Whisper ASR

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.7.9",
+	"version": "0.7.11",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",

src/audio.ts

@@ -1,6 +1,6 @@
 import {execSync, spawn} from 'node:child_process';
-import {mkdtempSync, rmSync} from 'node:fs';
-import fs from 'node:fs/promises';
+import {mkdtempSync} from 'node:fs';
+import fs, {rm} from 'node:fs/promises';
 import {tmpdir} from 'node:os';
 import Path, {join} from 'node:path';
 import {AbortablePromise, Ai} from './ai.ts';
@@ -12,7 +12,7 @@ export class Audio {
 
 	constructor(private ai: Ai) {
 		if(ai.options.whisper) {
-			this.whisperModel = ai.options.asr?.endsWith('.bin') ? ai.options.asr : ai.options.asr + '.bin';
+			this.whisperModel = ai.options.asr || 'ggml-base.en.bin';
 			this.downloadAsrModel();
 		}
 
@@ -76,13 +76,10 @@ print(json.dumps(segments))
 			if(aborted) return;
 			if(!p && !p3) throw new Error('Pyannote is not installed: pip install pyannote.audio');
 			const binary = p3 ? 'python3' : 'python';
-			let tmp: string | null = null;
 			return new Promise((resolve, reject) => {
-				tmp = join(mkdtempSync(join(tmpdir(), 'audio-')), 'converted.wav');
-				execSync(`ffmpeg -i "${file}" -ar 16000 -ac 1 -f wav "${tmp}"`, { stdio: 'ignore' });
 				if(aborted) return;
 				let output = '';
-				const proc = spawn(binary, ['-c', this.pyannote, tmp]);
+				const proc = spawn(binary, ['-c', this.pyannote, file]);
 				proc.stdout.on('data', (data: Buffer) => output += data.toString());
 				proc.stderr.on('data', (data: Buffer) => console.error(data.toString()));
 				proc.on('close', (code: number) => {
@@ -95,7 +92,7 @@ print(json.dumps(segments))
 				});
 				proc.on('error', reject);
 				abort = () => proc.kill('SIGTERM');
-			}).finally(() => { if(tmp) rmSync(Path.dirname(tmp), { recursive: true, force: true }); });
+			});
 		}));
 		return <any>Object.assign(p, {abort});
 	}
@@ -129,17 +126,22 @@ print(json.dumps(segments))
 	asr(file: string, options: { model?: string; diarization?: boolean | 'id' } = {}): AbortablePromise<string | null> {
 		if(!this.ai.options.whisper) throw new Error('Whisper not configured');
 
-		const transcript = this.runAsr(file, {model: options.model, diarization: !!options.diarization});
-		const diarization: any = options.diarization ? this.runDiarization(file) : Promise.resolve(null);
-		const abort = () => {
+		const tmp = join(mkdtempSync(join(tmpdir(), 'audio-')), 'converted.wav');
+		execSync(`ffmpeg -i "${file}" -ar 16000 -ac 1 -f wav "${tmp}"`, { stdio: 'ignore' });
+		const clean = () => rm(Path.dirname(tmp), { recursive: true, force: true }).catch(() => {});
+		const transcript = this.runAsr(tmp, {model: options.model, diarization: !!options.diarization});
+		const diarization: any = options.diarization ? this.runDiarization(tmp) : Promise.resolve(null);
+		let aborted = false, abort = () => {
+			aborted = true;
 			transcript.abort();
 			diarization?.abort?.();
+			clean();
 		};
 
 		const response = Promise.all([transcript, diarization]).then(async ([t, d]) => {
-			if(!options.diarization) return t;
+			if(aborted || !options.diarization) return t;
 			t = this.combineSpeakerTranscript(t, d);
-			if(options.diarization === 'id') {
+			if(!aborted && options.diarization === 'id') {
 				if(!this.ai.language.defaultModel) throw new Error('Configure an LLM for advanced ASR speaker detection');
 				let chunks = this.ai.language.chunk(t, 500, 0);
 				if(chunks.length > 4) chunks = [...chunks.slice(0, 3), <string>chunks.at(-1)];
@@ -150,7 +152,7 @@ print(json.dumps(segments))
 				Object.entries(names).forEach(([speaker, name]) => t = t.replaceAll(`[Speaker ${speaker}]`, `[${name}]`));
 			}
 			return t;
-		});
+		}).finally(() => clean());
 		return <any>Object.assign(response, {abort});
 	}