW. */
|
||||
const val LEASE_MS = 90_000L
|
||||
const val FLAG_TRUNCATED = 0x80
|
||||
const val CODEC_RAW = 1
|
||||
const val FORMAT_PCM16 = 0
|
||||
|
||||
/** One decoded 'SM' frame header; the samples are [payloadLength] bytes at [payloadOffset] of its packet. */
|
||||
class Frame(
|
||||
/** Low 16 bits of the temple's frame counter. */
|
||||
@JvmField val sequence: Int,
|
||||
/** The temple's own 1 ms tick when it packed the frame (unsigned 32-bit). */
|
||||
@JvmField val deviceTickMs: Long,
|
||||
@JvmField val channels: Int,
|
||||
@JvmField val sampleRate: Int,
|
||||
@JvmField val format: Int,
|
||||
@JvmField val codec: Int,
|
||||
/** The temple's own bearing estimate, in signed degrees (0 unless it was asked for one). */
|
||||
@JvmField val angleDegrees: Int,
|
||||
@JvmField val ssr: Int,
|
||||
/** The firmware cut the payload short of the samples it captured. */
|
||||
@JvmField val truncated: Boolean,
|
||||
@JvmField val payloadOffset: Int,
|
||||
@JvmField val payloadLength: Int,
|
||||
)
|
||||
|
||||
/** A control record with no payload: ['M','C', version, op]. */
|
||||
@JvmStatic
|
||||
fun control(op: Int): ByteArray = byteArrayOf('M'.code.toByte(), 'C'.code.toByte(), VERSION.toByte(), op.toByte())
|
||||
|
||||
@JvmStatic
|
||||
fun isConfigure(record: ByteArray?): Boolean =
|
||||
record != null && record.size >= 13 && record[0] == 'M'.code.toByte() && record[1] == 'C'.code.toByte() &&
|
||||
record[2] == VERSION.toByte() && record[3] == OP_CONFIGURE.toByte()
|
||||
|
||||
/** Decodes a stream frame's header, or returns null when [data] isn't one. */
|
||||
@JvmStatic
|
||||
fun decodeFrame(data: ByteArray): Frame? {
|
||||
if (data.size < HEADER_BYTES || data[0] != 'S'.code.toByte() || data[1] != 'M'.code.toByte() || data[2] != VERSION.toByte()) {
|
||||
return null
|
||||
}
|
||||
val length = minOf(u16(data, 19), data.size - HEADER_BYTES)
|
||||
return Frame(
|
||||
sequence = u16(data, 4),
|
||||
deviceTickMs = u16(data, 6).toLong() or (u16(data, 8).toLong() shl 16),
|
||||
channels = data[10].toInt() and 0xff,
|
||||
sampleRate = u16(data, 11) * 100,
|
||||
format = data[13].toInt() and 0xff,
|
||||
codec = data[14].toInt() and 0xff,
|
||||
angleDegrees = u16(data, 15).toShort().toInt(),
|
||||
ssr = u16(data, 17).toShort().toInt(),
|
||||
truncated = (data[3].toInt() and FLAG_TRUNCATED) != 0,
|
||||
payloadOffset = HEADER_BYTES,
|
||||
payloadLength = length,
|
||||
)
|
||||
}
|
||||
|
||||
/**
|
||||
* The frame's samples as interleaved PCM16 (whole sample frames only), or null when the
|
||||
* payload isn't raw 16-bit PCM or holds none.
|
||||
*/
|
||||
@JvmStatic
|
||||
fun pcm16(data: ByteArray, frame: Frame): ShortArray? {
|
||||
if (frame.codec != CODEC_RAW || frame.format != FORMAT_PCM16 || frame.channels < 1) return null
|
||||
val frames = frame.payloadLength / (2 * frame.channels)
|
||||
if (frames <= 0) return null
|
||||
val samples = ShortArray(frames * frame.channels)
|
||||
var at = frame.payloadOffset
|
||||
for (index in samples.indices) {
|
||||
samples[index] = AudioSegmentation.pcm16le(data[at], data[at + 1])
|
||||
at += 2
|
||||
}
|
||||
return samples
|
||||
}
|
||||
|
||||
private fun u16(data: ByteArray, at: Int): Int = (data[at].toInt() and 0xff) or ((data[at + 1].toInt() and 0xff) shl 8)
|
||||
}
|
||||
@@ -0,0 +1,484 @@
|
||||
package com.faceclaw.app
|
||||
|
||||
import kotlin.concurrent.Volatile
|
||||
import kotlin.jvm.JvmField
|
||||
import kotlin.jvm.JvmStatic
|
||||
import kotlin.math.abs
|
||||
|
||||
/** Which microphones a chunk of capture audio came from. */
|
||||
object MicStream {
|
||||
/** One mixed channel: the glasses' stock stream, or the phone's own microphone. */
|
||||
const val MONO = 0
|
||||
/** A temple's own microphones, on that temple's clock (channel 0 front, channel 1 rear). */
|
||||
const val LEFT = 1
|
||||
const val RIGHT = 2
|
||||
/** The mono feed derived from the temples for speech recognition: recorded, never replayed. */
|
||||
const val MIX = 3
|
||||
const val COUNT = 4
|
||||
|
||||
/** The stream's name in frame metadata. */
|
||||
@JvmStatic
|
||||
fun key(stream: Int): String = when (stream) {
|
||||
MONO -> "mic"
|
||||
LEFT -> "left"
|
||||
RIGHT -> "right"
|
||||
else -> "mix"
|
||||
}
|
||||
|
||||
/** The name of the stream's audio tracks in a recording. */
|
||||
@JvmStatic
|
||||
fun trackName(stream: Int): String = when (stream) {
|
||||
MONO -> "mic"
|
||||
LEFT -> "left temple"
|
||||
RIGHT -> "right temple"
|
||||
else -> "mix"
|
||||
}
|
||||
|
||||
/** The stream a recording's track or metadata names, or -1. */
|
||||
@JvmStatic
|
||||
fun named(name: String): Int = (0 until COUNT).firstOrNull { trackName(it) == name || key(it) == name } ?: -1
|
||||
}
|
||||
|
||||
/**
|
||||
* One chunk of microphone audio as a capture source delivered it (about 50 ms), before any
|
||||
* processing. Live captures, recordings and replays all trade in these.
|
||||
*/
|
||||
class MicChunk(
|
||||
@JvmField val stream: Int,
|
||||
@JvmField val channels: Int,
|
||||
@JvmField val sampleRate: Int,
|
||||
/** Interleaved PCM16; the first [frames] sample frames are the chunk. */
|
||||
@JvmField val pcm: ShortArray,
|
||||
@JvmField val frames: Int,
|
||||
/** When the chunk arrived: the phone's monotonic clock for a live capture, the offset into the file for a replay. */
|
||||
@JvmField val timeMs: Long,
|
||||
) {
|
||||
/** Whether the source reported a direction of arrival with the chunk. */
|
||||
@JvmField var hasDirection = false
|
||||
@JvmField var angleDegrees = 0
|
||||
@JvmField var ssr = 0
|
||||
/** The source's packet counter, or -1 when it has none. */
|
||||
@JvmField var sequence = -1
|
||||
/** The source's own clock when it captured the chunk (unsigned 32-bit ms), or -1 when it has none. */
|
||||
@JvmField var deviceTickMs = -1L
|
||||
/** Whose clock [deviceTickMs] is on: the stream the samples were captured by. */
|
||||
@JvmField var clockStream = stream
|
||||
/** The source cut the chunk short of what it captured. */
|
||||
@JvmField var truncated = false
|
||||
/** Sample frames known to be missing between the previous chunk of this stream and this one. */
|
||||
@JvmField var lostFramesBefore = 0
|
||||
|
||||
val durationMs: Long get() = frames * 1000L / sampleRate
|
||||
}
|
||||
|
||||
/**
|
||||
* Derives the mono feed speech recognition hears from the temples' chunks. This is the seam for
|
||||
* array processing (beamforming, noise cancellation): replays feed it the recorded chunks with
|
||||
* the times they first arrived at, so an implementation that decides from the chunks alone
|
||||
* behaves the same on a replay as it did live.
