Use one-off workers to process requests without blocking

src/audio.ts

@@ -1,133 +1,40 @@
-import {spawn} from 'node:child_process';
-import {pipeline} from '@xenova/transformers';
-import * as fs from 'node:fs';
+import {Worker} from 'worker_threads';
+import path from 'node:path';
 import {AbortablePromise, Ai} from './ai.ts';
-import wavefile from 'wavefile';
+import {canDiarization} from './asr.ts';
 
 export class Audio {
-	private whisperPipeline: any;
+	constructor(private ai: Ai) {}
 
-	constructor(private ai: Ai) { }
-
-	private combineSpeakerTranscript(chunks: any[], speakers: any[]): string {
-		const speakerMap = new Map();
-		let speakerCount = 0;
-		speakers.forEach((seg: any) => {
-			if(!speakerMap.has(seg.speaker)) speakerMap.set(seg.speaker, ++speakerCount);
-		});
-
-		const lines: string[] = [];
-		let currentSpeaker = -1;
-		let currentText = '';
-		chunks.forEach((chunk: any) => {
-			const time = chunk.timestamp[0];
-			const speaker = speakers.find((s: any) => time >= s.start && time <= s.end);
-			const speakerNum = speaker ? speakerMap.get(speaker.speaker) : 1;
-			if (speakerNum !== currentSpeaker) {
-				if(currentText) lines.push(`[speaker ${currentSpeaker}]: ${currentText.trim()}`);
-				currentSpeaker = speakerNum;
-				currentText = chunk.text;
-			} else {
-				currentText += chunk.text;
-			}
-		});
-		if(currentText) lines.push(`[speaker ${currentSpeaker}]: ${currentText.trim()}`);
-		return lines.join('\n');
-	}
-
-	async canDiarization(): Promise<boolean> {
-		return new Promise((resolve) => {
-			const proc = spawn('python3', ['-c', 'import pyannote.audio']);
-			proc.on('close', (code: number) => resolve(code === 0));
-			proc.on('error', () => resolve(false));
-		});
-	}
-
-	private async runDiarization(audioPath: string): Promise<any[]> {
-		if(!await this.canDiarization()) throw new Error('Pyannote is not installed: pip install pyannote.audio');
-		const script = `
-import sys
-import json
-from pyannote.audio import Pipeline
-
-os.environ['TORCH_HOME'] = "${this.ai.options.path}"
-pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1")
-diarization = pipeline(sys.argv[1])
-
-segments = []
-for turn, _, speaker in diarization.itertracks(yield_label=True):
-    segments.append({
-        "start": turn.start,
-        "end": turn.end,
-        "speaker": speaker
-    })
-
-print(json.dumps(segments))
-`;
-
-		return new Promise((resolve, reject) => {
-			let output = '';
-			const proc = spawn('python3', ['-c', script, audioPath]);
-			proc.stdout.on('data', (data: Buffer) => output += data.toString());
-			proc.stderr.on('data', (data: Buffer) => console.error(data.toString()));
-			proc.on('close', (code: number) => {
-				if(code === 0) {
-					try {
-						resolve(JSON.parse(output));
-					} catch (err) {
-						reject(new Error('Failed to parse diarization output'));
-					}
-				} else {
-					reject(new Error(`Python process exited with code ${code}`));
-				}
-			});
-
-			proc.on('error', reject);
-		});
-	}
-
-	asr(path: string, options: { model?: string; speaker?: boolean } = {}): AbortablePromise<string | null> {
+	asr(filepath: string, options: { model?: string; speaker?: boolean } = {}): AbortablePromise<string | null> {
 		const { model = this.ai.options.asr || 'whisper-base', speaker = false } = options;
 		let aborted = false;
 		const abort = () => { aborted = true; };
 
-		const p = new Promise<string | null>(async (resolve, reject) => {
-			try {
-				if(aborted) return resolve(null);
-				if(!this.whisperPipeline) this.whisperPipeline = await pipeline('automatic-speech-recognition', `Xenova/${model}`, { cache_dir: this.ai.options.path, quantized: true });
-
-				// Prepare audio file (convert to mono channel wave)
-				if(aborted) return resolve(null);
-				const wav = new wavefile.WaveFile(fs.readFileSync(path));
-				wav.toBitDepth('32f');
-				wav.toSampleRate(16000);
-				const samples = wav.getSamples();
-				let buffer;
-				if(Array.isArray(samples)) { // stereo to mono - average the channels
-					const left = samples[0];
-					const right = samples[1];
-					buffer = new Float32Array(left.length);
-					for (let i = 0; i < left.length; i++) buffer[i] = (left[i] + right[i]) / 2;
-				} else {
-					buffer = samples;
+		const p = new Promise<string | null>((resolve, reject) => {
+			const worker = new Worker(path.join(import.meta.dirname, 'asr.js'));
+			const handleMessage = ({ text, warning, error }: any) => {
+				worker.terminate();
+				if(aborted) return;
+				if(error) reject(new Error(error));
+				else {
+					if(warning) console.warn(warning);
+					resolve(text);
 				}
-
-				// Transcribe
-				if(aborted) return resolve(null);
-				const transcriptResult = await this.whisperPipeline(buffer, {return_timestamps: speaker ? 'word' : false});
-				if(!speaker) return resolve(transcriptResult.text?.trim() || null);
-
-				// Speaker Diarization
-				if(aborted) return resolve(null);
-				const speakers = await this.runDiarization(path);
-				if(aborted) return resolve(null);
-				const combined = this.combineSpeakerTranscript(transcriptResult.chunks || [], speakers);
-				resolve(combined);
-			} catch (err) {
-				reject(err);
-			}
+			};
+			const handleError = (err: Error) => {
+				worker.terminate();
+				if(!aborted) reject(err);
+			};
+			worker.on('message', handleMessage);
+			worker.on('error', handleError);
+			worker.on('exit', (code) => {
+				if(code !== 0 && !aborted) reject(new Error(`Worker exited with code ${code}`));
+			});
+			worker.postMessage({path: filepath, model, speaker, torchHome: this.ai.options.path,});
 		});
-
 		return Object.assign(p, { abort });
 	}
+
+	canDiarization = canDiarization;
 }