expose diarization support checking function

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.5.1",
+	"version": "0.6.1",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",

src/ai.ts

@@ -10,22 +10,16 @@ export type AbortablePromise<T> = Promise<T> & {
 export type AiOptions = {
 	/** Path to models */
 	path?: string;
+	/** ASR model: whisper-tiny, whisper-base */
+	asr?: string;
+	/** Embedding model: all-MiniLM-L6-v2, bge-small-en-v1.5, bge-large-en-v1.5 */
+	embedder?: string;
 	/** Large language models, first is default */
 	llm?: Omit<LLMRequest, 'model'> & {
 		models: {[model: string]: AnthropicConfig | OllamaConfig | OpenAiConfig};
 	}
-	/** Tesseract OCR configuration */
+	/** OCR model: eng, eng_best, eng_fast */
-	tesseract?: {
+	ocr?: string;
-		/** Model: eng, eng_best, eng_fast */
-		model?: string;
-	}
-	/** Whisper ASR configuration */
-	whisper?: {
-		/** Whisper binary location */
-		binary: string;
-		/** Model: `ggml-base.en.bin` */
-		model: string;
-	}
 }
 
 export class Ai {

src/antrhopic.ts

@@ -13,25 +13,25 @@ export class Anthropic extends LLMProvider {
 	}
 
 	private toStandard(history: any[]): LLMMessage[] {
-		for(let i = 0; i < history.length; i++) {
+		const timestamp = Date.now();
-			const orgI = i;
+		const messages: LLMMessage[] = [];
-			if(typeof history[orgI].content != 'string') {
+		for(let h of history) {
-				if(history[orgI].role == 'assistant') {
+			if(typeof h.content == 'string') {
-					history[orgI].content.filter((c: any) => c.type =='tool_use').forEach((c: any) => {
+				messages.push(<any>{timestamp, ...h});
-						history.splice(i + 1, 0, {role: 'tool', id: c.id, name: c.name, args: c.input, timestamp: Date.now()});
+			} else {
-					});
+				const textContent = h.content?.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
-				} else if(history[orgI].role == 'user') {
+				if(textContent) messages.push({timestamp, role: h.role, content: textContent});
-					history[orgI].content.filter((c: any) => c.type =='tool_result').forEach((c: any) => {
+				h.content.forEach((c: any) => {
-						const h = history.find((h: any) => h.id == c.tool_use_id);
+					if(c.type == 'tool_use') {
-						h[c.is_error ? 'error' : 'content'] = c.content;
+						messages.push({timestamp, role: 'tool', id: c.id, name: c.name, args: c.input, content: undefined});
+					} else if(c.type == 'tool_result') {
+						const m: any = messages.findLast(m => (<any>m).id == c.tool_use_id);
+						if(m) m[c.is_error ? 'error' : 'content'] = c.content;
+					}
 				});
 			}
-				history[orgI].content = history[orgI].content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
-				if(!history[orgI].content) history.splice(orgI, 1);
 		}
-			if(!history[orgI].timestamp) history[orgI].timestamp = Date.now();
+		return messages;
-		}
-		return history.filter(h => !!h.content);
 	}
 
 	private fromStandard(history: LLMMessage[]): any[] {
@@ -50,8 +50,8 @@ export class Anthropic extends LLMProvider {
 
 	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
 		const controller = new AbortController();
-		return Object.assign(new Promise<any>(async (res, rej) => {
+		return Object.assign(new Promise<any>(async (res) => {
-			const history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
+			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
 			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
@@ -73,7 +73,6 @@ export class Anthropic extends LLMProvider {
 			};
 
 			let resp: any, isFirstMessage = true;
-			const assistantMessages: string[] = [];
 			do {
 				resp = await this.client.messages.create(requestParams).catch(err => {
 					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
@@ -119,7 +118,6 @@ export class Anthropic extends LLMProvider {
 						if(options.stream) options.stream({tool: toolCall.name});
 						if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
 						try {
-							console.log(typeof tool.fn);
 							const result = await tool.fn(toolCall.input, options?.stream, this.ai);
 							return {type: 'tool_result', tool_use_id: toolCall.id, content: JSONSanitize(result)};
 						} catch (err: any) {
@@ -131,7 +129,7 @@ export class Anthropic extends LLMProvider {
 				}
 			} while (!controller.signal.aborted && resp.content.some((c: any) => c.type === 'tool_use'));
 			history.push({role: 'assistant', content: resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n')});
-			this.toStandard(history);
+			history = this.toStandard(history);
 
 			if(options.stream) options.stream({done: true});
 			if(options.history) options.history.splice(0, options.history.length, ...history);

src/audio.ts

@@ -1,50 +1,115 @@
 import { spawn } from 'node:child_process';
-import fs from 'node:fs/promises';
+import { pipeline } from '@xenova/transformers';
-import Path from 'node:path';
 import { AbortablePromise, Ai } from './ai.ts';
 
 export class Audio {
-	private downloads: {[key: string]: Promise<string>} = {};
+	private whisperPipeline: any;
-	private whisperModel!: string;
 
-	constructor(private ai: Ai) {
+	constructor(private ai: Ai) {}
-		if(ai.options.whisper?.binary) {
+
-			this.whisperModel = ai.options.whisper?.model.endsWith('.bin') ? ai.options.whisper?.model : ai.options.whisper?.model + '.bin';
+	private combineSpeakerTranscript(chunks: any[], speakers: any[]): string {
-			this.downloadAsrModel();
+		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');
 	}
 
