diff --git a/config/electron-builder.config.cjs b/config/electron-builder.config.cjs index fc33b98fec6..b4655ee824e 100644 --- a/config/electron-builder.config.cjs +++ b/config/electron-builder.config.cjs @@ -174,19 +174,6 @@ const commonExtraResources = [ nativeChatVisualsResource, emojiShortcodeDatasetResource ] -// Native speech addons need real disk paths; copy only the artifact target’s variant. -const macSpeechNativeResource = { - from: 'node_modules/sherpa-onnx-darwin-${arch}', - to: 'node_modules/sherpa-onnx-darwin-${arch}' -} -const linuxSpeechNativeResource = { - from: 'node_modules/sherpa-onnx-linux-${arch}', - to: 'node_modules/sherpa-onnx-linux-${arch}' -} -const winSpeechNativeResource = { - from: 'node_modules/sherpa-onnx-win-x64', - to: 'node_modules/sherpa-onnx-win-x64' -} // electron-builder replaces these defaults when `depends` is configured; retain // Electron's loader requirements alongside Orca's headless-host dependencies. const debElectronRuntimeDependencies = [ @@ -332,8 +319,8 @@ module.exports = { '!resources/native-chat-visuals/**', // extraResources already ships plugin bytes; exclude source duplicates. '!resources/plugins/launch/**', - // Platform extraResources selects speech natives; asar would duplicate it and ship other variants. - '!node_modules/{sherpa-onnx*,transcribe-cpp,@transcribe-cpp,koffi,@koromix}{,/**/*}', + // afterPack copies the target's speech engine to resources/node_modules; asar would duplicate it and ship other variants. + '!node_modules/{transcribe-cpp,@transcribe-cpp,koffi,@koromix}{,/**/*}', // CLI shim belongs beside resources/bin/orca.exe; unpacking source creates a second .cmd without its executable (#7351). '!resources/win32{,/**/*}' ], @@ -563,7 +550,6 @@ module.exports = { extraResources: [ ...commonExtraResources, ...windowsRuntimeResources, - winSpeechNativeResource, { from: 'resources/win32/bin/orca.cmd', to: 'bin/orca.cmd' @@ -620,7 +606,8 @@ module.exports = { rank: 'Alternate' })) ], - // Bundled ONNX Runtime (speech) needs macOS 13.4 (#22690); the update-manifest gate reads this from Info.plist. + // Set for ONNX Runtime (#22690, #39135), now gone; transcribe-cpp needs 11.0, Electron 12.0. Kept until + // Kelvin decides. The update-manifest gate reads this from Info.plist. minimumSystemVersion: '13.4', icon: 'resources/build/icon.icns', entitlements: 'resources/build/entitlements.mac.plist', @@ -643,7 +630,6 @@ module.exports = { extraResources: [ ...commonExtraResources, ...createPackagedRuntimeNodeModuleResources('darwin'), - macSpeechNativeResource, { from: 'resources/darwin/bin/orca', to: 'bin/orca' @@ -708,7 +694,6 @@ module.exports = { extraResources: [ ...commonExtraResources, ...createPackagedRuntimeNodeModuleResources('linux'), - linuxSpeechNativeResource, { from: 'native/wayland-clipboard/.build/${arch}/orca-wayland-clipboard', to: 'bin/orca-wayland-clipboard' diff --git a/config/main-sync-fs-baseline.txt b/config/main-sync-fs-baseline.txt index 6107621f3e8..55df3070cc5 100644 --- a/config/main-sync-fs-baseline.txt +++ b/config/main-sync-fs-baseline.txt @@ -499,7 +499,6 @@ 1 src/main/speech/speech-model-files.ts 2 src/main/speech/speech-model-http-download.ts 2 src/main/speech/stt-gpu-failure-marker.ts -1 src/main/speech/stt-worker-model-config.ts 2 src/main/sqlite/harden-database-files.ts 1 src/main/sqlite/row-cache/keyed-json-row-cache.ts 2 src/main/sqlite/row-cache/legacy-json-source.ts @@ -610,7 +609,7 @@ 7 src/shared/serve-bundle-park.ts 1 src/shared/serve-user-data-path.ts 3 src/shared/source-scan/source-tree-scan.ts -3 src/shared/speech-recognizer-package.ts +2 src/shared/speech-recognizer-package.ts 5 src/shared/startup-command-staging.ts 1 src/shared/terminal/pty/native-process/pty-process-ceiling-state.ts 8 src/shared/windows-breakaway-launcher.ts diff --git a/config/packaged-runtime-node-modules.cjs b/config/packaged-runtime-node-modules.cjs index e2372b8e99d..5a3893af9e5 100644 --- a/config/packaged-runtime-node-modules.cjs +++ b/config/packaged-runtime-node-modules.cjs @@ -82,7 +82,6 @@ const PACKAGED_MAIN_REQUIRED_FILES = [ const PACKAGED_MAIN_SOURCE_RE = /^out\/main\/.+\.js$/ const TYPE_DECLARATION_ARTIFACT_RE = /\.d\.(?:c|m)?ts(?:\.map)?$/ const JS_SOURCE_MAP_ARTIFACT_RE = /\.(?:c|m)?js\.map$/ -const VERSIONED_ONNXRUNTIME_DYLIB_RE = /^libonnxruntime\.\d[\d.]*\.dylib$/ const NODE_BUILTINS = new Set([ ...builtinModules, @@ -519,31 +518,6 @@ function prunePackagedRuntimeTypeAndSourceMapArtifacts(resourcesDir) { pruneMatchingFiles(nodeModulesDir, isPrunableTypeOrSourceMapArtifact) } -function prunePackagedSherpaOnnx(resourcesDir, electronPlatformName) { - if (electronPlatformName !== 'darwin') { - return - } - const nodeModulesDir = join(resourcesDir, 'node_modules') - if (!existsSync(nodeModulesDir)) { - return - } - for (const entry of readdirSync(nodeModulesDir, { withFileTypes: true })) { - if (!entry.isDirectory() || !entry.name.startsWith('sherpa-onnx-darwin-')) { - continue - } - const packageDir = join(nodeModulesDir, entry.name) - const packageEntries = readdirSync(packageDir) - const hasVersionedOnnxRuntime = packageEntries.some((filename) => - VERSIONED_ONNXRUNTIME_DYLIB_RE.test(filename) - ) - if (hasVersionedOnnxRuntime) { - // Why: darwin sherpa-onnx binaries link to the versioned ONNX Runtime - // install name; the unversioned dylib is a duplicate fallback copy. - rmSync(join(packageDir, 'libonnxruntime.dylib'), { force: true }) - } - } -} - function prunePackagedZodSources(resourcesDir) { // Why: Zod's src tree is TypeScript source only selected by the @zod/source // condition; packaged runtime import/require paths resolve to built JS. @@ -557,17 +531,6 @@ function assertPackagedNativeVariantsInstalled(electronPlatformName, electronArc const isInstalled = (name) => existsSync(join(nodeModulesDir, name, 'package.json')) const missing = [] - const rootOptionalDependencies = - JSON.parse(readFileSync(join(projectDir, 'package.json'), 'utf8')).optionalDependencies ?? {} - // Why win32 is always x64: winSpeechNativeResource packages sherpa-onnx-win-x64 for every - // Windows target (there is no sherpa-onnx-win-arm64; it runs under emulation). - const sherpaName = - electronPlatformName === 'win32' - ? 