|
||||
*/
|
||||
interface MicArrayMixer {
|
||||
/** The mono chunk [chunk] contributes to the feed, or null when it contributes nothing. */
|
||||
fun mix(chunk: MicChunk): MicChunk?
|
||||
|
||||
fun reset()
|
||||
}
|
||||
|
||||
/**
|
||||
* The feed is one temple at a time, its microphones averaged (the firmware's own stock downmix).
|
||||
* The temples sample on independent clocks, so their audio is never summed. The left temple
|
||||
* feeds whenever it is streaming; the right one stands in while it is not.
|
||||
*/
|
||||
class SingleTempleMixer : MicArrayMixer {
|
||||
companion object {
|
||||
/** A temple that has delivered nothing for this long no longer counts as streaming. */
|
||||
const val SILENT_MS = 300L
|
||||
}
|
||||
|
||||
private var leftAtMs = Long.MIN_VALUE
|
||||
|
||||
override fun reset() {
|
||||
leftAtMs = Long.MIN_VALUE
|
||||
}
|
||||
|
||||
override fun mix(chunk: MicChunk): MicChunk? {
|
||||
if (chunk.stream == MicStream.LEFT) {
|
||||
leftAtMs = chunk.timeMs
|
||||
} else if (chunk.stream != MicStream.RIGHT || (leftAtMs != Long.MIN_VALUE && chunk.timeMs - leftAtMs < SILENT_MS)) {
|
||||
return null
|
||||
}
|
||||
val channels = chunk.channels
|
||||
val mono = ShortArray(chunk.frames)
|
||||
for (frame in 0 until chunk.frames) {
|
||||
var sum = 0
|
||||
for (channel in 0 until channels) sum += chunk.pcm[frame * channels + channel]
|
||||
mono[frame] = (sum / channels).toShort()
|
||||
}
|
||||
val out = MicChunk(MicStream.MIX, 1, chunk.sampleRate, mono, chunk.frames, chunk.timeMs)
|
||||
out.hasDirection = chunk.hasDirection
|
||||
out.angleDegrees = chunk.angleDegrees
|
||||
out.ssr = chunk.ssr
|
||||
out.deviceTickMs = chunk.deviceTickMs
|
||||
out.clockStream = chunk.stream
|
||||
out.truncated = chunk.truncated
|
||||
return out
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Records a capture's chunks to a Matroska audio file (.mka), one PCM16 track per stream and
|
||||
* layout as each first appears, plus a "frames" text track with one JSON object per chunk that
|
||||
* has metadata, written just ahead of the chunk's audio. Format: notes/mic-recording-format.md.
|
||||
*
|
||||
* Block timestamps are capture times, not arrival times. A stream's own clock (a temple's tick,
|
||||
* or the count of samples from a source without one) says how far apart its chunks were
|
||||
* captured; the least-delayed recent arrival places that clock on the file's timeline. So a
|
||||
* chunk held up over Bluetooth still lands where it was captured, and lost audio leaves a gap.
|
||||
*
|
||||
* Safe to share between the capture thread and whoever finishes the recording.
|
||||
*/
|
||||
class MicRecordingWriter(
|
||||
output: SeekableOutput,
|
||||
/** A JSON document describing the capture, stored as the file's "capture.json" attachment. */
|
||||
configJson: String?,
|
||||
/** When the recording began (Unix epoch ms). */
|
||||
startedAtMs: Long,
|
||||
platform: ProtocolPlatform = protocolPlatform(),
|
||||
private val logWarn: (String) -> Unit = { PlatformLog.w("FaceclawMicRecording", it) },
|
||||
) {
|
||||
companion object {
|
||||
const val CONFIG_ATTACHMENT = "capture.json"
|
||||
const val META_TRACK = "frames"
|
||||
/** How far a chunk's arrival may sit from where its stream's clock puts it before the clock is re-placed. */
|
||||
const val DISCONTINUITY_MS = 5000L
|
||||
private const val WINDOW_BUCKETS = 8
|
||||
private const val BUCKET_MS = 1000L
|
||||
}
|
||||
|
||||
/**
|
||||
* Places one track's chunks on the file's timeline. `source` time is the stream's own clock
|
||||
* since this clock was anchored; the offset from it to arrival time is taken as the least
|
||||
* seen over the last few seconds, which is the least-delayed delivery.
|
||||
*/
|
||||
private class TrackClock(val number: Int, val sampleRate: Int) {
|
||||
private var anchored = false
|
||||
private var clockStream = -1
|
||||
private var tickBase = -1L
|
||||
private var lastTick = -1L
|
||||
private var frames = 0L
|
||||
private val minima = LongArray(WINDOW_BUCKETS)
|
||||
private var bucket = 0
|
||||
private var bucketStartMs = 0L
|
||||
var endMs = Long.MIN_VALUE
|
||||
private set
|
||||
|
||||
fun reset() {
|
||||
anchored = false
|
||||
}
|
||||
|
||||
private fun offset(): Long = minima.min()
|
||||
|
||||
fun stamp(chunk: MicChunk): Long {
|
||||
val arrival = chunk.timeMs
|
||||
val ticked = chunk.deviceTickMs >= 0
|
||||
if (anchored) {
|
||||
// Another clock (the feed moved to the other temple), or this one restarted.
|
||||
if (chunk.clockStream != clockStream || ticked != (tickBase >= 0)) anchored = false
|
||||
else if (ticked && (chunk.deviceTickMs - lastTick).toInt() < 0) anchored = false
|
||||
}
|
||||
var anchorNow = !anchored
|
||||
var source = 0L
|
||||
if (anchored) {
|
||||
frames += chunk.lostFramesBefore
|
||||
source = sourceTime(chunk)
|
||||
if (abs(arrival - source - offset()) > DISCONTINUITY_MS) anchorNow = true
|
||||
}
|
||||
if (anchorNow) {
|
||||
anchored = true
|
||||
clockStream = chunk.clockStream
|
||||
tickBase = if (ticked) chunk.deviceTickMs else -1L
|
||||
frames = 0
|
||||
minima.fill(Long.MAX_VALUE)
|
||||
bucket = 0
|
||||
bucketStartMs = arrival
|
||||
source = 0
|
||||
}
|
||||
if (arrival - bucketStartMs >= BUCKET_MS) {
|
||||
bucket = (bucket + 1) % WINDOW_BUCKETS
|
||||
minima[bucket] = Long.MAX_VALUE
|
||||
bucketStartMs = arrival
|
||||
}
|
||||
if (arrival - source < minima[bucket]) minima[bucket] = arrival - source
|
||||
// Never before the end of the track's previous block.
|
||||
val time = maxOf(source + offset(), endMs)
|
||||
frames += chunk.frames
|
||||
lastTick = chunk.deviceTickMs
|
||||
endMs = time + chunk.durationMs
|
||||
return time
|
||||
}
|
||||
|
||||
private fun sourceTime(chunk: MicChunk): Long =
|
||||
if (tickBase >= 0) (chunk.deviceTickMs - tickBase) and 0xffffffffL else frames * 1000L / sampleRate
|
||||
}
|
||||
|
||||
private val lock = platform.createLock()
|
||||
private val writer: MkaWriter
|
||||
private val clocks = HashMap<Long, TrackClock>()
|
||||
private var metaTrack = 0
|
||||
private var originMs = Long.MIN_VALUE
|
||||
private var lastTimeMs = 0L
|
||||
private var closed = false
|
||||
|
||||
/** The recording stopped early: the file could not be written. */
|
||||
@Volatile
|
||||
var failed = false
|
||||
private set
|
||||
|
||||
init {
|
||||
val attachments = if (configJson.isNullOrEmpty()) emptyList()
|
||||
else listOf(MkaAttachment(CONFIG_ATTACHMENT, "application/json", configJson.encodeToByteArray()))
|
||||
writer = MkaWriter(output, null, startedAtMs, attachments)
|
||||
}
|
||||
|
||||
/** Length of the recording so far, in ms. */
|
||||
val durationMs: Long get() = lock.withLock { writer.durationMs }
|
||||
|
||||
/** Appends [chunk] to its stream's track, adding the track on its first chunk. */
|
||||
fun write(chunk: MicChunk) {
|
||||
if (chunk.frames <= 0) return
|
||||
lock.withLock {
|
||||
if (closed) return
|
||||
try {
|
||||
val clock = clockFor(chunk) ?: return
|
||||
if (originMs == Long.MIN_VALUE) originMs = chunk.timeMs
|
||||
val time = clock.stamp(chunk) - originMs
|
||||
val meta = metadata(chunk)
|
||||
if (meta != null && metaTrack > 0) writer.writeText(metaTrack, time, maxOf(1L, chunk.durationMs), meta)
|
||||
val bytes = AudioSegmentation.pcm16ToLittleEndian(chunk.pcm, 0, chunk.frames * chunk.channels)
|
||||
writer.writeAudio(clock.number, time, bytes, 0, bytes.size)
|
||||
lastTimeMs = maxOf(lastTimeMs, time)
|
||||
} catch (error: Throwable) {
|
||||
fail(error)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Records a JSON object as an event at the recording's current position (a capture starting
|
||||
* over with another configuration, say), and has every stream's clock placed afresh by its
|
||||
* next chunk: what follows an event is not continuous with what came before.
