Add on-device Whisper as a second offline transcription option

sherpa-onnx's native library already reads an offline Whisper config
(verified against sherpa-onnx/jni/offline-recognizer.cc at v1.13.0);
only the Java wrapper was a package-private stub. This adds a public
OfflineWhisperModelConfig builder and wires base.en (int8) in as a
second on-device model next to Moonshine.

- FaceclawVoiceController: setOnboardModelKind() selects the model
  before start("onboard"); defaults to Moonshine. Whisper skips the
  700ms live-partial redecode (each call re-encodes the whole utterance
  plus ~10s of tail padding) and gates near-silent segments before
  decode.
- asr-model.ts: ASR_MODELS registry with per-model download state.
- Settings > Voice: "On-device (Moonshine)", "On-device (Whisper)",
  and "OpenAI (Whisper)" for the existing cloud provider (stored
  value unchanged); one model row per on-device model.
- Microphones live captions pass "moonshine" explicitly (unchanged).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Chris LeeandClaude Sonnet 5 committed 2026-09-14 22:46:22 -04:00
1 parent dcb071621d
commit 2d717a9cdb
9 files changed
+459 -162

No files matched your search

+4 -1
View File
@@ -730,7 +730,10 @@ class MicSession {
},
});
engine.setListener(this.captionListenerProxy);
engine.setAsrModelDir(isAsrModelReady() ? moonshineModelDir() : null);
// Live captions are a separate sherpa-onnx pipeline (FaceclawCaptionEngine)
// that only ever uses Moonshine; unaffected by the two-model on-device
// Whisper addition in native/voice-control.ts's push-to-talk path.
engine.setAsrModelDir(isAsrModelReady("moonshine") ? moonshineModelDir() : null);
engine.setSpeakerModelPath(micModelPath("speaker-embedding"));
engine.start();
this.captionEngine = engine;
+6 -3
View File
@@ -328,17 +328,20 @@ function voiceSpeakersMenu(): MenuLayer {
() => startMicModelDownload("speaker-embedding"),
),
modelItem(
// Live captions always use Moonshine (see the matching note in
// mic-session.ts); the second on-device model (Whisper) is a
// push-to-talk dictation option only, managed in Settings > Voice.
"Caption ASR model",
isAsrModelReady,
() => isAsrModelReady("moonshine"),
() => {
const state = asrModelState();
const state = asrModelState("moonshine");
if (state.status === "ready") return "ready";
if (state.status === "downloading") {
return `${Math.round((state.bytesDownloaded / state.totalBytes) * 100)}%`;
}
return "download (141 MB)";
},
startAsrModelDownload,
() => startAsrModelDownload("moonshine"),
),
modelItem(
"Re-diarization model",
+172 -95
View File
@@ -4,50 +4,107 @@ declare const com: any;
declare const java: any;
/**
* Download management for the on-device transcription model (Moonshine, via
* sherpa-onnx). The model is no longer bundled in the APK; it is fetched on
* demand into the same filesDir location that earlier releases copied the
* bundled files to, so upgraded installs that already used on-device
* transcription need no download. FaceclawVoiceController reads the files
* from this directory.
* Download management for the on-device transcription models (sherpa-onnx
* offline recognizers). Models are fetched on demand into filesDir rather
* than bundled in the APK; FaceclawVoiceController reads the files from the
* directory each model's `dirName` names.
*
* Mirrors the on-phone assistant model flow in llama.ts, except the model is
* three files rather than one; they download sequentially through
* Two models, same shape:
* - "moonshine": the original on-device option (three files). The model is
* no longer bundled in the APK; it is fetched into the same filesDir
* location that earlier releases copied the bundled files to, so
* upgraded installs that already used on-device transcription need no
* download.
* - "whisper-base-en": sherpa-onnx's offline Whisper backend, base.en,
* int8-quantized. Picked over tiny.en for materially better accuracy
* (Whisper's own tiny/base WER gap is real and well documented) on the
* reasoning that FaceclawVoiceController.java no longer re-decodes a
* Whisper segment on every live-partial tick — see its ASR_WHISPER_*
* comment — so base.en's extra compute is a one-time cost per utterance,
* not a cost paid ~1.4x/second while the user is speaking. tiny.en is an
* easy swap (this file's numbers only) if base.en proves too slow on a
* real phone; nobody has measured that yet. Files are csukuangfj's own
* (the sherpa-onnx maintainer's) HF mirror of the upstream sherpa-onnx
* release asset, matching the mirror-not-the-GH-tarball approach already
* used for Moonshine below. Every hash here was computed locally from a
* freshly downloaded copy of the file, not copied from a doc page.
