Better ASR IDing

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.7.1",
+	"version": "0.7.5",
 	"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/asr.ts

@@ -9,15 +9,20 @@ import wavefile from 'wavefile';
 
 let whisperPipeline: any;
 
-export async function canDiarization(): Promise<boolean> {
+export async function canDiarization(): Promise<string | null> {
-	return new Promise((resolve) => {
+	const checkPython = (cmd: string) => {
-		const proc = spawn('python', ['-c', 'import pyannote.audio']);
+		return new Promise<boolean>((resolve) => {
-		proc.on('close', (code: number) => resolve(code === 0));
+			const proc = spawn(cmd, ['-c', 'import pyannote.audio']);
-		proc.on('error', () => resolve(false));
+			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(audioPath: string, dir: string, token: string): Promise<any[]> {
+async function runDiarization(binary: string, audioPath: string, dir: string, token: string): Promise<any[]> {
 	const script = `
 import sys
 import json
@@ -37,7 +42,7 @@ print(json.dumps(segments))
 
 	return new Promise((resolve, reject) => {
 		let output = '';
-		const proc = spawn('python', ['-c', script, audioPath]);
+		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) => {
@@ -112,10 +117,10 @@ parentPort?.on('message', async ({ file, speaker, model, modelDir, token }) => {
 		const [f, buffer] = prepareAudioBuffer(file);
 
 		// Fetch transcript and speakers
-		const hasDiarization = speaker && await canDiarization();
+		const hasDiarization = await canDiarization();
 		const [transcript, speakers] = await Promise.all([
 			whisperPipeline(buffer, {return_timestamps: speaker ? 'word' : false}),
-			(!speaker || !token || !hasDiarization) ? Promise.resolve(): runDiarization(f, modelDir, token),
+			(!speaker || !token || !hasDiarization) ? Promise.resolve(): runDiarization(hasDiarization, f, modelDir, token),
 		]);
 		if(file != f) rmSync(f, { recursive: true, force: true });
 

src/audio.ts

@@ -42,8 +42,8 @@ export class Audio {
 				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"}', {
+				const names = await this.ai.language.json(chunks.join('\n'), '{1: "Detected Name", 2: "Second Name"}', {
-					system: 'Use this following transcript to identify speakers. Only identify speakers you are sure about',
+					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]) => {
@@ -56,5 +56,5 @@ export class Audio {
 		return Object.assign(p, { abort });
 	}
 
-	canDiarization = canDiarization;
+	canDiarization = () => canDiarization().then(resp => !!resp);
 }

src/llm.ts

@@ -255,11 +255,11 @@ 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 worker = new Worker(join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'));
@@ -279,13 +279,13 @@ class LLM {
 				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;
 	}
 
 	/**

vite.config.ts

@@ -1,6 +1,5 @@
 import {defineConfig} from 'vite';
 import dts from 'vite-plugin-dts';
-import {resolve} from 'path';
 
 export default defineConfig({
 	build: {