|
||||
*/
|
||||
fun event(json: String) {
|
||||
lock.withLock {
|
||||
if (closed) return
|
||||
try {
|
||||
for (clock in clocks.values) clock.reset()
|
||||
if (ensureMetaTrack()) writer.writeText(metaTrack, maxOf(lastTimeMs, writer.durationMs), 1, json)
|
||||
} catch (error: Throwable) {
|
||||
fail(error)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/** Completes the file. Returns the recording's length in ms. */
|
||||
fun finish(): Long {
|
||||
lock.withLock {
|
||||
if (!closed) {
|
||||
closed = true
|
||||
try {
|
||||
writer.finish()
|
||||
} catch (error: Throwable) {
|
||||
failed = true
|
||||
logWarn("finishing the recording failed: ${error.message}")
|
||||
}
|
||||
}
|
||||
return writer.durationMs
|
||||
}
|
||||
}
|
||||
|
||||
private fun fail(error: Throwable) {
|
||||
failed = true
|
||||
closed = true
|
||||
logWarn("recording stopped: ${error.message}")
|
||||
try {
|
||||
writer.finish()
|
||||
} catch (ignored: Throwable) {
|
||||
// The file is whatever made it to disk.
|
||||
}
|
||||
}
|
||||
|
||||
private fun ensureMetaTrack(): Boolean {
|
||||
if (metaTrack == 0) metaTrack = writer.addTextTrack(META_TRACK)
|
||||
return metaTrack > 0
|
||||
}
|
||||
|
||||
private fun clockFor(chunk: MicChunk): TrackClock? {
|
||||
val key = (chunk.stream.toLong() shl 48) or (chunk.channels.toLong() shl 32) or chunk.sampleRate.toLong()
|
||||
clocks[key]?.let { return it }
|
||||
// The frames track goes in with the first audio track, so it is there for that track's first chunk.
|
||||
ensureMetaTrack()
|
||||
// The mixed streams are what a player should pick; a temple's raw pair is for analysis.
|
||||
val preferred = chunk.stream == MicStream.MONO || chunk.stream == MicStream.MIX
|
||||
val number = writer.addAudioTrack(MicStream.trackName(chunk.stream), chunk.sampleRate, chunk.channels, preferred)
|
||||
if (number <= 0) {
|
||||
logWarn("recording has no room for another track; dropping ${MicStream.trackName(chunk.stream)} audio")
|
||||
return null
|
||||
}
|
||||
return TrackClock(number, chunk.sampleRate).also { clocks[key] = it }
|
||||
}
|
||||
|
||||
/** The chunk's frame-metadata JSON, or null when it has nothing to say beyond its audio. */
|
||||
private fun metadata(chunk: MicChunk): String? {
|
||||
if (chunk.stream == MicStream.MIX) return null
|
||||
if (!chunk.hasDirection && chunk.sequence < 0 && chunk.deviceTickMs < 0 && !chunk.truncated && chunk.lostFramesBefore == 0) return null
|
||||
val text = StringBuilder(96)
|
||||
text.append("{\"s\":\"").append(MicStream.key(chunk.stream)).append("\",\"rx\":").append(chunk.timeMs - originMs)
|
||||
if (chunk.sequence >= 0) text.append(",\"seq\":").append(chunk.sequence)
|
||||
if (chunk.deviceTickMs >= 0) text.append(",\"tick\":").append(chunk.deviceTickMs)
|
||||
if (chunk.hasDirection) text.append(",\"angle\":").append(chunk.angleDegrees).append(",\"ssr\":").append(chunk.ssr)
|
||||
if (chunk.truncated) text.append(",\"trunc\":true")
|
||||
if (chunk.lostFramesBefore > 0) text.append(",\"lost\":").append(chunk.lostFramesBefore)
|
||||
return text.append('}').toString()
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Reads a recording back as the chunks its capture delivered: the "mic" track as [MicStream.MONO]
|
||||
* chunks and the temple tracks as [MicStream.LEFT] / [MicStream.RIGHT] ones, each with the
|
||||
* metadata its frame carried. The "mix" track is skipped: a replay derives the feed again. A
|
||||
* Matroska file from elsewhere plays its first PCM16 track as the mono microphone.
|
||||
*
|
||||
* Throws IllegalArgumentException when the file has no audio a replay can use.
|
||||
*/
|
||||
class MicRecordingReader(input: SeekableInput) {
|
||||
companion object {
|
||||
const val SAMPLE_RATE = 16000
|
||||
/** Blocks longer than this many frames are handed out in pieces, like a live capture's chunks. */
|
||||
const val MAX_CHUNK_FRAMES = SAMPLE_RATE / 20
|
||||
}
|
||||
|
||||
private class Meta(val timeMs: Long, val fields: Map<String, Any?>)
|
||||
|
||||
private val reader = MkaReader(input)
|
||||
private val streams = HashMap<Int, Int>()
|
||||
private var metaTrack = -1
|
||||
private val pendingMeta = arrayOfNulls<Meta>(MicStream.COUNT)
|
||||
private val queued = ArrayDeque<MicChunk>()
|
||||
|
||||
val durationMs: Long get() = reader.durationMs
|
||||
|
||||
/** The capture's "capture.json" attachment, or null. */
|
||||
val configJson: String? =
|
||||
reader.attachments.firstOrNull { it.name == MicRecordingWriter.CONFIG_ATTACHMENT }?.data?.decodeToString()
|
||||
|
||||
/** What the replay plays, for a status line: "left temple 2ch + right temple 2ch". */
|
||||
val description: String
|
||||
|
||||
init {
|
||||
try {
|
||||
val audio = reader.tracks.filter { it.isPcm16 && it.channels in 1..8 }
|
||||
for (track in audio) {
|
||||
val stream = MicStream.named(track.name)
|
||||
if (stream == MicStream.MONO || stream == MicStream.LEFT || stream == MicStream.RIGHT) streams[track.number] = stream
|
||||
}
|
||||
if (streams.isEmpty()) {
|
||||
// Not one of ours: its first usable track stands in for the microphone.