*
* Mirrors the on-phone assistant model flow in llama.ts, except each model
* here is multiple files rather than one; they download sequentially through
* FaceclawModelDownloader (resume + pinned sha256 per file).
*/
export type AsrModelId = "moonshine" | "whisper-base-en";
type AsrModelFile = {
name: string;
sha256: string;
sizeBytes: number;
};
export const ASR_MODEL = {
label: "Moonshine (English)",
dirName: "sherpa-onnx-moonshine-base-en-quantized-2026-02-27",
// Hugging Face mirror of the sherpa-onnx release asset of the same name.
// The GitHub release only offers a tar.bz2, which the phone can't unpack;
// this mirror serves the same files (sha256-verified) individually.
baseUrl:
"https://huggingface.co/csukuangfj2/sherpa-onnx-moonshine-base-en-quantized-2026-02-27/resolve/main/",
files: [
{
name: "decoder_model_merged.ort",
sha256: "d9d7b333af34bc552580576ddcf248a1c6c839e0d3b43b09afb9376ed009899d",
sizeBytes: 109424400,
},
{
name: "encoder_model.ort",
sha256: "7c66495948d0d08ec1af454cd4b5514862ae6511e94712a60e6d83eaec8dc8cf",
sizeBytes: 31326816,
},
{
name: "tokens.txt",
sha256: "2870d843e14c1e187bf1913a521562a63b53933814bd7f2145120468f494a049",
sizeBytes: 549350,
},
] as AsrModelFile[],
totalBytes: 141300566,
type AsrModelDef = {
label: string;
dirName: string;
baseUrl: string;
files: AsrModelFile[];
totalBytes: number;
};
export const ASR_MODELS: Record<AsrModelId, AsrModelDef> = {
moonshine: {
label: "Moonshine (English)",
dirName: "sherpa-onnx-moonshine-base-en-quantized-2026-02-27",
// Hugging Face mirror of the sherpa-onnx release asset of the same name.
// The GitHub release only offers a tar.bz2, which the phone can't unpack;
// this mirror serves the same files (sha256-verified) individually.
baseUrl:
"https://huggingface.co/csukuangfj2/sherpa-onnx-moonshine-base-en-quantized-2026-02-27/resolve/main/",
files: [
{
name: "decoder_model_merged.ort",
sha256: "d9d7b333af34bc552580576ddcf248a1c6c839e0d3b43b09afb9376ed009899d",
sizeBytes: 109424400,
},
{
name: "encoder_model.ort",
sha256: "7c66495948d0d08ec1af454cd4b5514862ae6511e94712a60e6d83eaec8dc8cf",
sizeBytes: 31326816,
},
{
name: "tokens.txt",
sha256: "2870d843e14c1e187bf1913a521562a63b53933814bd7f2145120468f494a049",
sizeBytes: 549350,
},
],
totalBytes: 141300566,
},
"whisper-base-en": {
label: "Whisper (English, base)",
dirName: "sherpa-onnx-whisper-base-en-int8",
// csukuangfj/sherpa-onnx-whisper-base.en on Hugging Face: the sherpa-onnx
// maintainer's own mirror of the project's whisper export, individual
// files (no tar.bz2-unpack problem to begin with, but kept as the same
// per-file-with-pinned-hash shape as Moonshine above for consistency).
