Whisper ASR

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.6.7",
+	"version": "0.7.11",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -25,15 +25,14 @@
 		"watch": "npx vite build --watch"
 	},
 	"dependencies": {
-		"@anthropic-ai/sdk": "^0.67.0",
+		"@anthropic-ai/sdk": "^0.78.0",
 		"@tensorflow/tfjs": "^4.22.0",
 		"@xenova/transformers": "^2.17.2",
-		"@ztimson/node-utils": "^1.0.4",
+		"@ztimson/node-utils": "^1.0.7",
-		"@ztimson/utils": "^0.27.9",
+		"@ztimson/utils": "^0.28.13",
 		"cheerio": "^1.2.0",
-		"openai": "^6.6.0",
+		"openai": "^6.22.0",
-		"tesseract.js": "^6.0.1",
+		"tesseract.js": "^7.0.0"
-		"wavefile": "^11.0.0"
 	},
 	"devDependencies": {
 		"@types/node": "^24.8.1",

src/ai.ts

@@ -8,9 +8,11 @@ export type AbortablePromise<T> = Promise<T> & {
 };
 
 export type AiOptions = {
+	/** Token to pull models from hugging face */
+	hfToken?: string;
 	/** Path to models */
 	path?: string;
-	/** ASR model: whisper-tiny, whisper-base */
+	/** Whisper ASR model: ggml-tiny.en.bin, ggml-base.en.bin */
 	asr?: string;
 	/** Embedding model: all-MiniLM-L6-v2, bge-small-en-v1.5, bge-large-en-v1.5 */
 	embedder?: string;
@@ -20,6 +22,8 @@ export type AiOptions = {
 	}
 	/** OCR model: eng, eng_best, eng_fast */
 	ocr?: string;
+	/** Whisper binary */
+	whisper?: string;
 }
 
 export class Ai {

src/asr.ts

@@ -1,124 +0,0 @@
-import { pipeline } from '@xenova/transformers';
-import { parentPort } from 'worker_threads';
-import * as fs from 'node:fs';
-import wavefile from 'wavefile';
-import { spawn } from 'node:child_process';
-
-let whisperPipeline: any;
-
-export async function 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));
-	});
-}
-
-async function runDiarization(audioPath: string, torchHome: string): Promise<any[]> {
-	const script = `
-import sys
-import json
-import os
-from pyannote.audio import Pipeline
-
-os.environ['TORCH_HOME'] = "${torchHome}"
-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);
-	});
-}
-
-function 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}]: ${currentText.trim()}`);
-	return lines.join('\n');
-}
-
-parentPort?.on('message', async ({ file, speaker, model, modelDir }) => {
-	try {
-		if(!whisperPipeline) whisperPipeline = await pipeline('automatic-speech-recognition', `Xenova/${model}`, {cache_dir: modelDir, quantized: true});
-
-		// Prepare audio file (convert to mono channel wave)
-		const wav = new wavefile.WaveFile(fs.readFileSync(file));
-		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;
-		}
-
-		// Transcribe
-		const transcriptResult = await whisperPipeline(buffer, {return_timestamps: speaker ? 'word' : false});
-		if(!speaker) {
-			parentPort?.postMessage({ text: transcriptResult.text?.trim() || null });
-			return;
-		}
-
-		// Speaker Diarization
-		const hasDiarization = await canDiarization();
-		if(!hasDiarization) {
-			parentPort?.postMessage({ text: transcriptResult.text?.trim() || null, error: 'Speaker diarization unavailable' });
-			return;
-		}
-
-		const speakers = await runDiarization(file, modelDir);
-		const combined = combineSpeakerTranscript(transcriptResult.chunks || [], speakers);
-		parentPort?.postMessage({ text: combined });
-	} catch (err) {
-		parentPort?.postMessage({ error: (err as Error).message });
-	}
-});

src/audio.ts

@@ -1,40 +1,174 @@
-import {Worker} from 'worker_threads';
+import {execSync, spawn} from 'node:child_process';
-import Path from 'node:path';
+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';
-import {canDiarization} from './asr.ts';
 
 export class Audio {
-	constructor(private ai: Ai) {}
+	private downloads: {[key: string]: Promise<string>} = {};
+	private pyannote!: string;
+	private whisperModel!: string;
 
-	asr(file: string, options: { model?: string; speaker?: boolean } = {}): AbortablePromise<string | null> {
+	constructor(private ai: Ai) {
-		const { model = this.ai.options.asr || 'whisper-base', speaker = false } = options;
+		if(ai.options.whisper) {
-		let aborted = false;
+			this.whisperModel = ai.options.asr || 'ggml-base.en.bin';
-		const abort = () => { aborted = true; };
+			this.downloadAsrModel();
+		}
 