|
||||
val first = audio.firstOrNull { MicStream.named(it.name) != MicStream.MIX } ?: audio.firstOrNull()
|
||||
require(first != null) { "no 16-bit PCM audio track" }
|
||||
streams[first.number] = MicStream.MONO
|
||||
}
|
||||
for (number in streams.keys) {
|
||||
val rate = reader.track(number)!!.sampleRate
|
||||
require(rate == SAMPLE_RATE) { "unsupported sample rate ($rate Hz; replay needs ${SAMPLE_RATE / 1000} kHz)" }
|
||||
}
|
||||
metaTrack = reader.tracks.firstOrNull {
|
||||
it.type == MkaWriter.TRACK_SUBTITLE && it.name == MicRecordingWriter.META_TRACK
|
||||
}?.number ?: -1
|
||||
description = streams.entries.sortedBy { it.value }.joinToString(" + ") { (number, stream) ->
|
||||
val channels = reader.track(number)!!.channels
|
||||
MicStream.trackName(stream) + if (channels > 1 || stream != MicStream.MONO) " ${channels}ch" else ""
|
||||
}
|
||||
} catch (error: Throwable) {
|
||||
reader.close()
|
||||
throw error
|
||||
}
|
||||
}
|
||||
|
||||
/** Continue from the first chunk at or after [timeMs]. */
|
||||
fun seekTo(timeMs: Long) {
|
||||
queued.clear()
|
||||
pendingMeta.fill(null)
|
||||
reader.seekTo(timeMs)
|
||||
}
|
||||
|
||||
/** The next chunk in file order, or null at the end. */
|
||||
fun next(): MicChunk? {
|
||||
while (queued.isEmpty()) {
|
||||
val block = reader.next() ?: return null
|
||||
if (block.track == metaTrack) {
|
||||
noteMeta(block)
|
||||
continue
|
||||
}
|
||||
val stream = streams[block.track] ?: continue
|
||||
val track = reader.track(block.track) ?: continue
|
||||
queueChunks(block, stream, track)
|
||||
}
|
||||
return queued.removeFirst()
|
||||
}
|
||||
|
||||
fun close() = reader.close()
|
||||
|
||||
private fun noteMeta(block: MkaBlock) {
|
||||
try {
|
||||
val fields = Json.parseObject(block.data.decodeToString())
|
||||
val stream = MicStream.named(fields["s"].asString() ?: return)
|
||||
if (stream >= 0) pendingMeta[stream] = Meta(block.timeMs, fields)
|
||||
} catch (ignored: Throwable) {
|
||||
// An event or a frame we can't read: the audio plays without it.
|
||||
}
|
||||
}
|
||||
|
||||
private fun queueChunks(block: MkaBlock, stream: Int, track: MkaTrackInfo) {
|
||||
val sourceChannels = track.channels
|
||||
val totalFrames = block.data.size / (2 * sourceChannels)
|
||||
if (totalFrames <= 0) return
|
||||
// The mixed microphone is one channel however a foreign file stored it.
|
||||
val channels = if (stream == MicStream.MONO) 1 else sourceChannels
|
||||
val meta = pendingMeta[stream]?.takeIf { it.timeMs == block.timeMs }
|
||||
pendingMeta[stream] = null
|
||||
var start = 0
|
||||
while (start < totalFrames) {
|
||||
val frames = minOf(MAX_CHUNK_FRAMES, totalFrames - start)
|
||||
val pcm = ShortArray(frames * channels)
|
||||
for (frame in 0 until frames) {
|
||||
val at = (start + frame) * sourceChannels * 2
|
||||
if (channels == sourceChannels) {
|
||||
for (channel in 0 until channels) {
|
||||
pcm[frame * channels + channel] = AudioSegmentation.pcm16le(block.data[at + channel * 2], block.data[at + channel * 2 + 1])
|
||||
}
|
||||
} else {
|
||||
var sum = 0
|
||||
for (channel in 0 until sourceChannels) sum += AudioSegmentation.pcm16le(block.data[at + channel * 2], block.data[at + channel * 2 + 1])
|
||||
pcm[frame] = (sum / sourceChannels).toShort()
|
||||
}
|
||||
}
|
||||
val chunk = MicChunk(stream, channels, track.sampleRate, pcm, frames, block.timeMs + start * 1000L / track.sampleRate)
|
||||
if (meta != null && start == 0) applyMeta(chunk, meta.fields)
|
||||
queued.addLast(chunk)
|
||||
start += frames
|
||||
}
|
||||
}
|
||||
|
||||
private fun applyMeta(chunk: MicChunk, fields: Map<String, Any?>) {
|
||||
chunk.sequence = fields["seq"].asInt() ?: -1
|
||||
chunk.deviceTickMs = fields["tick"].asLong() ?: -1L
|
||||
val angle = fields["angle"].asInt()
|
||||
if (angle != null) {
|
||||
chunk.hasDirection = true
|
||||
chunk.angleDegrees = angle
|
||||
chunk.ssr = fields["ssr"].asInt() ?: 0
|
||||
}
|
||||
chunk.truncated = fields["trunc"].asBoolean() ?: false
|
||||
chunk.lostFramesBefore = fields["lost"].asInt() ?: 0
|
||||
}
|
||||
}
|
||||
+393
-46
@@ -35,6 +35,12 @@ interface VoiceCaptureHost {
|
||||
|
||||
/** The speaker-embedding model at [modelPath] (may be shared/cached across sessions). */
|
||||
fun speakerEmbedder(modelPath: String): SpeakerEmbedder
|
||||
|
||||
/** The glasses' per-temple microphone channel, or null when there are no glasses that have one. */
|
||||
fun micArrayLink(): MicArrayLink? = null
|
||||
|
||||
/** Opens a recording to replay; throws when it cannot be read. */
|
||||
fun openRecording(path: String): SeekableInput = openFileInput(path)
|
||||
}
|
||||
|
||||
/**
|
||||
@@ -44,6 +50,12 @@ interface VoiceCaptureHost {
|
||||
* REPLACE-semantics partials, and speaker verification at session end. The session owns its
|
||||
* worker thread (started through the shared [startThread]); listener callbacks go through the
|
||||
* host's [CallbackDispatcher] so the adapter decides the delivery thread.
|
||||
*
|
||||
* Audio comes from one of four sources, chosen at start(): the glasses' stock mixed stream, the
|
||||
* phone's microphone, the glasses' per-temple microphones ([setMicArray]), or a recording
|
||||
* played back in real time ([setReplay]). Each delivers [MicChunk]s; a live capture's chunks go
|
||||
* to the recorder ([setRecorder]) as they came, before any processing, which is what lets a
|
||||
* replay of the recording run the same processing again.
|
||||
*/
|
||||
class VoiceCaptureSession(
|
||||
private val host: VoiceCaptureHost,
|
||||
@@ -95,6 +107,19 @@ class VoiceCaptureSession(
|
||||
const val DEFAULT_VERIFY_THRESHOLD = 0.8f
|
||||
/** Samples per G2 packet; the decode buffer the worker reuses. */
|
||||
const val PACKET_SAMPLES = Lc3PacketFramer.SAMPLES_PER_PACKET
|
||||
// Mic array upkeep: the temples stop streaming unless their lease is renewed, and a
|
||||
// temple that goes quiet (a lost CONFIGURE, or it restarted) is configured again.
|
||||
const val ARRAY_CHECK_MS = 1000L
|
||||
const val ARRAY_RENEW_MS = MicArrayProtocol.LEASE_MS / 3
|
||||
const val ARRAY_TEMPLE_SILENT_MS = 5000L
|
||||
const val ARRAY_WAITING_MS = 2000L
|
||||
const val ARRAY_WAITING_STATUS = "Waiting for the glasses microphones..."
|
||||
// A replay skips stretches of the recording with no audio longer than this
|
||||
// (a recording made across a disconnect) instead of sitting through them.
|
||||
const val REPLAY_MAX_GAP_MS = 2000L
|
||||
const val REPLAY_POSITION_INTERVAL_MS = 1000L
|
||||
const val REPLAY_WAIT_SLICE_MS = 20L
|
||||
const val REPLAY_FINISHED_STATUS = "Replay finished."
|
||||
|
||||
/** Joins transcript pieces, attaching leading punctuation to the previous text. */
|
||||
@JvmStatic
|
||||
@@ -120,6 +145,8 @@ class VoiceCaptureSession(
|
||||
CLOUD, // decode locally, emit PCM for a cloud recognizer on the TS side
|
||||
}
|
||||
|
||||
private enum class Source { G2, PHONE, ARRAY, REPLAY }
|
||||
|
||||
private class AudioPacket(val data: ByteArray, arm: String?, val arrivalMs: Long) {
|
||||
val arm: String = arm ?: "?"
