Queue OCR & ASR work

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.6.5",
+	"version": "0.7.7",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -25,14 +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": {

src/ai.ts

@@ -8,6 +8,8 @@ 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 */

src/asr.ts

@@ -0,0 +1,137 @@
+import { pipeline } from '@xenova/transformers';
+import { parentPort } from 'worker_threads';
+import { spawn } from 'node:child_process';
+import { execSync } from 'node:child_process';
+import { mkdtempSync, rmSync, readFileSync } from 'node:fs';
+import { join } from 'node:path';
+import { tmpdir } from 'node:os';
+import wavefile from 'wavefile';
+
+let whisperPipeline: any;
+
+export async function canDiarization(): Promise<string | null> {
+	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));
+		});
+	};
+	if(await checkPython('python3')) return 'python3';
+	if(await checkPython('python')) return 'python';
+	return null;
+}
+
+async function runDiarization(binary: string, audioPath: string, dir: string, token: string): Promise<any[]> {
+	const script = `
+import sys
+import json
+import os
+from pyannote.audio import Pipeline
+
+os.environ['TORCH_HOME'] = r"${dir}"
+pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", token="${token}")
+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))
+`;
+
+	return new Promise((resolve, reject) => {
+		let output = '';
+		const proc = spawn(binary, ['-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}]: ${currentText.trim()}`);
+	return lines.join('\n');
+}
+
+function prepareAudioBuffer(file: string): [string, Float32Array] {
+	let wav: any, tmp;
+	try {
+		wav = new wavefile.WaveFile(readFileSync(file));
+	} catch(err) {
+		tmp = join(mkdtempSync(join(tmpdir(), 'audio-')), 'converted.wav');
+		execSync(`ffmpeg -i "${file}" -ar 16000 -ac 1 -f wav "${tmp}"`, { stdio: 'ignore' });
+		wav = new wavefile.WaveFile(readFileSync(tmp));
+	} finally {
+		wav.toBitDepth('32f');
+		wav.toSampleRate(16000);
+		const samples = wav.getSamples();
+		if(Array.isArray(samples)) {
+			const left = samples[0];
+			const right = samples[1];
+			const buffer = new Float32Array(left.length);
+			for (let i = 0; i < left.length; i++) buffer[i] = (left[i] + right[i]) / 2;
+			return [tmp || file, buffer];
+		}
+		return [tmp || file, samples];
+	}
+}
+
+parentPort?.on('message', async ({ file, speaker, model, modelDir, token }) => {
+	let tempFile = null;
+	try {
+		if(!whisperPipeline) whisperPipeline = await pipeline('automatic-speech-recognition', `Xenova/${model}`, {cache_dir: modelDir, quantized: true});
+
+		const [f, buffer] = prepareAudioBuffer(file);
+		tempFile = f !== file ? f : null;
+		const hasDiarization = await canDiarization();
+		const [transcript, speakers] = await Promise.all([
+			whisperPipeline(buffer, {return_timestamps: speaker ? 'word' : false}),
+			(!speaker || !token || !hasDiarization) ? Promise.resolve(): runDiarization(hasDiarization, f, modelDir, token),
+		]);
+
+		const text = transcript.text?.trim() || null;
+		if(!speaker) return parentPort?.postMessage({ text });
+		if(!token) return parentPort?.postMessage({ text, error: 'HuggingFace token required' });
+		if(!hasDiarization) return parentPort?.postMessage({ text, error: 'Speaker diarization unavailable' });
+
+		const combined = combineSpeakerTranscript(transcript.chunks || [], speakers || []);
+		parentPort?.postMessage({ text: combined });
+	} catch (err: any) {
+		parentPort?.postMessage({ error: err.stack || err.message });
+	} finally {
+		if(tempFile) rmSync(tempFile, { recursive: true, force: true });
+	}
+});

src/audio.ts

@@ -1,133 +1,82 @@
-import {spawn} from 'node:child_process';
+import {fileURLToPath} from 'url';
-import {pipeline} from '@xenova/transformers';
+import {Worker} from 'worker_threads';
-import * as fs from 'node:fs';
 import {AbortablePromise, Ai} from './ai.ts';
-import wavefile from 'wavefile';
+import {canDiarization} from './asr.ts';
+import {dirname, join} from 'path';
 
 export class Audio {
-	private whisperPipeline: any;
+	private busy = false;
+	private currentJob: any;
+	private queue: Array<{file: string, model: string, speaker: boolean | 'id', modelDir: string, token: string, resolve: any, reject: any}> = [];
+	private worker: Worker | null = null;
 
 	constructor(private ai: Ai) {}
 
-	private combineSpeakerTranscript(chunks: any[], speakers: any[]): string {
+	private processQueue() {
-		const speakerMap = new Map();
+		if(this.busy || !this.queue.length) return;
-		let speakerCount = 0;
-		speakers.forEach((seg: any) => {
-			if(!speakerMap.has(seg.speaker)) speakerMap.set(seg.speaker, ++speakerCount);
-		});
 