'sherpa-onnx-win-x64' - : `sherpa-onnx-${electronPlatformName}-${architecture}` - if (sherpaName in rootOptionalDependencies && !isInstalled(sherpaName)) { - missing.push(sherpaName) - } missing.push(...missingTranscribeCppPackages(readPackage, electronPlatformName, architecture)) // Why prefix, not equality: linux variants carry a libc suffix (watcher-linux-x64-glibc), @@ -636,7 +599,6 @@ function prunePackagedRuntimeNodeModules(resourcesDir, electronPlatformName, ele // Why before the filename walk: zod/src is deleted wholesale, so walking it first is wasted work. prunePackagedZodSources(resourcesDir) prunePackagedRuntimeTypeAndSourceMapArtifacts(resourcesDir) - prunePackagedSherpaOnnx(resourcesDir, electronPlatformName) } function pruneMatchingFiles(directory, shouldPrune) { @@ -665,7 +627,6 @@ module.exports = { prunePackagedParcelWatcher, prunePackagedRuntimeNodeModules, prunePackagedRuntimeTypeAndSourceMapArtifacts, - prunePackagedSherpaOnnx, prunePackagedZodSources, readPackage, verifyPackagedMainRuntimeDeps diff --git a/config/scripts/native-build/linux/verify-linux-glibc-floor.cjs b/config/scripts/native-build/linux/verify-linux-glibc-floor.cjs index a82ff3b1f3e..ab288f7ea91 100644 --- a/config/scripts/native-build/linux/verify-linux-glibc-floor.cjs +++ b/config/scripts/native-build/linux/verify-linux-glibc-floor.cjs @@ -43,9 +43,6 @@ const COMPAT_SLOT_GLIBC_FLOOR = glibcFloorProfile( [1, 3, 7] ) -// Lazy speech-worker sherpa prebuilt requires GCC11 GLIBCXX_3.4.29; exempt C++ floor only, retain glibc gate because it is not launch-critical. -const LIBSTDCXX_FLOOR_EXEMPT = /(?:^|[/\\])sherpa-onnx/ - // VER_FLG_WEAK needs may resolve null and use fallback; they must not raise required floor. const VER_FLG_WEAK = 0x2 @@ -118,15 +115,9 @@ function isLibstdcxxNode(name) { return name.startsWith('GLIBCXX_') || name.startsWith('CXXABI_') } -// Gate load-time requirements; exempt sherpa C++ needs only, never its glibc needs. -function findFloorViolations(needs, filePath = '', glibcFloor = DESKTOP_GLIBC_FLOOR) { - const exemptLibstdcxx = LIBSTDCXX_FLOOR_EXEMPT.test(filePath) - return needs.filter( - (need) => - !need.weak && - isVersionNodeAboveFloor(need.name, glibcFloor) && - !(exemptLibstdcxx && isLibstdcxxNode(need.name)) - ) +// Gate load-time requirements; weak needs can resolve to null and fall back. +function findFloorViolations(needs, glibcFloor = DESKTOP_GLIBC_FLOOR) { + return needs.filter((need) => !need.weak && isVersionNodeAboveFloor(need.name, glibcFloor)) } // Ubuntu20.04 keeps openpty/forkpty in libutil until glibc2.34 merge: require DT_NEEDED even when newer libc compat aliases satisfy link-time floor. @@ -402,7 +393,7 @@ function verifyLinuxGlibcFloor(rootDir, options = {}) { const { versionNeeds, neededLibraries, isStatic } = readDynamicInfo(filePath, objdumpPath) const isMuslTarget = isMuslTemplatePayload(filePath, neededLibraries, versionNeeds) // Remote musl payloads use their host's C++ runtime, not Ubuntu's libstdc++ or libutil. - const floorViolations = findFloorViolations(versionNeeds, filePath, glibcFloor).filter( + const floorViolations = findFloorViolations(versionNeeds, glibcFloor).filter( (need) => !isMuslTarget || !isLibstdcxxNode(need.name) ) // Inspect dynamic symbols only if relocated provider is absent from DT_NEEDED; healthy case avoids extra objdump. diff --git a/package.json b/package.json index 6970e3cfaad..4a038ca5fbd 100644 --- a/package.json +++ b/package.json @@ -398,11 +398,6 @@ "@transcribe-cpp/linux-x64-cpu-vulkan": "0.4.1", "@transcribe-cpp/win32-x64-cpu-vulkan": "0.4.1", "@vscode/windows-process-tree": "0.8.0", - "sherpa-onnx-darwin-arm64": "1.12.37", - "sherpa-onnx-darwin-x64": "1.12.37", - "sherpa-onnx-linux-arm64": "1.12.37", - "sherpa-onnx-linux-x64": "1.12.37", - "sherpa-onnx-win-x64": "1.12.37", "transcribe-cpp": "0.4.1" }, "lint-staged": { diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml index 153d8340ac6..16c159b227e 100644 --- a/pnpm-lock.yaml +++ b/pnpm-lock.yaml @@ -720,21 +720,6 @@ importers: '@vscode/windows-process-tree': specifier: 0.8.0 version: 0.8.0(patch_hash=a09f2e8b2b4e73ef20c317af91f9e06c09abf62fa2c891d99b16f53047f596fd) - sherpa-onnx-darwin-arm64: - specifier: 1.12.37 - version: 1.12.37 - sherpa-onnx-darwin-x64: - specifier: 1.12.37 - version: 1.12.37 - sherpa-onnx-linux-arm64: - specifier: 1.12.37 - version: 1.12.37 - sherpa-onnx-linux-x64: - specifier: 1.12.37 - version: 1.12.37 - sherpa-onnx-win-x64: - specifier: 1.12.37 - version: 1.12.37 transcribe-cpp: specifier: 0.4.1 version: 0.4.1 @@ -7660,31 +7645,6 @@ packages: resolution: {integrity: sha512-7++dFhtcx3353uBaq8DDR4NuxBetBzC7ZQOhmTQInHEd6bSrXdiEyzCvG07Z44UYdLShWUyXt5M/yhz8ekcb1A==} engines: {node: '>=8'} - sherpa-onnx-darwin-arm64@1.12.37: - resolution: {integrity: sha512-zpqbH+2TI6dvg7mxGm30Mnv17aJL3ZfRGshMiBK85dHBfhzqMbKUHNXCzsHhiKBQTzti6JqG1YoRbYrJwvdUjA==} - cpu: [arm64] - os: [darwin] - - sherpa-onnx-darwin-x64@1.12.37: - resolution: {integrity: sha512-2mB1BhEJJLG8Ow/YGAe7c3AL5fSkk9qDk3wmZRsfDW+Gno5eBaUusXseDI0ltjh2mZcW74l9F19ez4CyhwqNQQ==} - cpu: [x64] - os: [darwin] - - sherpa-onnx-linux-arm64@1.12.37: - resolution: {integrity: sha512-6yVnp7ucK4oJEjUfKbkyvJm+e6UdETTp7PnWFb1+Wr6nxSCN038BGSACCIKlVHEsQl+6huwrvymljkWrIro1mA==} - cpu: [arm64] - os: [linux] - - sherpa-onnx-linux-x64@1.12.37: - resolution: {integrity: sha512-WF69clWjTlBu6NIBdL8OLEEN4NwOlYjjB/AjCMs29pB6gtSzKJEKIjmTATfaSSExTGPeXsOriCXcMdHaMH4WNg==} - cpu: [x64] - os: [linux] - - sherpa-onnx-win-x64@1.12.37: - resolution: {integrity: sha512-1h1UrJxzA1vkOO2G8JqRQkF9SfEhM/LYW+yFki5izCY5N+eaaxYKdZLmVZI3KAhxk28pkE6dnMETYi0nj2/Cuw==} - cpu: [x64] - os: [win32] - shiki@4.5.0: resolution: {integrity: sha512-OcRW72ngQvc+MSVWBRIovC/BTW589iUE/Vxav7VeM4RcY9MUjqB+IR6ILG1k9nuW6LiG/nct9uma4/MBtnZbmQ==} engines: {node: '>=20'} @@ -15249,21 +15209,6 @@ snapshots: shebang-regex@3.0.0: {} - sherpa-onnx-darwin-arm64@1.12.37: - optional: true - - sherpa-onnx-darwin-x64@1.12.37: - optional: true - - sherpa-onnx-linux-arm64@1.12.37: - optional: true - - sherpa-onnx-linux-x64@1.12.37: - optional: true - - sherpa-onnx-win-x64@1.12.37: - optional: true - shiki@4.5.0: dependencies: '@shikijs/core': 4.5.0 diff --git a/src/cli/runtime/serve-speech-recognizer.test.ts