|
||||
}
|
||||
@@ -137,11 +164,33 @@ class VoiceCaptureSession(
|
||||
@Volatile
|
||||
private var workerThreadName: String? = null
|
||||
// Capture from the phone's own microphone instead of the G2 over BLE
|
||||
// (preview-only mode). Latched into activePhoneMic at start() (under `lock`)
|
||||
// (preview-only mode). Latched into activeSource at start() (under `lock`)
|
||||
// so a mid-session setter call can't switch pipelines underneath the worker.
|
||||
@Volatile
|
||||
private var usePhoneMic = false
|
||||
private var activePhoneMic = false
|
||||
// What the next capture reads (see start()); latched into activeSource and its
|
||||
// companions under `lock`, like usePhoneMic.
|
||||
@Volatile
|
||||
private var micArrayConfigure: ByteArray? = null
|
||||
@Volatile
|
||||
private var replayPath: String? = null
|
||||
@Volatile
|
||||
private var replayStartMs = 0L
|
||||
private var activeSource = Source.G2
|
||||
private var activeConfigure: ByteArray? = null
|
||||
private var activeReplayPath: String? = null
|
||||
private var activeReplayStartMs = 0L
|
||||
// Replay controls, picked up by the worker between chunks.
|
||||
@Volatile
|
||||
private var replaySeekMs = -1L
|
||||
@Volatile
|
||||
private var replayPaused = false
|
||||
@Volatile
|
||||
private var recorder: MicRecordingWriter? = null
|
||||
@Volatile
|
||||
private var mixer: MicArrayMixer = SingleTempleMixer()
|
||||
private var lastMissingPackets = 0L
|
||||
private var arrayRateWarned = false
|
||||
private var mode = Mode.CLOUD
|
||||
@Volatile
|
||||
private var onboardModelKind = VoiceModelKind.MOONSHINE
|
||||
@@ -193,6 +242,47 @@ class VoiceCaptureSession(
|
||||
this.usePhoneMic = usePhoneMic
|
||||
}
|
||||
|
||||
/**
|
||||
* Capture the next glasses session from the temples' own microphones instead of the stock
|
||||
* mixed stream: [configureRecord] is the mic_control CONFIGURE record sent to both temples
|
||||
* (it decides how many microphones each one streams). Null goes back to the mixed stream.
|
||||
*/
|
||||
fun setMicArray(configureRecord: ByteArray?) {
|
||||
this.micArrayConfigure = if (MicArrayProtocol.isConfigure(configureRecord)) configureRecord else null
|
||||
}
|
||||
|
||||
/** How the temples' chunks become the mono feed that is transcribed. */
|
||||
fun setMicArrayMixer(mixer: MicArrayMixer) {
|
||||
this.mixer = mixer
|
||||
}
|
||||
|
||||
/**
|
||||
* Have the next capture play the recording at [path] from [startMs] in real time, in place
|
||||
* of any microphone; null goes back to the microphones. A replay needs no glasses.
|
||||
*/
|
||||
fun setReplay(path: String?, startMs: Long) {
|
||||
this.replayPath = path
|
||||
this.replayStartMs = maxOf(0L, startMs)
|
||||
}
|
||||
|
||||
/** Continue the running replay from [positionMs] into its recording. */
|
||||
fun seekReplay(positionMs: Long) {
|
||||
this.replaySeekMs = maxOf(0L, positionMs)
|
||||
}
|
||||
|
||||
/** Hold the running replay where it is, or let it go on. A new capture starts unpaused. */
|
||||
fun setReplayPaused(paused: Boolean) {
|
||||
this.replayPaused = paused
|
||||
}
|
||||
|
||||
/**
|
||||
* Where live captures' chunks are recorded as they arrive, or null. Takes effect at once,
|
||||
* also mid-capture; the caller finishes the recording. A replay is never recorded.
|
||||
*/
|
||||
fun setRecorder(recorder: MicRecordingWriter?) {
|
||||
this.recorder = recorder
|
||||
}
|
||||
|
||||
fun setOnboardModelKind(kind: String?) {
|
||||
this.onboardModelKind = parseModelKind(kind)
|
||||
}
|
||||
@@ -258,13 +348,26 @@ class VoiceCaptureSession(
|
||||
emitStopped(captureId)
|
||||
return
|
||||
}
|
||||
if (!usePhoneMic && !host.isG2SessionReady()) {
|
||||
val path = replayPath
|
||||
val configure = micArrayConfigure
|
||||
val source = when {
|
||||
path != null -> Source.REPLAY
|
||||
usePhoneMic -> Source.PHONE
|
||||
configure != null -> Source.ARRAY
|
||||
else -> Source.G2
|
||||
}
|
||||
if ((source == Source.G2 || source == Source.ARRAY) && !host.isG2SessionReady()) {
|
||||
emitStatus("Voice control needs an active G2 connection.")
|
||||
emitStopped(captureId)
|
||||
return
|
||||
}
|
||||
mode = parseMode(requestedMode)
|
||||
activePhoneMic = usePhoneMic
|
||||
activeSource = source
|
||||
activeConfigure = configure
|
||||
activeReplayPath = path
|
||||
activeReplayStartMs = replayStartMs
|
||||
replaySeekMs = -1L
|
||||
replayPaused = false
|
||||
started = true
|
||||
audioStarted = false
|
||||
val finished = Latch(1, platform)
|
||||
@@ -281,21 +384,22 @@ class VoiceCaptureSession(
|
||||
*/
|
||||
fun isCapturing(): Boolean {
|
||||
var audioUp = false
|
||||
var phoneMic = false
|
||||
var source = Source.G2
|
||||
lock.withLock {
|
||||
if (!started) {
|
||||
return false
|
||||
}
|
||||
audioUp = audioStarted
|
||||
phoneMic = activePhoneMic
|
||||
source = activeSource
|
||||
}
|
||||
if (!audioUp) {
|
||||
// The worker is still bringing the mic up; report it as running so
|
||||
// a concurrent request shares it instead of restarting it.
|
||||
return true
|
||||
}
|
||||
if (phoneMic) {
|
||||
// The phone mic has no session to lose the enable to; it runs until stop().
|
||||
if (source == Source.PHONE || source == Source.REPLAY) {
|
||||
// Neither has a glasses session to lose the enable to: the phone mic runs
|
||||
// until stop(), a replay until its recording ends.
|
||||
return true
|
||||
}
|
||||
return host.isG2AudioCaptureActive()
|
||||
@@ -317,7 +421,7 @@ class VoiceCaptureSession(
|
||||
finished = workerFinished
|
||||
workerName = workerThreadName
|
||||
}
|
||||
stopG2Audio()
|
||||
stopSource()
|
||||
audioQueue.wake()
|
||||
if (finished != null && (workerName == null || currentThreadName() != workerName)) {
|
||||
finished.await(1500)
|
||||
@@ -356,28 +460,62 @@ class VoiceCaptureSession(
|
||||
val verifying = verifySpeakerModelPath != null && verifyWearerEmbedding != null
|
||||
verifyBuffer = if (verifying) ShortArray(VERIFY_MAX_SAMPLES) else null
|
||||
verifyCount = 0
|
||||
if (activePhoneMic) {
|
||||
val record = host.openPhoneMic()
|
||||
if (record == null) {
|
||||
emitStatus("Could not start the phone microphone.")
|
||||
return
|
||||
mixer.reset()
|
||||
lastMissingPackets = 0
|
||||
val listening = if (currentMode == Mode.CLOUD) "Listening (cloud)..." else "Listening..."
|
||||
when (activeSource) {
|
||||
Source.PHONE -> {
|
||||
val record = host.openPhoneMic()
|
||||
if (record == null) {
|
||||
emitStatus("Could not start the phone microphone.")
|
||||
return
|
||||
}
|
||||
lock.withLock { audioStarted = true }
|
||||
emitStatus(listening)
|
||||
try {
|
||||
processPhoneAudio(record)
|
||||
} finally {
|
||||
record.close()
|
||||
}
|
||||
}
|
||||
lock.withLock { audioStarted = true }
|
||||
emitStatus(if (currentMode == Mode.CLOUD) "Listening (cloud)..." else "Listening...")
