TTS

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.2.7",
+	"version": "0.3.0",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -42,6 +42,7 @@
 		"vite-plugin-dts": "^4.5.3"
 	},
 	"files": [
+		"bin",
 		"dist"
 	]
 }

src/ai.ts

@@ -3,9 +3,21 @@ import {LLM, LLMOptions} from './llm';
 import { Audio } from './audio.ts';
 import {Vision} from './vision.ts';
 
+export type AbortablePromise<T> = Promise<T> & {abort: () => any};
+
 export type AiOptions = LLMOptions & {
 	/** Path to models */
 	path?: string;
+	/** Piper TTS configuratoin */
+	piper?: {
+		/** Model URL: `https://huggingface.co/rhasspy/piper-voices/tree/main/.../model.onnx` */
+		model: string;
+	},
+	/** Tesseract OCR configuration */
+	tesseract?: {
+		/** Model: eng, eng_best, eng_fast */
+		model?: string;
+	}
 	/** Whisper ASR configuration */
 	whisper?: {
 		/** Whisper binary location */
@@ -13,11 +25,6 @@ export type AiOptions = LLMOptions & {
 		/** Model: `ggml-base.en.bin` */
 		model: string;
 	}
-	/** Tesseract OCR configuration */
-	tesseract?: {
-		/** Model: eng, eng_best, eng_fast */
-		model?: string;
-	}
 }
 
 export class Ai {

src/antrhopic.ts

@@ -1,8 +1,8 @@
 import {Anthropic as anthropic} from '@anthropic-ai/sdk';
 import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, deepCopy} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
 
 export class Anthropic extends LLMProvider {
 	client!: anthropic;
@@ -120,6 +120,7 @@ export class Anthropic extends LLMProvider {
 					original.push({role: 'assistant', content: resp.content});
 					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
 						const tool = tools.find(findByProp('name', toolCall.name));
+						if(options.stream) options.stream({tool: toolCall.name});
 						if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
 						try {
 							const result = await tool.fn(toolCall.input, this.ai);

src/audio.ts

@@ -1,30 +1,35 @@
 import {spawn} from 'node:child_process';
+import * as os from 'node:os';
+import {platform, arch} from 'node:os';
 import fs from 'node:fs/promises';
 import Path from 'node:path';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 
 export class Audio {
 	private downloads: {[key: string]: Promise<string>} = {};
 	private whisperModel!: string;
+	private piperBinary?: string;
 
 	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';
 			this.downloadAsrModel();
 		}
+		if(ai.options.piper?.model) {
+			if(!ai.options.piper.model.startsWith('http') || !ai.options.piper.model.endsWith('.onnx'))
+				throw new Error('Piper model should be a URL to an onnx file to download');
+			if(platform() != 'linux' || (arch() != 'x64' && arch() != 'arm64'))
+				throw new Error('Piper TTS only supported on Linux x64/arm64');
+			this.piperBinary = Path.join(import.meta.dirname, '../bin/piper');
+			this.downloadTtsModel();
+		}
 	}
 
-	/**
+	asr(path: string, model: string = this.whisperModel): AbortablePromise<string | null> {
-	 * Convert audio to text using Auditory Speech Recognition
-	 * @param {string} path Path to audio
-	 * @param model Whisper model
-	 * @returns {Promise<any>} Extracted text
-	 */
-	asr(path: string, model: string = this.whisperModel): {abort: () => void, response: Promise<string | null>} {
 		if(!this.ai.options.whisper?.binary) throw new Error('Whisper not configured');
 		let abort: any = () => {};
-		const response = new Promise<string | null>((resolve, reject) => {
+		const p = new Promise<string | null>(async (resolve, reject) => {
-			this.downloadAsrModel(model).then(m => {
+			const m = await this.downloadAsrModel(model);
 			let output = '';
 			const proc = spawn(<string>this.ai.options.whisper?.binary, ['-nt', '-np', '-m', m, '-f', path], {stdio: ['ignore', 'pipe', 'ignore']});
 			abort = () => proc.kill('SIGTERM');
@@ -35,16 +40,41 @@ export class Audio {
 				else reject(new Error(`Exit code ${code}`));
 			});
 		});
-		});
+		return Object.assign(p, {abort});
-		return {response, abort};
+	}
+
+	tts(text: string, outputPath?: string, model: string = <string>this.ai.options.piper?.model): AbortablePromise<Buffer | string> {
+		if(!this.piperBinary) throw new Error('Piper not configured');
+		if(!model) throw new Error('Invalid Piper model');
+		let abort: any = () => {};
+		const p = new Promise<Buffer | string>(async (resolve, reject) => {
+			const modelPath = await this.downloadTtsModel(model);
+			const tmpFile = outputPath || Path.join(os.tmpdir(), `piper_${Date.now()}.wav`);
+			const proc = spawn(<string>this.piperBinary, ['--model', modelPath, '--output_file', tmpFile], {
+				stdio: ['pipe', 'ignore', 'ignore'],
+				env: {...process.env, LD_LIBRARY_PATH: Path.dirname(<string>this.piperBinary)}
+			});
+			abort = () => proc.kill('SIGTERM');
+			proc.stdin.write(text);
+			proc.stdin.end();
+			proc.on('error', (err: Error) => reject(err));
+			proc.on('close', async (code: number) => {
+				if(code === 0) {
+					if(outputPath) {
+						resolve(outputPath);
+					} else {
+						const buffer = await fs.readFile(tmpFile);
+						await fs.unlink(tmpFile).catch(() => {});
+						resolve(buffer);
+					}
+				} else {
+					reject(new Error(`Exit code ${code}`));
+				}
+			});
+		});
+		return Object.assign(p, {abort});
 	}
 