b/src/cli/runtime/serve-speech-recognizer.test.ts index 9111b6a6ea8..32beb26e733 100644 --- a/src/cli/runtime/serve-speech-recognizer.test.ts +++ b/src/cli/runtime/serve-speech-recognizer.test.ts @@ -22,21 +22,23 @@ function resourcesWith(files: string[]): string { return resources } -describe("finding the app's speech engines (#31174)", () => { - it('answers nothing when the app ships no engine for this platform', () => { +describe("finding the app's speech engine (#31174)", () => { + it('answers nothing when the app ships no native package for this platform', () => { const resources = resourcesWith([ 'transcribe-cpp/package.json', - '@transcribe-cpp/linux-arm64-cpu-vulkan/contract.json', - 'sherpa-onnx-linux-arm64/sherpa-onnx.node' + '@transcribe-cpp/linux-arm64-cpu-vulkan/contract.json' ]) expect( findAppSpeechRecognizerDir(join(resources, 'app.asar.unpacked'), 'linux', 'x64') ).toBeUndefined() }) - // A loaded addon is locked on Windows; orcad holding the app's copy would block app updates. - it('never hands orcad the engines on Windows', () => { - const resources = resourcesWith(['sherpa-onnx-win-x64/sherpa-onnx.node']) + // A loaded library is locked on Windows; orcad holding the app's copy would block app updates. + it('never hands orcad the engine on Windows', () => { + const resources = resourcesWith([ + 'transcribe-cpp/package.json', + '@transcribe-cpp/win32-x64-cpu-vulkan/contract.json' + ]) expect( findAppSpeechRecognizerDir(join(resources, 'app.asar.unpacked'), 'win32', 'x64') ).toBeUndefined() diff --git a/src/cli/runtime/serve-speech-recognizer.ts b/src/cli/runtime/serve-speech-recognizer.ts index 3b5145f94ba..c8b94c18ba9 100644 --- a/src/cli/runtime/serve-speech-recognizer.ts +++ b/src/cli/runtime/serve-speech-recognizer.ts @@ -2,9 +2,9 @@ import { dirname, join } from 'node:path' import { hasSpeechEnginePackage } from '../../shared/speech-recognizer-package' /** - * The installed app's node_modules holding its on-device speech engines, for orcad's dictation - * (#31174). electron-builder puts them in `resources/node_modules`, beside the app root; a dev - * checkout keeps them in node_modules. + * The installed app's node_modules holding transcribe-cpp, for orcad's on-device dictation + * (#31174). electron-builder puts it in `resources/node_modules`, beside the app root; a dev + * checkout keeps it in node_modules. * * Not on Windows: a loaded addon locks its files for the server's life, which would make the next * app update fail to replace them. Windows orcad keeps cloud dictation only. @@ -17,9 +17,7 @@ export function findAppSpeechRecognizerDir( if (platform === 'win32') { return undefined } - return [join(dirname(appRoot), 'node_modules'), join(appRoot, 'node_modules')].find( - (dir) => - hasSpeechEnginePackage(dir, 'transcribe-cpp', platform, arch) || - hasSpeechEnginePackage(dir, 'sherpa', platform, arch) + return [join(dirname(appRoot), 'node_modules'), join(appRoot, 'node_modules')].find((dir) => + hasSpeechEnginePackage(dir, platform, arch) ) } diff --git a/src/main/ipc/speech.ts b/src/main/ipc/speech.ts index 6bdba861bbe..423c64e6dab 100644 --- a/src/main/ipc/speech.ts +++ b/src/main/ipc/speech.ts @@ -114,7 +114,7 @@ export function registerSpeechHandlers(store: Store): void { const getHotwordsFilePath = (content: string): string => { const digest = createHash('sha256').update(content).digest('hex').slice(0, 12) - // sherpa-onnx cannot read non-ASCII Windows paths; use the ASCII-safe model cache for hotwords. + // Native speech libraries may open paths with ANSI Windows APIs; keep hotwords in the ASCII-safe model cache. return join(getSpeechModelManager(store).getModelsDir(), `speech-hotwords-${digest}.txt`) } @@ -208,7 +208,7 @@ export function registerSpeechHandlers(store: Store): void { } if (hotwords && hotwords.length > 0) { - const content = `${hotwords.map((w) => `${w} :2.0`).join('\n')}\n` + const content = `${hotwords.join('\n')}\n` const hotwordsFilePath = getHotwordsFilePath(content) releaseHotwords = retainHotwordsFile(hotwordsFilePath) await writeFile(hotwordsFilePath, content, 'utf-8') diff --git a/src/main/runtime/speech/runtime-mobile-dictation-controller.ts b/src/main/runtime/speech/runtime-mobile-dictation-controller.ts index 0413eb9e043..e7af1ecf9be 100644 --- a/src/main/runtime/speech/runtime-mobile-dictation-controller.ts +++ b/src/main/runtime/speech/runtime-mobile-dictation-controller.ts @@ -54,7 +54,7 @@ export class RuntimeMobileDictationController { throw new Error('voice_model_not_selected') } const manifest = getCatalogModel(modelId) - if (manifest && isLocalSpeechModel(manifest) && !hasLocalSpeechRecognizer(manifest)) { + if (manifest && isLocalSpeechModel(manifest) && !hasLocalSpeechRecognizer()) { throw new Error(LOCAL_SPEECH_UNAVAILABLE_ERROR) } const modelState = await getSpeechModelManager(store).getModelState(modelId) diff --git a/src/main/runtime/speech/runtime-mobile-speech-catalog.ts b/src/main/runtime/speech/runtime-mobile-speech-catalog.ts index 4da3f086637..17c63be4f29 100644 --- a/src/main/runtime/speech/runtime-mobile-speech-catalog.ts +++ b/src/main/runtime/speech/runtime-mobile-speech-catalog.ts @@ -43,7 +43,7 @@ export class RuntimeMobileSpeechCatalog { const runnable = SPEECH_MODEL_CATALOG.filter( (manifest) => isOfferedSpeechModel(manifest) && - (!isLocalSpeechModel(manifest) || hasLocalSpeechRecognizer(manifest)) && + (!isLocalSpeechModel(manifest) || hasLocalSpeechRecognizer()) && // Phones offer an unready cloud row as "add an OpenAI key", which can't set this one up. (manifest.provider !