baseUrl: "https://huggingface.co/csukuangfj/sherpa-onnx-whisper-base.en/resolve/main/",
files: [
{
name: "base.en-encoder.int8.onnx",
sha256: "ef6b936f4c9b1d90a3b68634b60c4ed8576b26172b33c2535ec0e933c9edb823",
sizeBytes: 29120534,
},
{
name: "base.en-decoder.int8.onnx",
sha256: "f7162ad6db2dbef16cfaeaa7f945b9d7dd9c1b8d472f6aca82f2273d185e4d41",
sizeBytes: 130669978,
},
{
name: "base.en-tokens.txt",
sha256: "306cd27f03c1a714eca7108e03d66b7dc042abe8c258b44c199a7ed9838dd930",
sizeBytes: 835554,
},
],
totalBytes: 160626066,
},
};
export type AsrModelState = {
@@ -56,121 +113,141 @@ export type AsrModelState = {
totalBytes: number;
};
let downloader: any = null;
// Bytes of files already fully downloaded in this run, plus progress within
// the file currently downloading; drives the aggregate percentage.
let completedBytes = 0;
let currentFileBytes = 0;
const stateListeners = new Set<(state: AsrModelState) => void>();
type ModelRuntime = {
downloader: any;
// Bytes of files already fully downloaded in this run, plus progress within
// the file currently downloading; drives the aggregate percentage.
completedBytes: number;
currentFileBytes: number;
stateListeners: Set<(state: AsrModelState) => void>;
};
function modelDirPath(): string {
function freshRuntime(): ModelRuntime {
return { downloader: null, completedBytes: 0, currentFileBytes: 0, stateListeners: new Set() };
}
const runtimes: Record<AsrModelId, ModelRuntime> = {
moonshine: freshRuntime(),
"whisper-base-en": freshRuntime(),
};
function modelDirPath(id: AsrModelId): string {
const context = Utils.android.getApplicationContext();
return `${context.getFilesDir().getAbsolutePath()}/faceclaw-voice-asr/${ASR_MODEL.dirName}`;
return `${context.getFilesDir().getAbsolutePath()}/faceclaw-voice-asr/${ASR_MODELS[id].dirName}`;
}
function filePath(file: AsrModelFile): string {
return `${modelDirPath()}/${file.name}`;
function filePath(id: AsrModelId, file: AsrModelFile): string {
return `${modelDirPath(id)}/${file.name}`;
}
function isFilePresent(file: AsrModelFile): boolean {
function isFilePresent(id: AsrModelId, file: AsrModelFile): boolean {
try {
const javaFile = new java.io.File(filePath(file));
const javaFile = new java.io.File(filePath(id, file));
return javaFile.exists() && javaFile.length() > 0;
} catch {
return false;
}
}
export function isAsrModelReady(): boolean {
export function isAsrModelReady(id: AsrModelId): boolean {
if (!global.isAndroid) return false;
return ASR_MODEL.files.every(isFilePresent);
return ASR_MODELS[id].files.every((file) => isFilePresent(id, file));
}
export function asrModelState(): AsrModelState {
if (downloader) {
export function asrModelState(id: AsrModelId): AsrModelState {
const runtime = runtimes[id];
const totalBytes = ASR_MODELS[id].totalBytes;
if (runtime.downloader) {
return {
status: "downloading",
bytesDownloaded: completedBytes + currentFileBytes,
totalBytes: ASR_MODEL.totalBytes,
bytesDownloaded: runtime.completedBytes + runtime.currentFileBytes,
totalBytes,
};
}
return {
status: isAsrModelReady() ? "ready" : "absent",
status: isAsrModelReady(id) ? "ready" : "absent",
bytesDownloaded: 0,
totalBytes: ASR_MODEL.totalBytes,
totalBytes,
};
}
export function onAsrModelStateChanged(listener: (state: AsrModelState) => void): () => void {
stateListeners.add(listener);
return () => stateListeners.delete(listener);
export function onAsrModelStateChanged(id: AsrModelId, listener: (state: AsrModelState) => void): () => void {
const runtime = runtimes[id];
runtime.stateListeners.add(listener);
return () => runtime.stateListeners.delete(listener);
}
function notifyStateChanged(): void {
const state = asrModelState();
stateListeners.forEach((listener) => listener(state));
function notifyStateChanged(id: AsrModelId): void {
const state = asrModelState(id);
runtimes[id].stateListeners.forEach((listener) => listener(state));
}