-	/**
-	 * Downloads the specified Whisper model if it is not already present locally.
-	 *
-	 * @param {string} model Whisper model that will be downloaded
-	 * @return {Promise<string>} Absolute path to model file, resolves once downloaded
-	 */
 	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';
@@ -60,4 +90,24 @@ export class Audio {
 			});
 		return this.downloads[model];
 	}
+
+	async downloadTtsModel(model: string = <string>this.ai.options.piper?.model): Promise<string> {
+		if(!model) throw new Error('Invalid Piper model');
+		const m = <string>model.split('/').pop();
+		const p = Path.join(<string>this.ai.options.path, m);
+		const [onnxExists, jsonExists] = await Promise.all([
+			fs.stat(p).then(() => true).catch(() => false),
+			fs.stat(p + '.json').then(() => true).catch(() => false)
+		]);
+		if(onnxExists && jsonExists) return p;
+		if(!!this.downloads[m]) return this.downloads[m];
+		this.downloads[m] = Promise.all([
+			onnxExists ? Promise.resolve() : fetch(model).then(r => r.arrayBuffer()).then(b => fs.writeFile(p, Buffer.from(b))),
+			jsonExists ? Promise.resolve() : fetch(model + '.json').then(r => r.arrayBuffer()).then(b => fs.writeFile(p + '.json', Buffer.from(b)))
+		]).then(() => {
+			delete this.downloads[m];
+			return p;
+		});
+		return this.downloads[m];
+	}
 }

src/llm.ts

@@ -1,9 +1,9 @@
 import {JSONAttemptParse} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
 import {Ollama} from './ollama.ts';
 import {OpenAi} from './open-ai.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
 import {AiTool} from './tools.ts';
 import {Worker} from 'worker_threads';
 import {fileURLToPath} from 'url';
@@ -73,7 +73,7 @@ export type LLMRequest = {
 	/** LLM  model */
 	model?: string | [string, string];
 	/** Stream response */
-	stream?: (chunk: {text?: string, done?: true}) => any;
+	stream?: (chunk: {text?: string, tool?: string, done?: true}) => any;
 	/** Compress old messages in the chat to free up context */
 	compress?: {
 		/** Trigger chat compression once context exceeds the token count */

src/ollama.ts

@@ -1,7 +1,7 @@
 import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
 import {Ollama as ollama} from 'ollama';
 
 export class Ollama extends LLMProvider {
@@ -99,6 +99,7 @@ export class Ollama extends LLMProvider {
 					history.push(resp.message);
 					const results = await Promise.all(resp.message.tool_calls.map(async (toolCall: any) => {
 						const tool = tools.find(findByProp('name', toolCall.function.name));
+						if(options.stream) options.stream({tool: toolCall.function.name});
 						if(!tool) return {role: 'tool', tool_name: toolCall.function.name, content: '{"error": "Tool not found"}'};
 						const args = typeof toolCall.function.arguments === 'string' ? JSONAttemptParse(toolCall.function.arguments, {}) : toolCall.function.arguments;
 						try {

src/open-ai.ts

@@ -1,8 +1,8 @@
 import {OpenAI as openAI} from 'openai';
 import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
 
 export class OpenAi extends LLMProvider {
 	client!: openAI;
@@ -116,6 +116,7 @@ export class OpenAi extends LLMProvider {
 					history.push(resp.choices[0].message);
 					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
 						const tool = tools?.find(findByProp('name', toolCall.function.name));
+						if(options.stream) options.stream({tool: toolCall.function.name});
 						if(!tool) return {role: 'tool', tool_call_id: toolCall.id, content: '{"error": "Tool not found"}'};
 						try {
 							const args = JSONAttemptParse(toolCall.function.arguments, {});

src/provider.ts

@@ -1,7 +1,6 @@
+import {AbortablePromise} from './ai.ts';
 import {LLMMessage, LLMOptions, LLMRequest} from './llm.ts';
 
-export type AbortablePromise<T> = Promise<T> & {abort: () => void};
-
 export abstract class LLMProvider {
 	abstract ask(message: string, options: LLMRequest): AbortablePromise<LLMMessage[]>;
 }

src/vision.ts

@@ -1,5 +1,5 @@
 import {createWorker} from 'tesseract.js';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 
 export class Vision {
 
@@ -8,18 +8,16 @@ export class Vision {
 	/**
 	 * Convert image to text using Optical Character Recognition
 	 * @param {string} path Path to image
-	 * @returns {{abort: Function, response: Promise<string | null>}} Abort function & Promise of extracted text
+	 * @returns {AbortablePromise<string | null>} Promise of extracted text with abort method
 	 */
-	ocr(path: string): {abort: () => void, response: Promise<string | null>} {
+	ocr(path: string): AbortablePromise<string | null> {
 		let worker: any;
-		return {
+		const p = new Promise<string | null>(async res => {
-			abort: () => { worker?.terminate(); },
-			response: new Promise(async res => {
 			worker = await createWorker(this.ai.options.tesseract?.model || 'eng', 2, {cachePath: this.ai.options.path});
 			const {data} = await worker.recognize(path);
 			await worker.terminate();
 			res(data.text.trim() || null);
-			})
+		});
-		}
+		return Object.assign(p, {abort: () => worker?.terminate()});
 	}
 }