== 'openai-compatible' || stateById.get(manifest.id)?.status === 'ready') @@ -74,7 +74,7 @@ export class RuntimeMobileSpeechCatalog { async download(modelId: string): Promise<{ started: true }> { const store = this.requireStore() const manifest = getCatalogModel(modelId) - if (!manifest || !isLocalSpeechModel(manifest) || !hasLocalSpeechRecognizer(manifest)) { + if (!manifest || !isLocalSpeechModel(manifest) || !hasLocalSpeechRecognizer()) { throw new Error('voice_model_not_downloadable') } void getSpeechModelManager(store) diff --git a/src/main/speech/model-cache-path.ts b/src/main/speech/model-cache-path.ts index 8af6e92a1b7..05c4876e87c 100644 --- a/src/main/speech/model-cache-path.ts +++ b/src/main/speech/model-cache-path.ts @@ -59,7 +59,7 @@ export function getSpeechModelCacheDirCandidates( .filter((modelsDir) => !hasNonAsciiCharacters(modelsDir)) .map((modelsDir) => ({ modelsDir, migrationSourceDir: requestedModelsDir })) - // sherpa-onnx 1.12.x needs ASCII Windows paths; retain the requested path if shared caches fail. + // Native speech libraries may open paths with ANSI Windows APIs; retain the requested path if shared caches fail. return [...candidates, { modelsDir: requestedModelsDir, migrationSourceDir: null }] } diff --git a/src/main/speech/model-catalog.ts b/src/main/speech/model-catalog.ts index aaf558927ae..a02730cd21f 100644 --- a/src/main/speech/model-catalog.ts +++ b/src/main/speech/model-catalog.ts @@ -49,7 +49,6 @@ export const SPEECH_MODEL_CATALOG: SpeechModelManifest[] = [ ...getSpeechModelDownloadMetadata('zipformer-bilingual-zh-en'), sampleRate: 16000, streaming: true, - modelingUnit: 'cjkchar+bpe', // transcribe-cpp has no zipformer; SenseVoice covers Chinese and English. replacedBy: 'sense-voice-zh-en-ja-ko-yue' }, @@ -77,7 +76,6 @@ export const SPEECH_MODEL_CATALOG: SpeechModelManifest[] = [ ...getSpeechModelDownloadMetadata('zipformer-streaming-en-20m'), sampleRate: 16000, streaming: true, - modelingUnit: 'bpe', // The model itself loses the first words of every utterance. replacedBy: 'parakeet-tdt-0.6b-v2-int8' }, @@ -91,7 +89,6 @@ export const SPEECH_MODEL_CATALOG: SpeechModelManifest[] = [ ...getSpeechModelDownloadMetadata('zipformer-streaming-zh-14m'), sampleRate: 16000, streaming: true, - modelingUnit: 'cjkchar', // transcribe-cpp has no zipformer; SenseVoice beat Nemotron 3.5 on Korean and Chinese. replacedBy: 'sense-voice-zh-en-ja-ko-yue' }, @@ -105,7 +102,6 @@ export const SPEECH_MODEL_CATALOG: SpeechModelManifest[] = [ ...getSpeechModelDownloadMetadata('zipformer-streaming-korean'), sampleRate: 16000, streaming: true, - modelingUnit: 'bpe', // transcribe-cpp has no zipformer; SenseVoice beat Nemotron 3.5 on Korean and Chinese. replacedBy: 'sense-voice-zh-en-ja-ko-yue' }, diff --git a/src/main/speech/stt-gpu-failure-marker.ts b/src/main/speech/stt-gpu-failure-marker.ts index 5a79cd52f22..5320a34967d 100644 --- a/src/main/speech/stt-gpu-failure-marker.ts +++ b/src/main/speech/stt-gpu-failure-marker.ts @@ -15,7 +15,7 @@ function gpuFailureKey(): string { return cachedGpuFailureKey } let version = 'unknown' - const modulesDir = getSpeechEngineModulesDir('transcribe-cpp') + const modulesDir = getSpeechEngineModulesDir() if (modulesDir) { try { version = JSON.parse( diff --git a/src/main/speech/stt-offline-audio-chunker.ts b/src/main/speech/stt-offline-audio-chunker.ts index 935e6921cb7..09b45586aac 100644 --- a/src/main/speech/stt-offline-audio-chunker.ts +++ b/src/main/speech/stt-offline-audio-chunker.ts @@ -1,4 +1,4 @@ -// Bound ONNX allocations: Chromium kills the app on a single allocation >=2GiB (#7925). +// Parakeet has no long-form mode, and bounded chunks keep the decode left for stop short. export const OFFLINE_DECODE_CHUNK_SECONDS = 30 // Split near quiet tail pauses to avoid mid-word cuts; ~100ms windows exclude short intra-word closures. diff --git a/src/main/speech/stt-recognizer.ts b/src/main/speech/stt-recognizer.ts index 79ce45c2c29..2191fdf3ce0 100644 --- a/src/main/speech/stt-recognizer.ts +++ b/src/main/speech/stt-recognizer.ts @@ -1,23 +1,25 @@ -import type { RecognizerConfigInput } from './stt-worker-model-config' - -export type SttRecognizerInit = RecognizerConfigInput & { - /** The node_modules directory holding this model's engine package. */ +export type SttRecognizerInit = { + /** The node_modules directory holding transcribe-cpp. */ speechModulesDir: string + modelDir: string + files: string[] + sampleRate: number + hotwordsFilePath?: string + /** Absent = auto-detect. */ + language?: string } export type SttRecognizerEvent = - | { type: 'partial'; streamId: string; text: string } - | { type: 'final'; streamId: string; text: string } /** One decoded offline chunk; `samples` counts the stream's input-rate samples decoded so far. */ - | { type: 'chunk'; streamId: string; text: string; samples: number } + { type: 'chunk'; streamId: string; text: string; samples: number } export type SttRecognizerPost = (event: SttRecognizerEvent) => void /** One loaded model in the speech host child, shared by every dictation stream; calls never overlap. */ export type SttRecognizer = { - feed(streamId: string, samples: Float32Array, sampleRate: number): void | Promise + feed(streamId: string, samples: Float32Array, sampleRate: number): Promise /** Decode the stream's buffered audio, post its text and drop the stream; may throw after posting. */ - finish(streamId: string): void | Promise + finish(streamId: string): Promise /** Drop the stream's audio undecoded. */ discard(streamId: string): void } diff --git a/src/main/speech/stt-session-prewarm.ts b/src/main/speech/stt-session-prewarm.ts index 0c32f07b208..afc4f7633e0 100644 --- a/src/main/speech/stt-session-prewarm.ts +++ b/src/main/speech/stt-session-prewarm.ts @@ -45,7 +45,7 @@ export async function prewarmSttModel( const ready = manifest !== undefined && config !== null && - hasLocalSpeechRecognizer(manifest) && + hasLocalSpeechRecognizer() && (await state.modelManager.getModelState(modelId)).status === 'ready' for (;;) { // Never disturb a dictation; whatever it loaded is what the user is using. diff --git a/src/main/speech/stt-session-start.ts b/src/main/speech/stt-session-start.ts index 04af95f835e..841396bf72a 100644 --- a/src/main/speech/stt-session-start.ts +++ b/src/main/speech/stt-session-start.ts @@ -20,7 +20,6 @@ import { teardownIdleSttEngine, teardownSttEngine } from './stt-session-stop' -import { speechEngineForModelFiles } from '../../shared/speech-model-engine' import { getSpeechEngineModulesDir, LOCAL_SPEECH_UNAVAILABLE_ERROR } from './stt-worker-paths' import { attachSttWorkerLifecycle, @@ -171,8 +170,7 @@ export function spawnSttEngine( config: SttRecognizerConfig, manifest: SpeechModelManifest ): SttEngine { - const files = manifest.files ?? [] - const speechModulesDir = getSpeechEngineModulesDir(speechEngineForModelFiles(files)) + const speechModulesDir = getSpeechEngineModulesDir() if (!speechModulesDir) { throw new Error(LOCAL_SPEECH_UNAVAILABLE_ERROR) } @@ -209,12 +207,9 @@ export function spawnSttEngine( initializeSttWorker(worker, { speechModulesDir, modelDir: state.modelManager.getModelDir(config.modelId), - modelType: manifest.type, - streaming: manifest.streaming, sampleRate: manifest.sampleRate, - files, + files: manifest.files ?? [], hotwordsFilePath: config.hotwordsFilePath, - modelingUnit: manifest.modelingUnit, language: config.language }) return engine diff --git a/src/main/speech/stt-sherpa-recognizer.ts b/src/main/speech/stt-sherpa-recognizer.ts deleted file mode 100644 index bf4a6668fc4..00000000000 --- a/src/main/speech/stt-sherpa-recognizer.ts +++ /dev/null @@ -1,149 +0,0 @@ -/* oxlint-disable typescript-eslint/no-explicit-any -- sherpa-onnx native addon has no type definitions */ -import { join } from 'node:path' -import { speechRecognizerPackageName } from '../../shared/speech-recognizer-package' -import { resampleToRate } from './stt-audio-resample' -import { OfflineAudioChunker } from './stt-offline-audio-chunker' -import { buildRecognizerConfig } from './stt-worker-model-config' -import { createStopTailPadding } from './stt-stop-tail-padding' -import type { SttRecognizer, SttRecognizerInit, SttRecognizerPost } from './stt-recognizer' - -// One per dictation session; every session shares the one loaded recognizer. -type SessionStream = { - // Online streams live for the session; offline ones are made per decode. - online: any - offlineChunker: OfflineAudioChunker | null - // Primed on the session's first feed so stop never waits on another session's stream. - primed: boolean - // Streaming models report the same hypothesis on most 100 ms feeds; only changes are posted. - lastPartialText: string -} - -export function createSherpaRecognizer( - init: SttRecognizerInit, - post: SttRecognizerPost -): SttRecognizer { - // Use the platform native addon: the main sherpa-onnx WASM package cannot load host model files. - const sherpa: any = require( - join(init.speechModulesDir, speechRecognizerPackageName(process.platform, process.arch)) - ) - const { streaming: isStreaming, sampleRate: modelSampleRate } = init - const { online, config } = buildRecognizerConfig(init) - const recognizer = online - ? sherpa.createOnlineRecognizer(config) - : sherpa.createOfflineRecognizer(config) - const streams = new Map() - - const openStream = (streamId: string): SessionStream => { - const existing = streams.get(streamId) - if (existing) { - return existing - } - const stream: SessionStream = { - online: isStreaming ? sherpa.createOnlineStream(recognizer) : null, - offlineChunker: isStreaming ? null : new OfflineAudioChunker(modelSampleRate), - primed: false, - lastPartialText: '' - } - streams.set(streamId, stream) - return stream - } - - const decodeOfflineChunk = (samples: Float32Array): string => { - const stream = sherpa.createOfflineStream(recognizer) - sherpa.acceptWaveformOffline(stream, { sampleRate: modelSampleRate, samples }) - sherpa.decodeOfflineStream(recognizer, stream) - const result = JSON.parse(sherpa.getOfflineStreamResultAsJson(stream)) - return result?.text?.trim() ?? '' - } - - const feedOnline = (streamId: string, stream: SessionStream, samples: Float32Array): void => { - if (!stream.primed) { - // New streams start with empty encoder caches; zipformer en-20M dropped most of the first second - // (first word wrong on 71 of 73 LibriSpeech clips). A second of silence fills them first. - sherpa.acceptWaveformOnline(stream.online, { - sampleRate: modelSampleRate, - samples: new Float32Array(modelSampleRate) - }) - stream.primed = true - } - sherpa.acceptWaveformOnline(stream.online, { sampleRate: modelSampleRate, samples }) - while (sherpa.isOnlineStreamReady(recognizer, stream.online)) { - sherpa.decodeOnlineStream(recognizer, stream.online) - } - const result = JSON.parse(sherpa.getOnlineStreamResultAsJson(recognizer, stream.online)) - const text = result?.text?.trim() - if (text && text !== stream.lastPartialText) { - stream.lastPartialText = text - post({ type: 'partial', streamId, text }) - } - if (sherpa.isEndpoint(recognizer, stream.online)) { - if (text) { - post({ type: 'final', streamId, text }) - } - sherpa.reset(recognizer, stream.online) - stream.lastPartialText = '' - } - } - - const feedOffline = (streamId: string, stream: SessionStream, samples: Float32Array): void => { - // Bound ONNX allocations below Chromium's >=2GiB crash threshold (#7925); consumers append final segments. - const readyChunks = stream.offlineChunker?.push(new Float32Array(samples)) ?? [] - // push() already consumed these windows; keep decoding later ones after a failure. - let firstError: unknown = null - for (const chunk of readyChunks) { - try { - const text = decodeOfflineChunk(chunk) - if (text) { - post({ type: 'final', streamId, text }) - } - } catch (err) { - firstError ??= err - } - } - if (firstError) { - throw firstError - } - } - - const finishStream = (stream: SessionStream): string => { - if (stream.online) { - sherpa.acceptWaveformOnline(stream.online, { - sampleRate: modelSampleRate, - samples: createStopTailPadding(modelSampleRate) - }) - sherpa.inputFinished(stream.online) - while (sherpa.isOnlineStreamReady(recognizer, stream.online)) { - sherpa.decodeOnlineStream(recognizer, stream.online) - } - const result = JSON.parse(sherpa.getOnlineStreamResultAsJson(recognizer, stream.online)) - return result?.text?.trim() ?? '' - } - // flush() leaves less than one bounded chunk. - const remaining = stream.offlineChunker?.flush() - return remaining && remaining.length > 0 ? decodeOfflineChunk(remaining) : '' - } - - return { - feed(streamId, samples, sampleRate) { - const stream = openStream(streamId) - // Normalize rates before native decode: mixed rates on one recognizer abort the process. - const resampled = resampleToRate(samples, sampleRate || modelSampleRate, modelSampleRate) - if (isStreaming) { - feedOnline(streamId, stream, resampled) - } else { - feedOffline(streamId, stream, resampled) - } - }, - finish(streamId) { - const stream = streams.get(streamId) - streams.delete(streamId) - const text = stream ? finishStream(stream) : '' - if (text) { - post({ type: 'final', streamId, text }) - } - }, - discard(streamId) { - streams.delete(streamId) - } - } -} diff --git a/src/main/speech/stt-stop-tail-padding.ts b/src/main/speech/stt-stop-tail-padding.ts deleted file mode 100644 index 77981589252..00000000000 --- a/src/main/speech/stt-stop-tail-padding.ts +++ /dev/null @@ -1,6 +0,0 @@ -// Streaming encoders emit a word only once lookahead frames follow it; sherpa-onnx's own decoders pad 0.66 s before inputFinished. -const STOP_TAIL_PADDING_SECONDS = 0.66 - -export function createStopTailPadding(sampleRate: number): Float32Array { - return new Float32Array(Math.round(STOP_TAIL_PADDING_SECONDS * sampleRate)) -} diff --git a/src/main/speech/stt-transcribe-cpp-recognizer.ts b/src/main/speech/stt-transcribe-cpp-recognizer.ts index dddd649d07d..f0bfefd74b0 100644 --- a/src/main/speech/stt-transcribe-cpp-recognizer.ts +++ b/src/main/speech/stt-transcribe-cpp-recognizer.ts @@ -5,7 +5,6 @@ import { pathToFileURL } from 'node:url' import type * as TranscribeCpp from 'transcribe-cpp' import { resampleToRate } from './stt-audio-resample' import { OfflineAudioChunker } from './stt-offline-audio-chunker' -import { resolveFile } from './stt-worker-model-config' import type { SttRecognizer, SttRecognizerInit, SttRecognizerPost } from './stt-recognizer' // Past ~8 threads decode slowed 2-4x under load in the 2026-11 benchmark; E-cores and SMT siblings @@ -15,13 +14,13 @@ function performanceCoreThreads(): number { return Math.min(logical, 8, Math.max(4, Math.floor(logical / 2))) } -// The hotwords file is written in sherpa's format, one `term :score` per line. The host deletes it -// once dictation starts, so a CPU fallback's reload finds it gone and runs without. +// One term per line. The host deletes the file once dictation starts, so a CPU fallback's reload +// finds it gone and runs without. async function readVocabulary(path: string | undefined): Promise { const text = path ? await readFile(path, 'utf8').catch(() => '') : '' return text .split('\n') - .map((line) => line.replace(/\s+:[\d.]+\s*$/, '').trim()) + .map((line) => line.trim()) .filter(Boolean) } @@ -39,9 +38,11 @@ export async function createTranscribeCppRecognizer( ): Promise { const entry = join(init.speechModulesDir, 'transcribe-cpp', 'dist', 'index.js') const transcribe: typeof TranscribeCpp = await import(pathToFileURL(entry).href) - const model = await transcribe.TranscribeModel.load( - resolveFile(init.files, '.gguf', init.modelDir, '.gguf') - ) + const gguf = init.files.find((file) => file.endsWith('.gguf')) + if (!gguf) { + throw new Error(`No .gguf model file in: ${init.files.join(', ')}`) + } + const model = await transcribe.TranscribeModel.load(join(init.modelDir, gguf)) // The child runs one call at a time, so every stream can share one session. const session = model.createSession({ nThreads: performanceCoreThreads() }) const vocabulary = await readVocabulary(init.hotwordsFilePath) diff --git a/src/main/speech/stt-worker-cpu-fallback.test.ts b/src/main/speech/stt-worker-cpu-fallback.test.ts index ba8c2eed943..aefde690b93 100644 --- a/src/main/speech/stt-worker-cpu-fallback.test.ts +++ b/src/main/speech/stt-worker-cpu-fallback.test.ts @@ -40,8 +40,6 @@ function setup(): { type: 'init', speechModulesDir: '/app/node_modules', modelDir: '/models/parakeet', - modelType: 'transducer', - streaming: false, sampleRate: 16000, files: ['model-Q8_0.gguf'] }) diff --git a/src/main/speech/stt-worker-cpu-fallback.ts b/src/main/speech/stt-worker-cpu-fallback.ts index 1ac9cd664a5..2f360f64951 100644 --- a/src/main/speech/stt-worker-cpu-fallback.ts +++ b/src/main/speech/stt-worker-cpu-fallback.ts @@ -1,6 +1,5 @@ import { EventEmitter } from 'node:events' import type { MessagePortMain } from 'electron' -import { speechEngineForModelFiles } from '../../shared/speech-model-engine' import { OFFLINE_DECODE_CHUNK_SECONDS } from './stt-offline-audio-chunker' import type { SttWorkerProcess, SttWorkerRequest } from './stt-worker-process' @@ -44,7 +43,6 @@ class SttCpuFallbackWorker extends EventEmitter implements SttWorkerProcess { private detachChild: () => void = () => undefined private cpuOnly: boolean private init: Extract | null = null - private transcribeCpp = false private ready = false private exitRequested = false private markOnCpuReady = false @@ -63,7 +61,6 @@ class SttCpuFallbackWorker extends EventEmitter implements SttWorkerProcess { postMessage(msg: SttWorkerRequest): void { if (msg.type === 'init') { this.init = msg - this.transcribeCpp = speechEngineForModelFiles(msg.files) === 'transcribe-cpp' } else if (msg.type === 'feed' && this.canFallBack()) { const stream = this.backlog(msg.streamId) stream.feeds.push(msg) @@ -87,7 +84,7 @@ class SttCpuFallbackWorker extends EventEmitter implements SttWorkerProcess { } private canFallBack(): boolean { - return this.transcribeCpp && !this.cpuOnly && !this.exitRequested + return this.init !== null && !this.cpuOnly && !this.exitRequested } private backlog(streamId: string): StreamBacklog { diff --git a/src/main/speech/stt-worker-model-config.ts b/src/main/speech/stt-worker-model-config.ts deleted file mode 100644 index b441ff745c2..00000000000 --- a/src/main/speech/stt-worker-model-config.ts +++ /dev/null @@ -1,214 +0,0 @@ -import { readdirSync } from 'node:fs' -import { availableParallelism } from 'node:os' -import { join } from 'node:path' - -// ONNX filenames vary by model; find each role in the manifest. -export function resolveFile( - files: string[], - role: string, - modelDir: string, - ext = '.onnx' -): string { - const match = files.find((f) => f.includes(role) && f.endsWith(ext)) - if (!match) { - throw new Error(`No *${role}*${ext} found in model files: ${files.join(', ')}`) - } - return join(modelDir, match) -} - -export function resolveTokens(files: string[], modelDir: string): string { - const match = files.find((f) => f.endsWith('tokens.txt')) - if (!match) { - throw new Error(`No *tokens.txt found in model files: ${files.join(', ')}`) - } - return join(modelDir, match) -} - -// BPE hotword matching needs a vocab file that older caches may lack. -function discoverBpeVocab(modelDir: string): string | undefined { - try { - const entries = readdirSync(modelDir) - const vocabFile = entries.find((f) => f.endsWith('.vocab')) - return vocabFile ? join(modelDir, vocabFile) : undefined - } catch { - return undefined - } -} - -export type HotwordsConfig = { - decodingMethod: string - hotwordsFile?: string - hotwordsScore?: number - modelingUnit?: string - bpeVocab?: string -} - -export function buildHotwordsConfig(opts: { - modelDir: string - modelType: string - hotwordsFilePath?: string - modelingUnit?: string -}): HotwordsConfig { - if (opts.modelType !