|
||||
try {
|
||||
processPhoneAudio(record)
|
||||
} finally {
|
||||
record.close()
|
||||
Source.REPLAY -> {
|
||||
val reader = try {
|
||||
MicRecordingReader(host.openRecording(activeReplayPath!!))
|
||||
} catch (error: Throwable) {
|
||||
emitStatus("Could not open the recording: " + error.message)
|
||||
emitReplay("error", 0, 0)
|
||||
return
|
||||
}
|
||||
try {
|
||||
lock.withLock { audioStarted = true }
|
||||
logInfo("replaying " + reader.description + ", " + formatFixed(reader.durationMs / 1000.0, 1) + " s")
|
||||
emitStatus(listening)
|
||||
processReplay(reader)
|
||||
} finally {
|
||||
reader.close()
|
||||
}
|
||||
}
|
||||
} else {
|
||||
packetDecoder = host.createPacketDecoder()
|
||||
if (!startG2Audio()) {
|
||||
emitStatus("Could not start G2 microphone input.")
|
||||
return
|
||||
Source.ARRAY -> {
|
||||
val link = host.micArrayLink()
|
||||
val configure = activeConfigure
|
||||
if (link == null || configure == null || !startArray(link, configure)) {
|
||||
emitStatus("Could not start the glasses microphone array.")
|
||||
return
|
||||
}
|
||||
lock.withLock { audioStarted = true }
|
||||
emitStatus(listening)
|
||||
processArray(link, configure, listening)
|
||||
}
|
||||
Source.G2 -> {
|
||||
packetDecoder = host.createPacketDecoder()
|
||||
if (!startG2Audio()) {
|
||||
emitStatus("Could not start G2 microphone input.")
|
||||
return
|
||||
}
|
||||
lock.withLock { audioStarted = true }
|
||||
emitStatus(listening)
|
||||
processG2Audio()
|
||||
}
|
||||
lock.withLock { audioStarted = true }
|
||||
emitStatus(if (currentMode == Mode.CLOUD) "Listening (cloud)..." else "Listening...")
|
||||
processG2Audio()
|
||||
}
|
||||
// Verification result must precede the final transcript so the
|
||||
// TS bridge can suppress a non-wearer command before it is acted
|
||||
@@ -392,7 +530,7 @@ class VoiceCaptureSession(
|
||||
logError("Voice control failed", error)
|
||||
emitStatus("Voice control failed: " + error.message)
|
||||
} finally {
|
||||
stopG2Audio()
|
||||
stopSource()
|
||||
releaseRecognizer()
|
||||
releaseDecoder()
|
||||
lock.withLock {
|
||||
@@ -432,23 +570,214 @@ class VoiceCaptureSession(
|
||||
continue
|
||||
}
|
||||
decodedSamples += count.toLong()
|
||||
// The beam filter drops packets whose firmware direction-of-arrival
|
||||
// falls outside the listening wedge (isolating the aimed talker for
|
||||
// every consumer, recognition included); the spectral noise
|
||||
// suppressor then cleans what remains before it reaches the
|
||||
// recognizer, cloud PCM, endpointing, or speaker verification.
|
||||
val angleDegrees = currentDecoder.lastAngleDegrees
|
||||
val ssr = currentDecoder.lastSsr
|
||||
if (beamFilterEnabled && ssr > 0 && !withinBeam(angleDegrees)) {
|
||||
emitFrameMeta(angleDegrees, ssr)
|
||||
maybeEmitAudioStats(false)
|
||||
continue
|
||||
}
|
||||
processPcmChunk(pcm, count, angleDegrees, ssr, currentDecoder.hasFrameMeta)
|
||||
val chunk = MicChunk(MicStream.MONO, 1, SAMPLE_RATE, pcm, count, packet.arrivalMs)
|
||||
chunk.hasDirection = currentDecoder.hasFrameMeta
|
||||
chunk.angleDegrees = currentDecoder.lastAngleDegrees
|
||||
chunk.ssr = currentDecoder.lastSsr
|
||||
// Packets the decoder found missing ahead of this one: a gap in a recording.
|
||||
val missing = currentDecoder.missingPackets
|
||||
chunk.lostFramesBefore = ((missing - lastMissingPackets) * currentDecoder.maxSamplesPerPacket).toInt()
|
||||
lastMissingPackets = missing
|
||||
handleMonoChunk(chunk)
|
||||
maybeEmitAudioStats(false)
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* A mixed-microphone chunk from any source: recorded as it came, then on to processing
|
||||
* unless the beam filter drops it. The filter drops chunks whose firmware
|
||||
* direction-of-arrival falls outside the listening wedge (isolating the aimed talker for
|
||||
* every consumer, recognition included); the spectral noise suppressor then cleans what
|
||||
* remains before it reaches the recognizer, cloud PCM, endpointing, or speaker verification.
|
||||
*/
|
||||
private fun handleMonoChunk(chunk: MicChunk) {
|
||||
recordChunk(chunk)
|
||||
if (beamFilterEnabled && chunk.ssr > 0 && !withinBeam(chunk.angleDegrees)) {
|
||||
emitFrameMeta(chunk.angleDegrees, chunk.ssr)
|
||||
return
|
||||
}
|
||||
processPcmChunk(chunk.pcm, chunk.frames, chunk.angleDegrees, chunk.ssr, chunk.hasDirection)
|
||||
}
|
||||
|
||||
/**
|
||||
* A chunk of one temple's own microphones: recorded as it came, then mixed into the mono
|
||||
* feed, which is recorded too (as what was transcribed) and processed like a mixed chunk.
|
||||
* The beam filter does not apply: a temple's bearing is along its own pair, not the
|
||||
* head-frame direction the filter is aimed in.
|
||||
*/
|
||||
private fun handleArrayChunk(chunk: MicChunk) {
|
||||
recordChunk(chunk)
|
||||
val feed = mixer.mix(chunk) ?: return
|
||||
recordChunk(feed)
|
||||
if (feed.sampleRate != SAMPLE_RATE || feed.channels != 1) {
|
||||
if (!arrayRateWarned) logWarn("mic array feed is ${feed.channels}ch ${feed.sampleRate} Hz; recognition needs mono $SAMPLE_RATE Hz")
|
||||
arrayRateWarned = true
|
||||
return
|
||||
}
|
||||
processPcmChunk(feed.pcm, feed.frames, feed.angleDegrees, feed.ssr, feed.hasDirection)
|
||||
}
|
||||
|
||||
private fun recordChunk(chunk: MicChunk) {
|
||||
if (activeSource != Source.REPLAY) recorder?.write(chunk)
|
||||
}
|
||||
|
||||
private fun handleChunk(chunk: MicChunk) {
|
||||
if (chunk.stream == MicStream.MONO) handleMonoChunk(chunk) else handleArrayChunk(chunk)
|
||||
}
|
||||
|
||||
private fun startArray(link: MicArrayLink, configure: ByteArray): Boolean {
|
||||
resetAudioStats()
|
||||
audioQueue.clear()
|
||||
if (!startArrayForwarding(link)) return false
|
||||
return link.sendControl(configure, "configure", true, true)
|
||||
}
|
||||
|
||||
private fun startArrayForwarding(link: MicArrayLink): Boolean =
|
||||
link.startForwarding(object : FaceclawAudioPacketListener {
|
||||
override fun onAudioPacket(data: ByteArray?, arm: String?, arrivalMs: Long) {
|
||||
queueAudioPacket(data, arm, arrivalMs)
|
||||
}
|
||||
})
|
||||
|
||||
/**
|
||||
* Mic array capture loop: each temple's 'SM' frames become chunks of that temple's stream.
|
||||
* Between frames it keeps the streams alive: renews the temples' lease, configures a temple
|
||||
* again when it has gone quiet, and takes forwarding back when the glasses session dropped it.
|
||||
*/
|
||||
private fun processArray(link: MicArrayLink, configure: ByteArray, listening: String) {
|
||||
val startedAt = platform.elapsedRealtimeMs()
|
||||
var checkedAt = startedAt
|
||||
var renewedAt = startedAt
|
||||
// Left, right.