-		const lines: string[] = [];
+		this.busy = true;
-		let currentSpeaker = -1;
+		const job = this.queue.shift()!;
-		let currentText = '';
+		if(!this.worker) {
-		chunks.forEach((chunk: any) => {
+			this.worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'asr.js'));
-			const time = chunk.timestamp[0];
+			this.worker.on('message', this.handleMessage.bind(this));
-			const speaker = speakers.find((s: any) => time >= s.start && time <= s.end);
+			this.worker.on('error', this.handleError.bind(this));
-			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> {
+		this.currentJob = job;
-		return new Promise((resolve) => {
+		this.worker.postMessage({file: job.file, model: job.model, speaker: job.speaker, modelDir: job.modelDir, token: job.token});
-			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[]> {
+	private handleMessage({text, warning, error}: any) {
-		if(!await this.canDiarization()) throw new Error('Pyannote is not installed: pip install pyannote.audio');
+		const job = this.currentJob!;
-		const script = `
+		this.busy = false;
-import sys
+		if(error) job.reject(new Error(error));
-import json
+		else {
-from pyannote.audio import Pipeline
+			if(warning) console.warn(warning);
-
+			job.resolve(text);
-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 {
+		this.processQueue();
-					reject(new Error(`Python process exited with code ${code}`));
-				}
-			});
-
-			proc.on('error', reject);
-		});
 	}
 
-	asr(path: string, options: { model?: string; speaker?: boolean } = {}): AbortablePromise<string | null> {
+	private handleError(err: Error) {
+		if(this.currentJob) {
+			this.currentJob.reject(err);
+			this.busy = false;
+			this.processQueue();
+		}
+	}
+
+	asr(file: string, options: { model?: string; speaker?: boolean | 'id' } = {}): AbortablePromise<string | null> {
 		const { model = this.ai.options.asr || 'whisper-base', speaker = false } = options;
 		let aborted = false;
 		const abort = () => { aborted = true; };
-
+		let p = new Promise<string | null>((resolve, reject) => {
-		const p = new Promise<string | null>(async (resolve, reject) => {
+			this.queue.push({file, model, speaker, modelDir: <string>this.ai.options.path, token: <string>this.ai.options.hfToken,
-			try {
+				resolve: (text: string | null) => !aborted && resolve(text),
-				if(aborted) return resolve(null);
+				reject: (err: Error) => !aborted && reject(err)
-				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;
-				}
-
-				// 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);
-			}
 			});
+			this.processQueue();
+		});
+
+		if(options.speaker == 'id') {
+			if(!this.ai.language.defaultModel) throw new Error('Configure an LLM for advanced ASR speaker detection');
+			p = p.then(async transcript => {
+				if(!transcript) return transcript;
+				let chunks = this.ai.language.chunk(transcript, 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]) => {
+					transcript = (<string>transcript).replaceAll(`[Speaker ${speaker}]`, `[${name}]`);
+				});
+				return transcript;
+			})
+		}
 
 		return Object.assign(p, { abort });
 	}
+
+	canDiarization = () => canDiarization().then(resp => !!resp);
 }

src/embedder.ts

@@ -3,12 +3,9 @@ import { parentPort } from 'worker_threads';
 
 let embedder: any;
 
-parentPort?.on('message', async ({ id, text, model, path }) => {
+parentPort?.on('message', async ({text, model, modelDir }) => {
-	if(!embedder) embedder = await pipeline('feature-extraction', 'Xenova/' + model, {
+	if(!embedder) embedder = await pipeline('feature-extraction', 'Xenova/' + model, {quantized: true, cache_dir: modelDir});
-		quantized: true,
-		cache_dir: path,
-	});
 	const output = await embedder(text, { pooling: 'mean', normalize: true });
 	const embedding = Array.from(output.data);
-	parentPort?.postMessage({ id, embedding });
+	parentPort?.postMessage({embedding});
 });

src/index.ts

@@ -1,5 +1,6 @@
 export * from './ai';
 export * from './antrhopic';
+export * from './asr';
 export * from './audio';
 export * from './embedder'
 export * from './llm';