-		const p = new Promise<string | null>((resolve, reject) => {
+		this.pyannote = `
-			const worker = new Worker(Path.join(import.meta.dirname, 'asr.js'));
+import sys
-			const handleMessage = ({ text, warning, error }: any) => {
+import json
-				worker.terminate();
+import os
+from pyannote.audio import Pipeline
+
+os.environ['TORCH_HOME'] = r"${ai.options.path}"
+pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", token="${ai.options.hfToken}")
+output = pipeline(sys.argv[1])
+
+segments = []
+for turn, speaker in output.speaker_diarization:
+    segments.append({"start": turn.start, "end": turn.end, "speaker": speaker})
+
+print(json.dumps(segments))
+`;
+	}
+
+	private runAsr(file: string, opts: {model?: string, diarization?: boolean} = {}): AbortablePromise<any> {
+		let proc: any;
+		const p = new Promise<any>((resolve, reject) => {
+			this.downloadAsrModel(opts.model).then(m => {
+				let output = '';
+				const args = [opts.diarization ? '-owts' : '-nt', '-m', m, '-f', file];
+				proc = spawn(<string>this.ai.options.whisper, args, {stdio: ['ignore', 'pipe', 'ignore']});
+				proc.on('error', (err: Error) => reject(err));
+				proc.stdout.on('data', (data: Buffer) => output += data.toString());
+				proc.on('close', (code: number) => {
+					if(code === 0) {
+						if(opts.diarization) {
+							try { resolve(JSON.parse(output)); }
+							catch(e) { reject(new Error('Failed to parse whisper JSON')); }
+						} else {
+							resolve(output.trim() || null);
+						}
+					} else {
+						reject(new Error(`Exit code ${code}`));
+					}
+				});
+			});
+		});
+		return <any>Object.assign(p, {abort: () => proc?.kill('SIGTERM')});
+	}
+
+	private runDiarization(file: string): AbortablePromise<any> {
+		let aborted = false, abort = () => { aborted = true; };
+		const checkPython = (cmd: string) => {
+			return new Promise<boolean>((resolve) => {
+				const proc = spawn(cmd, ['-c', 'import pyannote.audio']);
+				proc.on('close', (code: number) => resolve(code === 0));
+				proc.on('error', () => resolve(false));
+			});
+		};
+		const p = Promise.all<any>([
+			checkPython('python'),
+			checkPython('python3'),
+		]).then(<any>(async ([p, p3]: [boolean, boolean]) => {
 			if(aborted) return;
-				if(error) reject(new Error(error));
+			if(!p && !p3) throw new Error('Pyannote is not installed: pip install pyannote.audio');
-				else {
+			const binary = p3 ? 'python3' : 'python';
-					if(warning) console.warn(warning);
+			return new Promise((resolve, reject) => {
-					resolve(text);
+				if(aborted) return;
+				let output = '';
+				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) => {
+					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}`));
 					}
-			};
-			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({file, model, speaker, modelDir: this.ai.options.path});
+				proc.on('error', reject);
+				abort = () => proc.kill('SIGTERM');
 			});
-		return Object.assign(p, { abort });
+		}));
+		return <any>Object.assign(p, {abort});
 	}
 
-	canDiarization = canDiarization;
+	private combineSpeakerTranscript(transcript: 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 = '';
+		transcript.transcription.forEach((word: any) => {
+			const time = word.offsets.from / 1000; // Convert ms to seconds
+			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 = word.text;
+			} else {
+				currentText += ' ' + word.text;
+			}
+		});
+		if(currentText) lines.push(`[Speaker ${currentSpeaker}]: ${currentText.trim()}`);
+		return lines.join('\n');
+	}
+
+	asr(file: string, options: { model?: string; diarization?: boolean | 'id' } = {}): AbortablePromise<string | null> {
+		if(!this.ai.options.whisper) throw new Error('Whisper not configured');
+
+		const tmp = join(mkdtempSync(join(tmpdir(), 'audio-')), 'converted.wav');
+		execSync(`ffmpeg -i "${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(aborted || !options.diarization) return t;
+			t = this.combineSpeakerTranscript(t, d);
+			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)];
+				const names = await this.ai.language.json(chunks.join('\n'), '{1: "Detected Name", 2: "Second Name"}', {
+					system: 'Use the following transcript to identify speakers. Only identify speakers you are positive about, dont mention speakers you are unsure about in your response',
+					temperature: 0.1,
+				});
+				Object.entries(names).forEach(([speaker, name]) => t = t.replaceAll(`[Speaker ${speaker}]`, `[${name}]`));
+			}
+			return t;
+		}).finally(() => clean());
+		return <any>Object.assign(response, {abort});
+	}
+
+	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
+		if(!this.ai.options.whisper) throw new Error('Whisper not configured');
+		if(!model.endsWith('.bin')) model += '.bin';
+		const p = Path.join(<string>this.ai.options.path, model);
+		if(await fs.stat(p).then(() => true).catch(() => false)) return p;
+		if(!!this.downloads[model]) return this.downloads[model];
+		this.downloads[model] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${model}`)
+			.then(resp => resp.arrayBuffer())
+			.then(arr => Buffer.from(arr)).then(async buffer => {
+				await fs.writeFile(p, buffer);
+				delete this.downloads[model];
+				return p;
+			});
+		return this.downloads[model];
+	}
 }

