Compare commits
4 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
| 9b831f7d95 | |||
| 498b326e45 | |||
| 56e4efec94 | |||
| a07f069ad0 |
1038
package-lock.json
generated
1038
package-lock.json
generated
File diff suppressed because it is too large
Load Diff
12
package.json
12
package.json
@@ -1,6 +1,6 @@
|
||||
{
|
||||
"name": "@ztimson/ai-utils",
|
||||
"version": "0.7.2",
|
||||
"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/utils": "^0.27.9",
|
||||
"@ztimson/node-utils": "^1.0.7",
|
||||
"@ztimson/utils": "^0.28.13",
|
||||
"cheerio": "^1.2.0",
|
||||
"openai": "^6.6.0",
|
||||
"tesseract.js": "^6.0.1",
|
||||
"openai": "^6.22.0",
|
||||
"tesseract.js": "^7.0.0",
|
||||
"wavefile": "^11.0.0"
|
||||
},
|
||||
"devDependencies": {
|
||||
|
||||
25
src/asr.ts
25
src/asr.ts
@@ -9,15 +9,20 @@ import wavefile from 'wavefile';
|
||||
|
||||
let whisperPipeline: any;
|
||||
|
||||
export async function canDiarization(): Promise<boolean> {
|
||||
return new Promise((resolve) => {
|
||||
const proc = spawn('python', ['-c', 'import pyannote.audio']);
|
||||
proc.on('close', (code: number) => resolve(code === 0));
|
||||
proc.on('error', () => resolve(false));
|
||||
});
|
||||
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(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 });
|
||||
|
||||
|
||||
@@ -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"}', {
|
||||
system: 'Use this following transcript to identify speakers. Only identify speakers you are sure about',
|
||||
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]) => {
|
||||
@@ -56,5 +56,5 @@ export class Audio {
|
||||
return Object.assign(p, { abort });
|
||||
}
|
||||
|
||||
canDiarization = canDiarization;
|
||||
canDiarization = () => canDiarization().then(resp => !!resp);
|
||||
}
|
||||
|
||||
25
src/llm.ts
25
src/llm.ts
@@ -255,12 +255,11 @@ class LLM {
|
||||
/**
|
||||
* Create a vector representation of a string
|
||||
* @param {object | string} target Item that will be embedded (objects get converted)
|
||||
* @param {maxTokens?: number, overlapTokens?: number, parellel?: number} opts Options for embedding such as chunk sizes and parallel processing
|
||||
* @param {maxTokens?: number, overlapTokens?: number} opts Options for embedding such as chunk sizes
|
||||
* @returns {Promise<Awaited<{index: number, embedding: number[], text: string, tokens: number}>[]>} Chunked embeddings
|
||||
*/
|
||||
async embedding(target: object | string, opts: {maxTokens?: number, overlapTokens?: number, parallel?: number} = {}) {
|
||||
let {maxTokens = 500, overlapTokens = 50, parallel = 1} = opts;
|
||||
|
||||
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'));
|
||||
@@ -280,19 +279,13 @@ class LLM {
|
||||
worker.postMessage({text, model: this.ai.options?.embedder || 'bge-small-en-v1.5', modelDir: this.ai.options.path});
|
||||
});
|
||||
};
|
||||
let i = 0, chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
|
||||
const next: Function = () => {
|
||||
const index = i++;
|
||||
if(index >= chunks.length) return;
|
||||
const text = chunks[index];
|
||||
return embed(text).then(embedding => {
|
||||
results.push({index, embedding, text, tokens: this.estimateTokens(text)});
|
||||
return next();
|
||||
})
|
||||
const chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
|
||||
for(let i = 0; i < chunks.length; i++) {
|
||||
const text= chunks[i];
|
||||
const embedding = await embed(text);
|
||||
results.push({index: i, embedding, text, tokens: this.estimateTokens(text)});
|
||||
}
|
||||
|
||||
await Promise.all(Array(parallel).fill(null).map(() => next()));
|
||||
return results.toSorted((a, b) => a.index - b.index);
|
||||
return results;
|
||||
}
|
||||
|
||||
/**
|
||||
|
||||
@@ -1,6 +1,5 @@
|
||||
import {defineConfig} from 'vite';
|
||||
import dts from 'vite-plugin-dts';
|
||||
import {resolve} from 'path';
|
||||
|
||||
export default defineConfig({
|
||||
build: {
|
||||
|
||||
Reference in New Issue
Block a user