export function startAsrModelDownload(): void {
if (!global.isAndroid || downloader || isAsrModelReady()) return;
completedBytes = ASR_MODEL.files.filter(isFilePresent).reduce((sum, f) => sum + f.sizeBytes, 0);
downloadNextFile();
notifyStateChanged();
export function startAsrModelDownload(id: AsrModelId): void {
const runtime = runtimes[id];
if (!global.isAndroid || runtime.downloader || isAsrModelReady(id)) return;
runtime.completedBytes = ASR_MODELS[id].files
.filter((file) => isFilePresent(id, file))
.reduce((sum, f) => sum + f.sizeBytes, 0);
downloadNextFile(id);
notifyStateChanged(id);
}
function downloadNextFile(): void {
const nextFile = ASR_MODEL.files.find((file) => !isFilePresent(file));
function downloadNextFile(id: AsrModelId): void {
const runtime = runtimes[id];
const def = ASR_MODELS[id];
const nextFile = def.files.find((file) => !isFilePresent(id, file));
if (!nextFile) {
downloader = null;
notifyStateChanged();
runtime.downloader = null;
notifyStateChanged(id);
return;
}
currentFileBytes = 0;
runtime.currentFileBytes = 0;
const listener = new com.faceclaw.app.FaceclawModelDownloaderListener({
onProgress: (bytes: number, _total: number) => {
currentFileBytes = Number(bytes);
notifyStateChanged();
runtime.currentFileBytes = Number(bytes);
notifyStateChanged(id);
},
onDone: () => {
completedBytes += nextFile.sizeBytes;
currentFileBytes = 0;
downloadNextFile();
runtime.completedBytes += nextFile.sizeBytes;
runtime.currentFileBytes = 0;
downloadNextFile(id);
},
onError: (message: string) => {
console.error(`Voice model download failed (${nextFile.name}): ${message}`);
downloader = null;
notifyStateChanged();
console.error(`Voice model download failed (${id}/${nextFile.name}): ${message}`);
runtime.downloader = null;
notifyStateChanged(id);
},
});
downloader = new com.faceclaw.app.FaceclawModelDownloader(
`${ASR_MODEL.baseUrl}${nextFile.name}`,
filePath(nextFile),
runtime.downloader = new com.faceclaw.app.FaceclawModelDownloader(
`${def.baseUrl}${nextFile.name}`,
filePath(id, nextFile),
nextFile.sha256,
nextFile.sizeBytes,
listener,
);
downloader.start();
runtime.downloader.start();
}
/** Stops the download; already-fetched bytes are kept and resumed next time. */
export function cancelAsrModelDownload(): void {
if (!downloader) return;
downloader.cancel();
downloader = null;
notifyStateChanged();
export function cancelAsrModelDownload(id: AsrModelId): void {
const runtime = runtimes[id];
if (!runtime.downloader) return;
runtime.downloader.cancel();
runtime.downloader = null;
notifyStateChanged(id);
}
export function deleteAsrModel(): void {
export function deleteAsrModel(id: AsrModelId): void {
if (!global.isAndroid) return;
cancelAsrModelDownload();
cancelAsrModelDownload(id);
try {
for (const file of ASR_MODEL.files) {
new java.io.File(filePath(file)).delete();
new java.io.File(`${filePath(file)}.part`).delete();
for (const file of ASR_MODELS[id].files) {
new java.io.File(filePath(id, file)).delete();
new java.io.File(`${filePath(id, file)}.part`).delete();
}
new java.io.File(modelDirPath()).delete();
new java.io.File(modelDirPath(id)).delete();
} catch (error) {
console.error(`Voice model delete failed: ${String(error)}`);
console.error(`Voice model delete failed (${id}): ${String(error)}`);
}
notifyStateChanged();
notifyStateChanged(id);
}
+10 -2
View File
@@ -26,7 +26,12 @@ export type VoiceControlState = {
detail: string;
};
export type VoiceProviderKind = "onboard" | "elevenlabs" | "whisper" | "soniox";
// "whisper" (no "onboard-" prefix) is the OpenAI CLOUD provider (gpt-realtime-whisper,
// see openai-stt.ts) -- an unfortunate pre-existing name collision with the model
// family, not with "onboard-whisper" below, which is the on-device sherpa-onnx
// Whisper backend added alongside "onboard" (Moonshine). Kept as-is rather than
// renamed, since it's a persisted settings value on real installs already.
