forked from Blendman974/faceclaw
Merge multilingual Whisper transcription
# Conflicts: # app/g2/dashboard-controller.ts
This commit is contained in:
21 files changed
+375
-39
No files matched your search
@@ -25,18 +25,21 @@ internal object AndroidSpeechEngines {
|
||||
private const val FEATURE_DIM = 80
|
||||
|
||||
/** Sherpa-onnx configuration for the on-device model files in [modelDir]. */
|
||||
fun recognizerConfig(modelDir: File, kind: VoiceModelKind): OfflineRecognizerConfig {
|
||||
fun recognizerConfig(
|
||||
modelDir: File, kind: VoiceModelKind,
|
||||
whisper: WhisperModelFiles = WhisperModelFiles.forLanguage(null),
|
||||
): OfflineRecognizerConfig {
|
||||
val modelConfig = OfflineModelConfig.builder()
|
||||
.setNumThreads(1)
|
||||
if (kind == VoiceModelKind.WHISPER) {
|
||||
modelConfig
|
||||
.setWhisper(OfflineWhisperModelConfig.builder()
|
||||
.setEncoder(File(modelDir, "base.en-encoder.int8.onnx").absolutePath)
|
||||
.setDecoder(File(modelDir, "base.en-decoder.int8.onnx").absolutePath)
|
||||
.setLanguage("en")
|
||||
.setEncoder(File(modelDir, whisper.encoder).absolutePath)
|
||||
.setDecoder(File(modelDir, whisper.decoder).absolutePath)
|
||||
.setLanguage(whisper.language)
|
||||
.setTask("transcribe")
|
||||
.build())
|
||||
.setTokens(File(modelDir, "base.en-tokens.txt").absolutePath)
|
||||
.setTokens(File(modelDir, whisper.tokens).absolutePath)
|
||||
} else {
|
||||
modelConfig
|
||||
.setMoonshine(OfflineMoonshineModelConfig.builder()
|
||||
|
||||
@@ -32,17 +32,6 @@ class FaceclawVoiceController(context: Context) {
|
||||
"decoder_model_merged.ort",
|
||||
"tokens.txt"
|
||||
)
|
||||
// Second on-device model: sherpa-onnx's offline Whisper backend (base.en,
|
||||
// int8-quantized -- see the model-choice note in asr-model.ts). Directory
|
||||
// shared with the TS-side download flow, same convention as ASR_MODEL_DIR.
|
||||
// Whisper re-encodes the whole buffer on every call, so the shared session
|
||||
// skips live partials for it; see VoiceCaptureSession.processRecognizer.
|
||||
private const val ASR_WHISPER_MODEL_DIR = "sherpa-onnx-whisper-base-en-int8"
|
||||
private val ASR_WHISPER_MODEL_FILES = arrayOf(
|
||||
"base.en-encoder.int8.onnx",
|
||||
"base.en-decoder.int8.onnx",
|
||||
"base.en-tokens.txt"
|
||||
)
|
||||
// 50 ms chunks match the G2 packet cadence the rest of the pipeline
|
||||
// (endpointing, transcript pacing) is tuned for.
|
||||
private const val PHONE_MIC_CHUNK_SAMPLES = SAMPLE_RATE / 20
|
||||
@@ -84,6 +73,7 @@ class FaceclawVoiceController(context: Context) {
|
||||
private val mainHandler = Handler(Looper.getMainLooper())
|
||||
@Volatile
|
||||
private var communicator: FaceclawBleCommunicator? = null
|
||||
private var whisperModel = WhisperModelFiles.forLanguage(null)
|
||||
|
||||
private val host = object : VoiceCaptureHost {
|
||||
override val dispatcher = CallbackDispatcher { action -> mainHandler.post { action() } }
|
||||
@@ -122,7 +112,7 @@ class FaceclawVoiceController(context: Context) {
|
||||
|
||||
override fun loadTranscriber(kind: VoiceModelKind): OfflineTranscriber {
|
||||
val modelDir = findAsrModelDir(kind) ?: throw IllegalStateException("voice model missing")
|
||||
return AndroidSpeechEngines.SherpaTranscriber(AndroidSpeechEngines.recognizerConfig(modelDir, kind))
|
||||
return AndroidSpeechEngines.SherpaTranscriber(AndroidSpeechEngines.recognizerConfig(modelDir, kind, whisperModel))
|
||||
}
|
||||
|
||||
override fun speakerEmbedder(modelPath: String): SpeakerEmbedder {
|
||||
@@ -168,6 +158,11 @@ class FaceclawVoiceController(context: Context) {
|
||||
session.setOnboardModelKind(kind)
|
||||
}
|
||||
|
||||
/** Language for the next on-device Whisper capture; empty enables auto-detection. */
|
||||
fun setOnboardLanguage(language: String?) {
|
||||
whisperModel = WhisperModelFiles.forLanguage(language)
|
||||
}
|
||||
|
||||
/**
|
||||
* When true, watch the decoded PCM and fire {@code onSpeechEnd} once the
|
||||
* speaker stops. Used by hands-free ("Hey Even") capture, which has no
|
||||
@@ -263,8 +258,8 @@ class FaceclawVoiceController(context: Context) {
|
||||
*/
|
||||
private fun findAsrModelDir(kind: VoiceModelKind): File? {
|
||||
val whisper = kind == VoiceModelKind.WHISPER
|
||||
val dirName = if (whisper) ASR_WHISPER_MODEL_DIR else ASR_MODEL_DIR
|
||||
val fileNames = if (whisper) ASR_WHISPER_MODEL_FILES else ASR_MODEL_FILES
|
||||
val dirName = if (whisper) whisperModel.dirName else ASR_MODEL_DIR
|
||||
val fileNames = if (whisper) whisperModel.files else ASR_MODEL_FILES.toList()
|
||||
val modelDir = File(appContext.filesDir, ASR_ROOT + File.separator + dirName)
|
||||
for (fileName in fileNames) {
|
||||
val file = File(modelDir, fileName)
|
||||
|
||||
Reference in new issue
Block a user