== 'transducer' || !opts.hotwordsFilePath) { - return { decodingMethod: 'greedy_search' } - } - - const unit = opts.modelingUnit - if (unit?.includes('bpe')) { - const bpeVocab = discoverBpeVocab(opts.modelDir) - if (!bpeVocab) { - return { decodingMethod: 'greedy_search' } - } - return { - decodingMethod: 'modified_beam_search', - hotwordsFile: opts.hotwordsFilePath, - hotwordsScore: 1.5, - modelingUnit: unit, - bpeVocab - } - } - - return { - decodingMethod: 'modified_beam_search', - hotwordsFile: opts.hotwordsFilePath, - hotwordsScore: 1.5, - modelingUnit: unit - } -} - -export type RecognizerConfigInput = { - modelDir: string - modelType: string - streaming: boolean - sampleRate: number - files: string[] - hotwordsFilePath?: string - modelingUnit?: string - // Absent = auto-detect; only Whisper and SenseVoice read it. - language?: string -} - -// The user waits on offline decode at stop (and per 30 s chunk); 4 threads cut it ~35% at equal memory. -const OFFLINE_NUM_THREADS = Math.min(4, availableParallelism()) - -// Online configs go to createOnlineRecognizer, the rest to createOfflineRecognizer. -export function buildRecognizerConfig(msg: RecognizerConfigInput): { - online: boolean - config: Record -} { - const { modelDir, modelType, streaming, sampleRate, files } = msg - const tokens = resolveTokens(files, modelDir) - const hotwords = buildHotwordsConfig(msg) - - if (streaming && modelType === 'transducer') { - const config = { - featConfig: { sampleRate, featureDim: 80 }, - modelConfig: { - transducer: { - encoder: resolveFile(files, 'encoder', modelDir), - decoder: resolveFile(files, 'decoder', modelDir), - joiner: resolveFile(files, 'joiner', modelDir) - }, - tokens, - numThreads: 1, - provider: 'cpu', - debug: 0 - }, - ...hotwords, - enableEndpoint: 1, - rule1MinTrailingSilence: 2.4, - rule2MinTrailingSilence: 1.2, - rule3MinUtteranceLength: 20 - } - return { online: true, config } - } - if (streaming && modelType === 'paraformer') { - const config = { - featConfig: { sampleRate, featureDim: 80 }, - modelConfig: { - paraformer: { - encoder: resolveFile(files, 'encoder', modelDir), - decoder: resolveFile(files, 'decoder', modelDir) - }, - tokens, - numThreads: 1, - provider: 'cpu', - debug: 0 - }, - decodingMethod: 'greedy_search', - enableEndpoint: 1, - rule1MinTrailingSilence: 2.4, - rule2MinTrailingSilence: 1.2, - rule3MinUtteranceLength: 20 - } - return { online: true, config } - } - if (modelType === 'whisper') { - const config = { - featConfig: { sampleRate, featureDim: 80 }, - modelConfig: { - whisper: { - encoder: resolveFile(files, 'encoder', modelDir), - decoder: resolveFile(files, 'decoder', modelDir), - language: msg.language ?? '' - }, - tokens, - numThreads: OFFLINE_NUM_THREADS, - provider: 'cpu', - debug: 0 - }, - decodingMethod: 'greedy_search' - } - return { online: false, config } - } - if (modelType === 'nemo-ctc') { - const config = { - featConfig: { sampleRate, featureDim: 80 }, - modelConfig: { - nemoCtc: { - model: resolveFile(files, 'model', modelDir) - }, - tokens, - numThreads: OFFLINE_NUM_THREADS, - provider: 'cpu', - debug: 0 - }, - decodingMethod: 'greedy_search' - } - return { online: false, config } - } - if (modelType === 'senseVoice') { - const config = { - featConfig: { sampleRate, featureDim: 80 }, - modelConfig: { - senseVoice: { - model: resolveFile(files, 'model', modelDir), - language: msg.language ?? '', - useInverseTextNormalization: 1 - }, - tokens, - numThreads: OFFLINE_NUM_THREADS, - provider: 'cpu', - debug: 0 - }, - decodingMethod: 'greedy_search' - } - return { online: false, config } - } - const config = { - featConfig: { sampleRate, featureDim: 80 }, - modelConfig: { - transducer: { - encoder: resolveFile(files, 'encoder', modelDir), - decoder: resolveFile(files, 'decoder', modelDir), - joiner: resolveFile(files, 'joiner', modelDir) - }, - tokens, - numThreads: OFFLINE_NUM_THREADS, - provider: 'cpu', - debug: 0 - }, - ...hotwords - } - return { online: false, config } -} diff --git a/src/main/speech/stt-worker-paths.ts b/src/main/speech/stt-worker-paths.ts index f8e3edf8ede..4fa7d3a3f1d 100644 --- a/src/main/speech/stt-worker-paths.ts +++ b/src/main/speech/stt-worker-paths.ts @@ -1,6 +1,5 @@ import { join } from 'node:path' import { getAppEnvironment } from '../../shared/app-environment' -import { speechEngineForModelFiles, type SpeechEngine } from '../../shared/speech-model-engine' import { hasSpeechEnginePackage, SPEECH_RECOGNIZER_DIR_ENV @@ -23,10 +22,7 @@ function speechNodeModulesDirCandidates(): string[] { ? process.resourcesPath : undefined if (getAppEnvironment().isPackaged() && resourcesPath) { - return [ - join(resourcesPath, 'node_modules'), - join(resourcesPath, 'app.asar.unpacked', 'node_modules') - ] + return [join(resourcesPath, 'node_modules')] } if (getAppEnvironment().isPackaged()) { // orcad: only the installed app's copy, which `orca serve` names. @@ -36,23 +32,18 @@ function speechNodeModulesDirCandidates(): string[] { return require.resolve.paths('transcribe-cpp') ?? [] } -// The app's engine packages don't change while it runs, so each engine is looked up once. -const speechEngineModulesDirs = new Map() +// The app's engine packages don't change while it runs, so they are looked up once. +let speechEngineModulesDir: string | null | undefined -/** The node_modules directory holding `engine`, or null on hosts without it. */ -export function getSpeechEngineModulesDir(engine: SpeechEngine): string | null { - let dir = speechEngineModulesDirs.get(engine) - if (dir === undefined) { - dir = - speechNodeModulesDirCandidates().find((candidate) => - hasSpeechEnginePackage(candidate, engine) - ) ?? null - speechEngineModulesDirs.set(engine, dir) +/** The node_modules directory holding transcribe-cpp, or null on hosts without it. */ +export function getSpeechEngineModulesDir(): string | null { + if (speechEngineModulesDir === undefined) { + speechEngineModulesDir = + speechNodeModulesDirCandidates().find((dir) => hasSpeechEnginePackage(dir)) ?? null } - return dir + return speechEngineModulesDir } -/** Whether this host ships the engine `model` runs on. */ -export function hasLocalSpeechRecognizer(model: { files?: readonly string[] }): boolean { - return getSpeechEngineModulesDir(speechEngineForModelFiles(model.files ?? [])) !