export type VoiceProviderKind = "onboard" | "onboard-whisper" | "elevenlabs" | "whisper" | "soniox";
export type VoiceTranscriptEvent = CloudSttTranscriptEvent & {
/**
@@ -314,6 +319,9 @@ export class FaceclawVoiceControlBridge {
this.cloudClient = null;
this.started = true;
// Which on-device model to load; a no-op setter for every provider except
// "onboard-whisper" (FaceclawVoiceController defaults to Moonshine).
this.controller?.setOnboardModelKind(options.provider === "onboard-whisper" ? "whisper" : "moonshine");
this.controller?.start("onboard");
}
@@ -323,7 +331,7 @@ export class FaceclawVoiceControlBridge {
* than failing the capture outright.
*/
private createCloudClient(options: PushToTalkOptions): CloudSttClient | null {
if (options.provider === "onboard") return null;
if (options.provider === "onboard" || options.provider === "onboard-whisper") return null;
const sttOptions = {
apiKey: "",
onTranscript: (event: CloudSttTranscriptEvent) =>
+12 -5
View File
@@ -523,12 +523,19 @@ export const ringConnectionModeSetting = new ConfigSettingEnum<RingConnectionMod
"How R1 ring input reaches the phone. Only via glasses: the ring's own link to the glasses carries its gestures, and the phone never opens a Bluetooth connection to the ring. Direct: also connect to the ring from the phone (currently unreliable). Takes effect on the next connection to the glasses.",
});
export type VoiceProvider = "onboard" | "elevenlabs" | "whisper" | "soniox";
// "whisper" (no "onboard-" prefix) is OpenAI's CLOUD realtime model
// (gpt-realtime-whisper); "onboard-whisper" is the on-device sherpa-onnx
// Whisper backend. Same underlying model family, two different places it
// runs -- see the same note in native/voice-control.ts. The "whisper" value
// keeps its name (it's a persisted setting on real installs) but its label
// below now says "OpenAI" to tell the two apart in the picker.
export type VoiceProvider = "onboard" | "onboard-whisper" | "elevenlabs" | "whisper" | "soniox";
const voiceProviderLabels: Record<VoiceProvider, string> = {
onboard: "On-device",
onboard: "On-device (Moonshine)",
"onboard-whisper": "On-device (Whisper)",
elevenlabs: "ElevenLabs",
whisper: "Whisper",
whisper: "OpenAI (Whisper)",
soniox: "Soniox",
};
@@ -537,7 +544,7 @@ export const voiceProviderSetting = new ConfigSettingEnum<VoiceProvider>({
label: "Transcription Provider",
storageKey: "voice.provider",
defaultValue: "onboard",
values: ["onboard", "elevenlabs", "whisper", "soniox"],
values: ["onboard", "onboard-whisper", "elevenlabs", "whisper", "soniox"],
formatValue: (value) => voiceProviderLabels[value] ?? value,
isDisabled: (value) => {
if (value === "elevenlabs") return elevenLabsApiKeySetting.get().trim().length === 0;
@@ -545,7 +552,7 @@ export const voiceProviderSetting = new ConfigSettingEnum<VoiceProvider>({
if (value === "soniox") return sonioxApiKeySetting.get().trim().length === 0;
return false;
},
description: "Speech-to-text engine for voice input. ElevenLabs, Whisper, and Soniox are cloud services that need an API key, with significantly better accuracy than on-device transcription. On-device transcription needs the voice model downloaded (below).",
description: "Speech-to-text engine for voice input. ElevenLabs, OpenAI, and Soniox are cloud services that need an API key, with significantly better accuracy than on-device transcription. The two On-device options need their voice model downloaded (below) and never leave the phone.",
});
const wakeWordActionLabels: Record<WakeWordAction, string> = {
+31 -25
View File
@@ -13,12 +13,13 @@ import {
startLocalModelDownload,
} from "../../native/llama";
import {
ASR_MODEL,
ASR_MODELS,
asrModelState,
cancelAsrModelDownload,
