From 29cc22dc0ae1487db510fa3b0463d176df2a5f7f Mon Sep 17 00:00:00 2001 From: Kelvin Amoaba <97001695+AmoabaKelvin@users.noreply.github.com> Date: Mon, 9 Nov 2026 02:04:41 +0000 Subject: [PATCH] fix(speech): keep the engine's file reads off the main thread's hot paths The engine lookup and the GPU failure marker are read once per process instead of on every dictation start and mobile catalog request, the marker is written asynchronously, and the child reads the hotwords file with fs/promises. The baseline moves the existing engine presence checks from stt-worker-paths to the shared package table and adds the once-per-process marker reads. --- config/main-sync-fs-baseline.txt | 3 ++- src/main/speech/stt-gpu-failure-marker.ts | 18 +++++++++++------- .../speech/stt-transcribe-cpp-recognizer.ts | 12 +++++------- src/main/speech/stt-worker-paths.ts | 13 ++++++++++++- 4 files changed, 30 insertions(+), 16 deletions(-) diff --git a/config/main-sync-fs-baseline.txt b/config/main-sync-fs-baseline.txt index e106b4f0b3b..e2c5517cf6e 100644 --- a/config/main-sync-fs-baseline.txt +++ b/config/main-sync-fs-baseline.txt @@ -497,8 +497,8 @@ 2 src/main/speech/speech-model-download-cleanup.ts 1 src/main/speech/speech-model-download-transport.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/speech/stt-worker-paths.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 @@ -609,6 +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 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/src/main/speech/stt-gpu-failure-marker.ts b/src/main/speech/stt-gpu-failure-marker.ts index 2f04cc6e6d6..5a79cd52f22 100644 --- a/src/main/speech/stt-gpu-failure-marker.ts +++ b/src/main/speech/stt-gpu-failure-marker.ts @@ -1,4 +1,5 @@ -import { readFileSync, writeFileSync } from 'node:fs' +import { readFileSync } from 'node:fs' +import { writeFile } from 'node:fs/promises' import { release } from 'node:os' import { join } from 'node:path' import type { SttGpuPolicy } from './stt-worker-cpu-fallback' @@ -39,20 +40,23 @@ function readMarkerKey(file: string): string | null { /** `gpuEnabled` is the user's Voice setting; the marker records a GPU that failed on this machine. */ export function createSttGpuPolicy(modelsDir: string, gpuEnabled: () => boolean): SttGpuPolicy { const file = join(modelsDir, GPU_FAILURE_MARKER_FILE) + // Read once per process; this process's own failures update it in memory. + let failedKey: string | null | undefined return { allowGpu: () => { if (!gpuEnabled()) { return false } - const marker = readMarkerKey(file) - return marker === null || marker !== gpuFailureKey() + if (failedKey === undefined) { + failedKey = readMarkerKey(file) + } + return failedKey !== gpuFailureKey() }, markGpuFailed: () => { - try { - writeFileSync(file, JSON.stringify({ key: gpuFailureKey() })) - } catch (error) { + failedKey = gpuFailureKey() + writeFile(file, JSON.stringify({ key: failedKey })).catch((error: unknown) => { console.warn('[speech] Could not record the GPU failure:', error) - } + }) } } } diff --git a/src/main/speech/stt-transcribe-cpp-recognizer.ts b/src/main/speech/stt-transcribe-cpp-recognizer.ts index c04b205f962..5230e3a3446 100644 --- a/src/main/speech/stt-transcribe-cpp-recognizer.ts +++ b/src/main/speech/stt-transcribe-cpp-recognizer.ts @@ -1,4 +1,4 @@ -import { existsSync, readFileSync } from 'node:fs' +import { readFile } from 'node:fs/promises' import { availableParallelism } from 'node:os' import { join } from 'node:path' import { pathToFileURL } from 'node:url' @@ -17,11 +17,9 @@ function performanceCoreThreads(): number { // 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. -function readVocabulary(path: string | undefined): string[] { - if (!path || !existsSync(path)) { - return [] - } - return readFileSync(path, 'utf8') +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()) .filter(Boolean) @@ -46,7 +44,7 @@ export async function createTranscribeCppRecognizer( ) // The child runs one call at a time, so every stream can share one session. const session = model.createSession({ nThreads: performanceCoreThreads() }) - const vocabulary = readVocabulary(init.hotwordsFilePath) + const vocabulary = await readVocabulary(init.hotwordsFilePath) const options: TranscribeCpp.TranscribeOptions = { timestamps: 'none', ...(init.language && model.capabilities.languages.includes(init.language) diff --git a/src/main/speech/stt-worker-paths.ts b/src/main/speech/stt-worker-paths.ts index 6b7b3a4fd1b..f8e3edf8ede 100644 --- a/src/main/speech/stt-worker-paths.ts +++ b/src/main/speech/stt-worker-paths.ts @@ -36,9 +36,20 @@ 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 node_modules directory holding `engine`, or null on hosts without it. */ export function getSpeechEngineModulesDir(engine: SpeechEngine): string | null { - return speechNodeModulesDirCandidates().find((dir) => hasSpeechEnginePackage(dir, engine)) ?? null + let dir = speechEngineModulesDirs.get(engine) + if (dir === undefined) { + dir = + speechNodeModulesDirCandidates().find((candidate) => + hasSpeechEnginePackage(candidate, engine) + ) ?? null + speechEngineModulesDirs.set(engine, dir) + } + return dir } /** Whether this host ships the engine `model` runs on. */