== null +export function hasLocalSpeechRecognizer(): boolean { + return getSpeechEngineModulesDir() !== null } diff --git a/src/main/speech/stt-worker.ts b/src/main/speech/stt-worker.ts index 45410d83c70..6922c958303 100644 --- a/src/main/speech/stt-worker.ts +++ b/src/main/speech/stt-worker.ts @@ -1,8 +1,6 @@ import type { MessagePortMain } from 'electron' -import { speechEngineForModelFiles } from '../../shared/speech-model-engine' import type { SpeechAudioPortMessage } from '../../shared/speech-types' import type { SttRecognizer, SttRecognizerInit, SttRecognizerPost } from './stt-recognizer' -import { createSherpaRecognizer } from './stt-sherpa-recognizer' import { createTranscribeCppRecognizer } from './stt-transcribe-cpp-recognizer' import { connectSttWorkerParentPort } from './stt-worker-parent-port' import type { SttWorkerRequest } from './stt-worker-process' @@ -29,10 +27,7 @@ function enqueue(task: () => void | Promise): void { async function handleInit(init: SttRecognizerInit): Promise { try { const post: SttRecognizerPost = (event) => parentPort.post(event) - recognizer = - speechEngineForModelFiles(init.files) === 'transcribe-cpp' - ? await createTranscribeCppRecognizer(init, post) - : createSherpaRecognizer(init, post) + recognizer = await createTranscribeCppRecognizer(init, post) parentPort.post({ type: 'ready' }) } catch (err) { parentPort.post({ type: 'error', error: String(err) }) diff --git a/src/renderer/src/components/settings/voice/VoicePane.tsx b/src/renderer/src/components/settings/voice/VoicePane.tsx index 4b2918d67fc..9b008cdd441 100644 --- a/src/renderer/src/components/settings/voice/VoicePane.tsx +++ b/src/renderer/src/components/settings/voice/VoicePane.tsx @@ -17,7 +17,6 @@ import { VoiceModelUnloadSetting } from './VoiceModelUnloadSetting' import { VoiceSpeechModelSection } from './VoiceSpeechModelSection' import { SettingsSwitchRow } from '../form-controls/SettingsFormControls' import { VoiceSpeechGpuSetting } from './VoiceSpeechGpuSetting' -import { speechEngineForModelFiles } from '../../../../../shared/speech-model-engine' import { matchesSettingsSearch } from '../search/settings-search' import { getOpenaiTranscriptionSearchEntry } from './voice-pane-search' import { useTranslate } from '@/i18n/use-translate' @@ -301,16 +300,15 @@ export function VoicePane({ settings, updateSettings }: VoicePaneProps): React.J onUpdateVoiceSettings={updateVoiceSettings} /> - {selectedModel && - speechEngineForModelFiles(selectedModel.files ?? []) === 'transcribe-cpp' && ( - <> - - - - )} + {selectedModel?.provider === 'local' && ( + <> + + + + )} {showOpenAiSettingsRow && ( <> diff --git a/src/shared/speech-model-engine.ts b/src/shared/speech-model-engine.ts deleted file mode 100644 index 6dd770b6695..00000000000 --- a/src/shared/speech-model-engine.ts +++ /dev/null @@ -1,6 +0,0 @@ -export type SpeechEngine = 'sherpa' | 'transcribe-cpp' - -/** GGUF models run on transcribe-cpp; everything else on sherpa-onnx. */ -export function speechEngineForModelFiles(files: readonly string[]): SpeechEngine { - return files.some((file) => file.endsWith('.gguf')) ? 'transcribe-cpp' : 'sherpa' -} diff --git a/src/shared/speech-recognizer-package.ts b/src/shared/speech-recognizer-package.ts index e4afe9a2d1b..27551971bce 100644 --- a/src/shared/speech-recognizer-package.ts +++ b/src/shared/speech-recognizer-package.ts @@ -1,6 +1,5 @@ import { existsSync } from 'node:fs' import { join } from 'node:path' -import type { SpeechEngine } from './speech-model-engine' import transcribeCppPackages from './transcribe-cpp-packages.json' /** Desktop and orca serve share the app’s speech engine packages; orcad slots omit them. */ @@ -8,26 +7,15 @@ import transcribeCppPackages from './transcribe-cpp-packages.json' /** Set by `orca serve` for orcad: the app's node_modules directory holding the speech engine packages. */ export const SPEECH_RECOGNIZER_DIR_ENV = 'ORCA_SPEECH_RECOGNIZER_DIR' -export function speechRecognizerPackageName(platform: string, arch: string): string { - return platform === 'win32' && arch === 'x64' - ? 'sherpa-onnx-win-x64' - : `sherpa-onnx-${platform}-${arch}` -} - // Packaging reads the same table. Intel Macs run transcribe-cpp on the CPU (its x64 build has no GPU backend). const TRANSCRIBE_CPP_PACKAGES: Record = transcribeCppPackages -/** Whether `nodeModulesDir` holds a loadable copy of `engine` for this host. */ +/** Whether `nodeModulesDir` holds transcribe-cpp and this host's native package for it. */ export function hasSpeechEnginePackage( nodeModulesDir: string, - engine: SpeechEngine, platform: string = process.platform, arch: string = process.arch ): boolean { - if (engine === 'sherpa') { - const sherpa = speechRecognizerPackageName(platform, arch) - return existsSync(join(nodeModulesDir, sherpa, 'sherpa-onnx.node')) - } const native = TRANSCRIBE_CPP_PACKAGES[`${platform}-${arch}`] return ( native !== undefined && diff --git a/src/shared/speech-types.ts b/src/shared/speech-types.ts index d6e123207e2..f5dddb46505 100644 --- a/src/shared/speech-types.ts +++ b/src/shared/speech-types.ts @@ -7,8 +7,6 @@ export type SpeechModelType = | 'openai' export type SpeechModelProvider = 'local' | 'openai' | 'openai-compatible' -export type ModelingUnit = 'bpe' | 'cjkchar' | 'cjkchar+bpe' - export type SpeechModelDownloadFile = { name: string url: string @@ -30,7 +28,6 @@ export type SpeechModelManifest = { legacyDownloadFiles?: SpeechModelDownloadFile[] sampleRate: number streaming: boolean - modelingUnit?: ModelingUnit recommended?: boolean /** Retired: never offered for selection, and a saved selection moves to this model on load. */ replacedBy?: string