|
||||
val armedAt = longArrayOf(startedAt, startedAt)
|
||||
val frameAt = longArrayOf(startedAt, startedAt)
|
||||
var waiting = false
|
||||
while (started) {
|
||||
val packet = takeAudioPacket()
|
||||
val now = platform.elapsedRealtimeMs()
|
||||
if (packet != null) {
|
||||
val side = if (packet.arm == "L") 0 else if (packet.arm == "R") 1 else -1
|
||||
val frame = if (side >= 0) MicArrayProtocol.decodeFrame(packet.data) else null
|
||||
val pcm = if (frame != null) MicArrayProtocol.pcm16(packet.data, frame) else null
|
||||
// Anything else is a stock-format packet from before arming took effect.
|
||||
if (frame != null && pcm != null) {
|
||||
frameAt[side] = now
|
||||
val frames = pcm.size / frame.channels
|
||||
decodedSamples += frames.toLong()
|
||||
val stream = if (side == 0) MicStream.LEFT else MicStream.RIGHT
|
||||
val chunk = MicChunk(stream, frame.channels, frame.sampleRate, pcm, frames, packet.arrivalMs)
|
||||
chunk.hasDirection = frame.angleDegrees != 0 || frame.ssr != 0
|
||||
chunk.angleDegrees = frame.angleDegrees
|
||||
chunk.ssr = frame.ssr
|
||||
chunk.sequence = frame.sequence
|
||||
chunk.deviceTickMs = frame.deviceTickMs
|
||||
chunk.truncated = frame.truncated
|
||||
handleArrayChunk(chunk)
|
||||
}
|
||||
}
|
||||
if (now - checkedAt >= ARRAY_CHECK_MS) {
|
||||
checkedAt = now
|
||||
if (!link.isForwarding()) {
|
||||
if (host.isG2SessionReady() && startArrayForwarding(link) && link.sendControl(configure, "configure", true, true)) {
|
||||
logInfo("mic array forwarding re-established")
|
||||
armedAt.fill(now)
|
||||
renewedAt = now
|
||||
}
|
||||
} else {
|
||||
if (now - renewedAt >= ARRAY_RENEW_MS) {
|
||||
link.sendControl(MicArrayProtocol.control(MicArrayProtocol.OP_RENEW), "renew", true, true)
|
||||
renewedAt = now
|
||||
}
|
||||
for (side in 0..1) {
|
||||
if (now - maxOf(frameAt[side], armedAt[side]) < ARRAY_TEMPLE_SILENT_MS) continue
|
||||
logWarn("mic array: " + (if (side == 0) "left" else "right") + " temple silent; configuring it again")
|
||||
link.sendControl(configure, "configure-" + (if (side == 0) "left" else "right"), side == 1, side == 0)
|
||||
armedAt[side] = now
|
||||
}
|
||||
}
|
||||
val silent = now - maxOf(frameAt[0], frameAt[1]) >= ARRAY_WAITING_MS
|
||||
if (silent != waiting) {
|
||||
waiting = silent
|
||||
emitStatus(if (silent) ARRAY_WAITING_STATUS else listening)
|
||||
}
|
||||
}
|
||||
maybeEmitAudioStats(false)
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Replay loop: hands the recording's chunks to the pipeline at the pace they were captured,
|
||||
* so recognition (its partials, a cloud provider's stream) sees what a live capture would.
|
||||
* Jumps and pauses asked for from outside take effect between chunks.
|
||||
*/
|
||||
private fun processReplay(reader: MicRecordingReader) {
|
||||
val duration = reader.durationMs
|
||||
var position = minOf(activeReplayStartMs, duration)
|
||||
reader.seekTo(position)
|
||||
emitReplay("start", position, duration)
|
||||
var reportedAt = position
|
||||
// Wall-clock time at which file time `anchorFile` plays; unset after a jump, pause or gap.
|
||||
var anchorWall = -1L
|
||||
var anchorFile = 0L
|
||||
var pending: MicChunk? = null
|
||||
while (started) {
|
||||
val seek = replaySeekMs
|
||||
if (seek >= 0) {
|
||||
replaySeekMs = -1L
|
||||
// What played so far is a finished utterance; the rest starts clean.
|
||||
finishReplayedUtterance()
|
||||
position = minOf(seek, duration)
|
||||
reader.seekTo(position)
|
||||
pending = null
|
||||
anchorWall = -1L
|
||||
reportedAt = position
|
||||
emitReplay("seek", position, duration)
|
||||
}
|
||||
if (replayPaused) {
|
||||
anchorWall = -1L
|
||||
sleepMs(REPLAY_WAIT_SLICE_MS)
|
||||
continue
|
||||
}
|
||||
val chunk = pending ?: reader.next()
|
||||
if (chunk == null) {
|
||||
emitReplay("end", duration, duration)
|
||||
emitStatus(REPLAY_FINISHED_STATUS)
|
||||
return
|
||||
}
|
||||
pending = chunk
|
||||
val now = platform.elapsedRealtimeMs()
|
||||
if (anchorWall < 0 || chunk.timeMs - position > REPLAY_MAX_GAP_MS) {
|
||||
anchorWall = now
|
||||
anchorFile = chunk.timeMs
|
||||
}
|
||||
val wait = anchorWall + (chunk.timeMs - anchorFile) - now
|
||||
if (wait > 0) {
|
||||
// Short slices, so a stop, jump or pause is noticed promptly.
|
||||
sleepMs(minOf(wait, REPLAY_WAIT_SLICE_MS))
|
||||
continue
|
||||
}
|
||||
pending = null
|
||||
position = chunk.timeMs
|
||||
decodedSamples += chunk.frames.toLong()
|
||||
handleChunk(chunk)
|
||||
if (position - reportedAt >= REPLAY_POSITION_INTERVAL_MS) {
|
||||
reportedAt = position
|
||||
emitReplay("position", position, duration)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/** Ends the utterance a replay was in the middle of when it jumped, so the audio on both sides isn't joined. */
|
||||
private fun finishReplayedUtterance() {
|
||||
mixer.reset()
|
||||
endpointDetector.reset()
|
||||
suppressor?.reset()
|
||||
if (mode != Mode.ONBOARD) return
|
||||
decodeTranscript(true)
|
||||
resetTranscriptState()
|
||||
lastTranscript = ""
|
||||
}
|
||||
|
||||
/**
|
||||
* Per-chunk processing shared by the G2 and phone-mic paths, downstream of
|
||||
* decode and the beam filter. hasFrameMeta is false for the phone mic,
|
||||
@@ -495,7 +824,7 @@ class VoiceCaptureSession(
|
||||
continue
|
||||
}
|
||||
decodedSamples += read.toLong()
|
||||
processPcmChunk(pcm, read, 0, 0, false)
|
||||
handleMonoChunk(MicChunk(MicStream.MONO, 1, SAMPLE_RATE, pcm, read, platform.elapsedRealtimeMs()))
|
||||
}
|
||||
}
|
||||
|
||||
@@ -681,6 +1010,11 @@ class VoiceCaptureSession(
|
||||
host.dispatcher.post { currentListener.onFrameMeta(angleDegrees, ssr) }
|
||||
}
|
||||
|
||||
private fun emitReplay(event: String, positionMs: Long, durationMs: Long) {
|
||||
val currentListener = listener ?: return
|
||||
host.dispatcher.post { currentListener.onReplay(event, positionMs, durationMs) }
|
||||
}
|
||||
|
||||
private fun emitSpeechEnd() {
|
||||
val currentListener = listener ?: return
|
||||
host.dispatcher.post { currentListener.onSpeechEnd() }
|
||||
@@ -691,8 +1025,21 @@ class VoiceCaptureSession(
|
||||
host.dispatcher.post { currentListener.onStopped(captureId) }
|
||||
}
|
||||
|
||||
private fun stopG2Audio() {
|
||||
host.g2AudioSource()?.stop()
|
||||
/**
|
||||
* Release what the running capture's source holds on the glasses. Runs on stop() and again
|
||||
* as the worker winds down (it may have started the source after stop() looked), so
|
||||
* everything here is safe to repeat.