src/llm.ts

@@ -75,22 +75,10 @@ export type LLMRequest = {
 }
 
 class LLM {
-	private embedWorker: Worker | null = null;
+	defaultModel!: string;
-	private embedQueue = new Map<number, { resolve: (value: number[]) => void; reject: (error: any) => void }>();
+	models: {[model: string]: LLMProvider} = {};
-	private embedId = 0;
-	private models: {[model: string]: LLMProvider} = {};
-	private defaultModel!: string;
 
 	constructor(public readonly ai: Ai) {
-		this.embedWorker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
-		this.embedWorker.on('message', ({ id, embedding }) => {
-			const pending = this.embedQueue.get(id);
-			if (pending) {
-				pending.resolve(embedding);
-				this.embedQueue.delete(id);
-			}
-		});
-
 		if(!ai.options.llm?.models) return;
 		Object.entries(ai.options.llm.models).forEach(([model, config]) => {
 			if(!this.defaultModel) this.defaultModel = model;
@@ -196,7 +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}`)]);
@@ -262,30 +255,37 @@ 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) {
+	async embedding(target: object | string, opts: {maxTokens?: number, overlapTokens?: number} = {}) {
+		let {maxTokens = 500, overlapTokens = 50} = opts;
 		const embed = (text: string): Promise<number[]> => {
 			return new Promise((resolve, reject) => {
-				const id = this.embedId++;
+				const worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
-				this.embedQueue.set(id, { resolve, reject });
+				const handleMessage = ({ embedding }: any) => {
-				this.embedWorker?.postMessage({
+					worker.terminate();
-					id,
+					resolve(embedding);
-					text,
+				};
-					model: this.ai.options?.embedder || 'bge-small-en-v1.5',
+				const handleError = (err: Error) => {
-					path: this.ai.options.path
+					worker.terminate();
+					reject(err);
+				};
+				worker.on('message', handleMessage);
+				worker.on('error', handleError);
+				worker.on('exit', (code) => {
+					if(code !== 0) reject(new Error(`Worker exited with code ${code}`));
 				});
+				worker.postMessage({text, model: this.ai.options?.embedder || 'bge-small-en-v1.5', modelDir: this.ai.options.path});
 			});
 		};
-		const chunks = this.chunk(target, maxTokens, overlapTokens);
+		const chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
-		return Promise.all(chunks.map(async (text, index) => ({
+		for(let i = 0; i < chunks.length; i++) {
-			index,
+			const text= chunks[i];
-			embedding: await embed(text),
+			const embedding = await embed(text);
-			text,
+			results.push({index: i, embedding, text, tokens: this.estimateTokens(text)});
-			tokens: this.estimateTokens(text),
+		}
-		})));
+		return results;
 	}
 
 	/**
@@ -317,12 +317,16 @@ 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

@@ -2,22 +2,43 @@ import {createWorker} from 'tesseract.js';
 import {AbortablePromise, Ai} from './ai.ts';
 
 export class Vision {
+	private worker: any = null;
+	private queue: Array<{ path: string, resolve: any, reject: any }> = [];
+	private busy = false;
 
 	constructor(private ai: Ai) {}
 
+	private async processQueue() {
+		if(this.busy || !this.queue.length) return;
+		this.busy = true;
+		const job = this.queue.shift()!;
+		if(!this.worker) this.worker = await createWorker(this.ai.options.ocr || 'eng', 2, {cachePath: this.ai.options.path});
+		try {
+			const {data} = await this.worker.recognize(job.path);
+			job.resolve(data.text.trim() || null);
+		} catch(err) {
+			job.reject(err);
+		}
+		this.busy = false;
+		this.processQueue();
+	}
+
 	/**
 	 * Convert image to text using Optical Character Recognition
 	 * @param {string} path Path to image
 	 * @returns {AbortablePromise<string | null>} Promise of extracted text with abort method
 	 */
 	ocr(path: string): AbortablePromise<string | null> {
-		let worker: any;
+		let aborted = false;
-		const p = new Promise<string | null>(async res => {
+		const abort = () => { aborted = true; };
-			worker = await createWorker(this.ai.options.ocr || 'eng', 2, {cachePath: this.ai.options.path});
+		const p = new Promise<string | null>((resolve, reject) => {
-			const {data} = await worker.recognize(path);
+			this.queue.push({
-			await worker.terminate();
+				path,
-			res(data.text.trim() || null);
+				resolve: (text: string | null) => !aborted && resolve(text),
+				reject: (err: Error) => !aborted && reject(err)
 			});
-		return Object.assign(p, {abort: () => worker?.terminate()});
+			this.processQueue();
+		});
+		return Object.assign(p, {abort});
 	}
 }

vite.config.ts

@@ -1,11 +1,11 @@
 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',
 			},