src/embedder.ts

@@ -1,11 +1,13 @@
 import { pipeline } from '@xenova/transformers';
-import { parentPort } from 'worker_threads';
 
-let embedder: any;
+const [modelDir, model] = process.argv.slice(2);
 
-parentPort?.on('message', async ({ id, text, model, modelDir }) => {
+let text = '';
-	if(!embedder) embedder = await pipeline('feature-extraction', 'Xenova/' + model, {quantized: true, cache_dir: modelDir});
+process.stdin.on('data', chunk => text += chunk);
+process.stdin.on('end', async () => {
+	const embedder = await pipeline('feature-extraction', 'Xenova/' + model, {quantized: true, cache_dir: modelDir});
 	const output = await embedder(text, { pooling: 'mean', normalize: true });
 	const embedding = Array.from(output.data);
-	parentPort?.postMessage({ id, embedding });
+	console.log(JSON.stringify({embedding}));
+	process.exit();
 });

src/index.ts

@@ -1,8 +1,6 @@
 export * from './ai';
 export * from './antrhopic';
-export * from './asr';
 export * from './audio';
-export * from './embedder'
 export * from './llm';
 export * from './open-ai';
 export * from './provider';

src/llm.ts

@@ -4,9 +4,9 @@ import {Anthropic} from './antrhopic.ts';
 import {OpenAi} from './open-ai.ts';
 import {LLMProvider} from './provider.ts';
 import {AiTool} from './tools.ts';
-import {Worker} from 'worker_threads';
 import {fileURLToPath} from 'url';
 import {dirname, join} from 'path';
+import { spawn } from 'node:child_process';
 
 export type AnthropicConfig = {proto: 'anthropic', token: string};
 export type OllamaConfig = {proto: 'ollama', host: string};
@@ -75,8 +75,8 @@ export type LLMRequest = {
 }
 
 class LLM {
-	private models: {[model: string]: LLMProvider} = {};
+	defaultModel!: string;
-	private defaultModel!: string;
+	models: {[model: string]: LLMProvider} = {};
 