|
||||
*/
|
||||
private fun stopSource() {
|
||||
val source = lock.withLock { activeSource }
|
||||
when (source) {
|
||||
Source.G2 -> host.g2AudioSource()?.stop()
|
||||
Source.ARRAY -> host.micArrayLink()?.let { link ->
|
||||
link.sendControl(MicArrayProtocol.control(MicArrayProtocol.OP_STOP), "stop", true, true)
|
||||
link.stopForwarding()
|
||||
}
|
||||
Source.PHONE, Source.REPLAY -> {}
|
||||
}
|
||||
maybeEmitAudioStats(true)
|
||||
}
|
||||
|
||||
@@ -711,7 +1058,7 @@ class VoiceCaptureSession(
|
||||
if (!started || data == null) {
|
||||
return
|
||||
}
|
||||
if ("L" != arm) {
|
||||
if ("L" != arm && activeSource != Source.ARRAY) {
|
||||
wrongArmPackets++
|
||||
}
|
||||
audioQueue.put(AudioPacket(data, arm, arrivalMs)) {
|
||||
|
||||
@@ -67,6 +67,24 @@ interface PacketAudioSource {
|
||||
fun isActive(): Boolean
|
||||
}
|
||||
|
||||
/**
|
||||
* The custom firmware's mic_control channel on a glasses session: each temple streams its own
|
||||
* microphones as 'SM' frames (see [MicArrayProtocol]) on the characteristic the stock mixed
|
||||
* stream uses, while it holds a configuration that the phone keeps renewing.
|
||||
*/
|
||||
interface MicArrayLink {
|
||||
/** Deliver the audio characteristic's packets to [listener]; false when no session is up. */
|
||||
fun startForwarding(listener: FaceclawAudioPacketListener): Boolean
|
||||
|
||||
fun stopForwarding()
|
||||
|
||||
/** Whether forwarding still holds: it lives in the glasses session, which can drop it. */
|
||||
fun isForwarding(): Boolean
|
||||
|
||||
/** Queue a control record for the chosen temples; false when nothing was queued. */
|
||||
fun sendControl(record: ByteArray, label: String, rightTemple: Boolean, leftTemple: Boolean): Boolean
|
||||
}
|
||||
|
||||
/** The phone's own microphone delivering 16 kHz mono PCM16 in blocking reads. */
|
||||
interface PcmAudioSource {
|
||||
/** Blocking read into [pcm]; returns samples read, 0 to retry, or a negative error. */
|
||||
|
||||
+9
@@ -41,4 +41,13 @@ interface FaceclawVoiceControllerListener {
|
||||
* speaker model could not be loaded.
|
||||
*/
|
||||
fun onSpeakerVerified(isWearer: Boolean, similarity: Float): Unit
|
||||
|
||||
/**
|
||||
* Progress of a capture that replays a recording (see VoiceCaptureSession.setReplay):
|
||||
* "start" once the file is open, "position" about once per second of it played, "seek" after
|
||||
* a jump took effect (the transcript of what played before it has been emitted as final),
|
||||
* "end" when the file ran out, and "error" when it could not be opened (the status says
|
||||
* why). Positions are milliseconds into the file.
|
||||
*/
|
||||
fun onReplay(event: String?, positionMs: Long, durationMs: Long): Unit
|
||||
}
|
||||
@@ -0,0 +1,37 @@
|
||||
package com.faceclaw.app
|
||||
|
||||
import platform.Foundation.NSData
|
||||
|
||||
/**
|
||||
* Records the iOS voice bridge's audio to a Matroska audio file (.mka), the format Android's
|
||||
* captures are recorded in (see [MicRecordingWriter]). The bridge has one stream to record:
|
||||
* 16 kHz mono PCM16, as its recognizer hands it to TypeScript.
|
||||
*/
|
||||
class IosMicRecorder internal constructor(private val writer: MicRecordingWriter) {
|
||||
/** The recording stopped early: the file could not be written. */
|
||||
val failed: Boolean get() = writer.failed
|
||||
|
||||
/** Appends a chunk of 16 kHz mono PCM16 (little-endian), as having arrived now. */
|
||||
fun append(pcm16le: NSData) {
|
||||
val bytes = pcm16le.byteArray()
|
||||
val samples = ShortArray(bytes.size / 2) { AudioSegmentation.pcm16le(bytes[it * 2], bytes[it * 2 + 1]) }
|
||||
writer.write(MicChunk(MicStream.MONO, 1, 16000, samples, samples.size, IosProtocolPlatform.elapsedRealtimeMs()))
|
||||
}
|
||||
|
||||
/** Completes the file. Returns the recording's length in ms. */
|
||||
fun finish(): Long = writer.finish()
|
||||
}
|
||||
|
||||
object IosMicRecordings {
|
||||
/**
|
||||
* A recorder writing a new file at [path], or null when the file cannot be created.
|
||||
* [configJson] describes the capture and is stored in the file.
|
||||
*/
|
||||
fun open(path: String, configJson: String?): IosMicRecorder? =
|
||||
try {
|
||||
IosMicRecorder(MicRecordingWriter(openFileOutput(path), configJson, currentTimeMillis(), IosProtocolPlatform))
|
||||
} catch (error: Throwable) {
|
||||
PlatformLog.w("FaceclawMicRecording", "could not create $path: ${error.message}")
|
||||
null
|
||||
}
|
||||
}
|
||||
@@ -114,3 +114,54 @@ actual fun fileLength(path: String): Long = memScoped {
|
||||
actual fun deleteFile(path: String): Boolean = unlink(path) == 0
|
||||
|
||||
actual fun renameFile(from: String, to: String): Boolean = rename(from, to) == 0
|
||||
|
||||
actual fun openFileOutput(path: String): SeekableOutput {
|
||||
val file = fopen(path, "wb+") ?: error("Cannot write $path")
|
||||
return object : SeekableOutput {
|
||||
override var size = 0L
|
||||
private set
|
||||
|
||||
private fun put(bytes: ByteArray, offset: Int, length: Int) {
|
||||
if (length == 0) return
|
||||
bytes.usePinned {
|
||||
check(fwrite(it.addressOf(offset), 1u, length.convert(), file) == length.convert<platform.posix.size_t>()) { "Short write to $path" }
|
||||
}
|
||||
}
|
||||
|
||||
override fun write(bytes: ByteArray, offset: Int, length: Int) {
|
||||
put(bytes, offset, length)
|
||||
size += length
|
||||
// Each write is a complete element; hand it to the OS so a crash keeps it.
|
||||
platform.posix.fflush(file)
|
||||
}
|
||||
|
||||
override fun writeAt(position: Long, bytes: ByteArray) {
|
||||
require(position >= 0 && position + bytes.size <= size)
|
||||
check(platform.posix.fseek(file, position, platform.posix.SEEK_SET) == 0)
|
||||
put(bytes, 0, bytes.size)
|
||||
check(platform.posix.fseek(file, size, platform.posix.SEEK_SET) == 0)
|
||||
platform.posix.fflush(file)
|
||||
}
|
||||
|
||||
override fun close() {
|
||||
fclose(file)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
actual fun openFileInput(path: String): SeekableInput {
|
||||
val file = fopen(path, "rb") ?: error("Cannot read $path")
|
||||
val fileSize = fileLength(path)
|
||||
return object : SeekableInput {
|
||||
override val length = fileSize
|
||||
|
||||
override fun readAt(position: Long, buffer: ByteArray, offset: Int, length: Int): Int {
|
||||
if (length == 0 || platform.posix.fseek(file, position, platform.posix.SEEK_SET) != 0) return 0
|
||||
return buffer.usePinned { platform.posix.fread(it.addressOf(offset), 1u, length.convert(), file).toInt() }
|
||||
}
|
||||
|
||||
override fun close() {
|
||||
fclose(file)
|
||||
}
|
||||
}
|
||||
}
|
||||
Reference in new issue
Block a user