deleteAsrModel,
onAsrModelStateChanged,
startAsrModelDownload,
type AsrModelId,
} from "../../native/asr-model";
import { TextViewerLayer } from "../../apps/files/text-viewer";
import type { LayerContext } from "../layers";
@@ -119,7 +120,8 @@ function settingsSections(): SettingsSection[] {
items: [
enumSettingMenuItem(wakeWordActionSetting),
enumSettingMenuItem(voiceProviderSetting),
asrModelMenuItem(),
asrModelMenuItem("moonshine"),
asrModelMenuItem("whisper-base-en"),
],
},
{
@@ -344,48 +346,52 @@ function localModelMenuItem(): MenuItem {
};
}
const ASR_MODEL_MB = `${Math.round(ASR_MODEL.totalBytes / 1e6)}MB`;
function asrModelMb(id: AsrModelId): string {
return `${Math.round(ASR_MODELS[id].totalBytes / 1e6)}MB`;
}
let asrModelRenderUnsub: (() => void) | null = null;
const asrModelRenderUnsub: Partial<Record<AsrModelId, () => void>> = {};
function watchAsrModelDownload(ctx: LayerContext): void {
asrModelRenderUnsub?.();
asrModelRenderUnsub = onAsrModelStateChanged((state) => {
function watchAsrModelDownload(id: AsrModelId, ctx: LayerContext): void {
asrModelRenderUnsub[id]?.();
asrModelRenderUnsub[id] = onAsrModelStateChanged(id, (state) => {
ctx.actions.requestRender();
if (state.status !== "downloading") {
asrModelRenderUnsub?.();
asrModelRenderUnsub = null;
asrModelRenderUnsub[id]?.();
asrModelRenderUnsub[id] = undefined;
}
});
}
function asrModelStatusText(): string {
const state = asrModelState();
function asrModelStatusText(id: AsrModelId): string {
const state = asrModelState(id);
if (state.status === "ready") return "downloaded";
if (state.status === "downloading") {
const pct = state.totalBytes > 0 ? Math.floor((state.bytesDownloaded / state.totalBytes) * 100) : 0;
return `${pct}% of ${ASR_MODEL_MB}`;
return `${pct}% of ${asrModelMb(id)}`;
}
return "not downloaded";
}
/** Download/cancel/delete management for the on-device transcription model. */
function asrModelMenuItem(): MenuItem {
/** Download/cancel/delete management for one on-device transcription model. */
function asrModelMenuItem(id: AsrModelId): MenuItem {
const def = ASR_MODELS[id];
const rowLabel = `On-device model: ${def.label}`;
return {
label: "On-device voice model",
label: rowLabel,
description:
`${ASR_MODEL.label} (${ASR_MODEL_MB} download). ` +
`${def.label} (${asrModelMb(id)} download). ` +
"Transcribes voice input on the phone itself, with no API key or cloud service. " +
"Required for the On-device transcription provider; the cloud providers work without it. " +
"Required for its matching Transcription Provider option; the other providers work without it. " +
"An interrupted download resumes where it left off.",
onSelect: (ctx) => {
const state = asrModelState();
const state = asrModelState(id);
const action: MenuItem =
state.status === "downloading"
? {
label: "Cancel download",
onSelect: (innerCtx) => {
cancelAsrModelDownload();
cancelAsrModelDownload(id);
innerCtx.stack.pop();
},
}
@@ -393,22 +399,22 @@ function asrModelMenuItem(): MenuItem {
? {
label: "Delete model",
onSelect: (innerCtx) => {
deleteAsrModel();
deleteAsrModel(id);
innerCtx.stack.pop();
},
}
: {
label: `Download (${ASR_MODEL_MB})`,
label: `Download (${asrModelMb(id)})`,
onSelect: (innerCtx) => {
startAsrModelDownload();
watchAsrModelDownload(innerCtx);
startAsrModelDownload(id);
watchAsrModelDownload(id, innerCtx);
innerCtx.stack.pop();
},
};
openModalMenu(ctx, "On-device voice model", [action], 0);
openModalMenu(ctx, rowLabel, [action], 0);
},
render: ({ image, x, y, width }) => {
drawRightValueMenuItem(image, getDefaultSmallFont(), x, y, width, "On-device voice model", asrModelStatusText());
drawRightValueMenuItem(image, getDefaultSmallFont(), x, y, width, rowLabel, asrModelStatusText(id));
},
};
}