 	constructor(public readonly ai: Ai) {
 		if(!ai.options.llm?.models) return;
@@ -184,7 +184,12 @@ class LLM {
 		const system = history[0].role == 'system' ? history[0] : null,
 			recent = keep == 0 ? [] : history.slice(-keep),
 			process = (keep == 0 ? history : history.slice(0, -keep)).filter(h => h.role === 'assistant' || h.role === 'user');
-		const summary: any = await this.json(`Create the smallest summary possible, no more than 500 tokens. Create a list of NEW facts (split by subject [pro]noun and fact) about what you learned from this conversation that you didn't already know or get from a tool call or system prompt. Focus only on new information about people, topics, or facts. Avoid generating facts about the AI. Match this format: {summary: string, facts: [[subject, fact]]}\n\n${process.map(m => `${m.role}: ${m.content}`).join('\n\n')}`, {model: options?.model, temperature: options?.temperature || 0.3});
+
+		const summary: any = await this.json(process.map(m => `${m.role}: ${m.content}`).join('\n\n'), '{summary: string, facts: [[subject, fact]]}', {
+			system: 'Create the smallest summary possible, no more than 500 tokens. Create a list of NEW facts (split by subject [pro]noun and fact) about what you learned from this conversation that you didn\'t already know or get from a tool call or system prompt. Focus only on new information about people, topics, or facts. Avoid generating facts about the AI.',
+			model: options?.model,
+			temperature: options?.temperature || 0.3
+		});
 		const timestamp = new Date();
 		const memory = await Promise.all((summary?.facts || [])?.map(async ([owner, fact]: [string, string]) => {
 			const e = await Promise.all([this.embedding(owner), this.embedding(`${owner}: ${fact}`)]);
@@ -250,37 +255,57 @@ class LLM {
 	/**
 	 * Create a vector representation of a string
 	 * @param {object | string} target Item that will be embedded (objects get converted)
-	 * @param {number} maxTokens Chunking size. More = better context, less = more specific (Search by paragraphs or lines)
+	 * @param {maxTokens?: number, overlapTokens?: number} opts Options for embedding such as chunk sizes
-	 * @param {number} overlapTokens Includes previous X tokens to provide continuity to AI (In addition to max tokens)
 	 * @returns {Promise<Awaited<{index: number, embedding: number[], text: string, tokens: number}>[]>} Chunked embeddings
 	 */
-	embedding(target: object | string, maxTokens = 500, overlapTokens = 50) {
+	embedding(target: object | string, opts: {maxTokens?: number, overlapTokens?: number} = {}): AbortablePromise<any[]> {
+		let {maxTokens = 500, overlapTokens = 50} = opts;
+		let aborted = false;
+		const abort = () => { aborted = true; };
+
 		const embed = (text: string): Promise<number[]> => {
 			return new Promise((resolve, reject) => {
-				const worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
+				if(aborted) return reject(new Error('Aborted'));
-				const handleMessage = ({ embedding }: any) => {
+
-					worker.terminate();
+				const args: string[] = [
-					resolve(embedding);
+					join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'),
-				};
+					<string>this.ai.options.path,
-				const handleError = (err: Error) => {
+					this.ai.options?.embedder || 'bge-small-en-v1.5'
-					worker.terminate();
+				];
-					reject(err);
+				const proc = spawn('node', args, {stdio: ['pipe', 'pipe', 'ignore']});
-				};
+				proc.stdin.write(text);
-				worker.on('message', handleMessage);
+				proc.stdin.end();
-				worker.on('error', handleError);
+
-				worker.on('exit', (code) => {
+				let output = '';
-					if(code !== 0) reject(new Error(`Worker exited with code ${code}`));
+				proc.stdout.on('data', (data: Buffer) => output += data.toString());
+				proc.on('close', (code: number) => {
+					if(aborted) return reject(new Error('Aborted'));
+					if(code === 0) {
+						try {
+							const result = JSON.parse(output);
+							resolve(result.embedding);
+						} catch(err) {
+							reject(new Error('Failed to parse embedding output'));
+						}
+					} else {
+						reject(new Error(`Embedder process exited with code ${code}`));
+					}
 				});
-				worker.postMessage({text, model: this.ai.options?.embedder || 'bge-small-en-v1.5', path: this.ai.options.path});
+				proc.on('error', reject);
 			});
 		};
-		const chunks = this.chunk(target, maxTokens, overlapTokens);
+
-		return Promise.all(chunks.map(async (text, index) => ({
+		const p = (async () => {
-			index,
+			const chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
-			embedding: await embed(text),
+			for(let i = 0; i < chunks.length; i++) {
-			text,
+				if(aborted) break;
-			tokens: this.estimateTokens(text),
+				const text = chunks[i];
-		})));
+				const embedding = await embed(text);
+				results.push({index: i, embedding, text, tokens: this.estimateTokens(text)});
+			}
+			return results;
+		})();
+		return Object.assign(p, { abort });
 	}
 
 	/**
@@ -312,12 +337,16 @@ class LLM {
 
 	/**
 	 * Ask a question with JSON response
-	 * @param {string} message Question
+	 * @param {string} text Text to process
+	 * @param {string} schema JSON schema the AI should match
 	 * @param {LLMRequest} options Configuration options and chat history
 	 * @returns {Promise<{} | {} | RegExpExecArray | null>}
 	 */
-	async json(message: string, options?: LLMRequest): Promise<any> {
+	async json(text: string, schema: string, options?: LLMRequest): Promise<any> {
-		let resp = await this.ask(message, {system: 'Respond using a JSON blob matching any provided examples', ...options});
+		let resp = await this.ask(text, {...options, system: (options?.system ? `${options.system}\n` : '') + `Only respond using a JSON code block matching this schema:
+\`\`\`json
+${schema}
+\`\`\``});
 		if(!resp) return {};
 		const codeBlock = /```(?:.+)?\s*([\s\S]*?)```/.exec(resp);
 		const jsonStr = codeBlock ? codeBlock[1].trim() : resp;

src/vision.ts

@@ -3,7 +3,7 @@ import {AbortablePromise, Ai} from './ai.ts';
 
 export class Vision {
 
-	constructor(private ai: Ai) { }
+	constructor(private ai: Ai) {}
 
 	/**
 	 * Convert image to text using Optical Character Recognition

vite.config.ts

@@ -1,12 +1,10 @@
 import {defineConfig} from 'vite';
 import dts from 'vite-plugin-dts';
-import {resolve} from 'path';
 
 export default defineConfig({
 	build: {
 		lib: {
 			entry: {
-				asr: './src/asr.ts',
 				index: './src/index.ts',
 				embedder: './src/embedder.ts',
 			},