-	asr(path: string, model: string = this.whisperModel): AbortablePromise<string | null> {
+	async canDiarization(): Promise<boolean> {
-		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
+		return new Promise((resolve) => {
-		let abort: any = () => {};
+			const proc = spawn('python3', ['-c', 'import pyannote.audio']);
-		const p = new Promise<string | null>(async (resolve, reject) => {
+			proc.on('close', (code: number) => resolve(code === 0));
-			const m = await this.downloadAsrModel(model);
+			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(<string>this.ai.options.whisper?.binary, ['-nt', '-np', '-m', m, '-f', path], {stdio: ['ignore', 'pipe', 'ignore']});
+			const proc = spawn('python3', ['-c', script, audioPath]);
-			abort = () => proc.kill('SIGTERM');
-			proc.on('error', (err: Error) => reject(err));
 			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) resolve(output.trim() || null);
+				if(code === 0) {
-				else reject(new Error(`Exit code ${code}`));
+					try {
+						resolve(JSON.parse(output));
+					} catch (err) {
+						reject(new Error('Failed to parse diarization output'));
+					}
+				} else {
+					reject(new Error(`Python process exited with code ${code}`));
+				}
 			});
+
+			proc.on('error', reject);
 		});
+	}
+
+	asr(path: string, options: { model?: string; speaker?: boolean } = {}): AbortablePromise<string | null> {
+		const { model = this.ai.options.asr || 'whisper-base', speaker = false } = options;
+		let aborted = false;
+		const abort = () => { aborted = true; };
+
+		const p = new Promise<string | null>(async (resolve, reject) => {
+			try {
+				if(aborted) return resolve(null);
+				if(!this.whisperPipeline) this.whisperPipeline = await pipeline('automatic-speech-recognition', `Xenova/${model}`, { cache_dir: this.ai.options.path, quantized: true });
+
+				// Transcript
+				if(aborted) return resolve(null);
+				const transcriptResult = await this.whisperPipeline(path, {return_timestamps: speaker ? 'word' : false, chunk_length_s: 30,});
+				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);
+			}
+		});
+
 		return Object.assign(p, { abort });
 	}
-
-	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
-		if(!this.ai.options.whisper?.binary) 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 +1,14 @@
 import { pipeline } from '@xenova/transformers';
 import { parentPort } from 'worker_threads';
 
-let model: any;
+let embedder: any;
 
-parentPort?.on('message', async ({ id, text }) => {
+parentPort?.on('message', async ({ id, text, model, path }) => {
-	if(!model) model = await pipeline('feature-extraction', 'Xenova/all-MiniLM-L6-v2');
+	if(!embedder) embedder = await pipeline('feature-extraction', 'Xenova/' + model, {
-	const output = await model(text, { pooling: 'mean', normalize: true });
+		quantized: true,
+		cache_dir: path,
+	});
+	const output = await embedder(text, { pooling: 'mean', normalize: true });
 	const embedding = Array.from(output.data);
 	parentPort?.postMessage({ id, embedding });
 });

src/llm.ts

@@ -271,7 +271,12 @@ class LLM {
 			return new Promise((resolve, reject) => {
 				const id = this.embedId++;
 				this.embedQueue.set(id, { resolve, reject });
-				this.embedWorker?.postMessage({ id, text });
+				this.embedWorker?.postMessage({
+					id,
+					text,
+					model: this.ai.options?.embedder || 'bge-small-en-v1.5',
+					path: this.ai.options.path
+				});
 			});
 		};
 		const chunks = this.chunk(target, maxTokens, overlapTokens);

src/open-ai.ts

@@ -68,7 +68,7 @@ export class OpenAi extends LLMProvider {
 		const controller = new AbortController();
 		return Object.assign(new Promise<any>(async (res, rej) => {
 			if(options.system && options.history?.[0]?.role != 'system') options.history?.splice(0, 0, {role: 'system', content: options.system, timestamp: Date.now()});
-			const history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
+			let history = this.fromStandard([...options.history || [], {role: 'user', content: message, timestamp: Date.now()}]);
 			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
@@ -133,7 +133,7 @@ export class OpenAi extends LLMProvider {
 				}
 			} while (!controller.signal.aborted && resp.choices?.[0]?.message?.tool_calls?.length);
 			history.push({role: 'assistant', content: resp.choices[0].message.content || ''});
-			this.toStandard(history);
+			history = this.toStandard(history);
 
 			if(options.stream) options.stream({done: true});
 			if(options.history) options.history.splice(0, options.history.length, ...history);

src/vision.ts

@@ -13,7 +13,7 @@ export class Vision {
 	ocr(path: string): AbortablePromise<string | null> {
 		let worker: any;
 		const p = new Promise<string | null>(async res => {
-			worker = await createWorker(this.ai.options.tesseract?.model || 'eng', 2, {cachePath: this.ai.options.path});
+			worker = await createWorker(this.ai.options.ocr || 'eng', 2, {cachePath: this.ai.options.path});
 			const {data} = await worker.recognize(path);
 			await worker.terminate();
 			res(data.text.trim() || null);