whisper-rs-node v0.1.1

whisper-rs-node

Node.js add-on to whisper-rs (which is bindings to whisper.cpp).

Features:

• Pre-compiled so that you don't have to prepare compilation environments for these platforms.
• Near native performance as the core part is whisper-rs.

Pre-compiled platforms:

Installation

npm i whisper-rs-node

Download models

Download according to whisper.cpp/models/README.md.

Quick start

import {
  convertFileBufferToAudioSamples,
  convertIntegerToFloatAudio,
  convertStereoToMonoAudio,
  WhisperContext,
  newGreedyFullParams,
} from 'whisper-rs-node';

// Read model file. You can download model files according to the README.md
const model = fs.readFileSync(
  path.resolve(__dirname, './PATH_TO_YOUR_MODEL.bin')
);

// Create WhipserContext and WhisperState that are used to run the model.
const ctx = new WhisperContext(model);
const state = ctx.createState();

// Read and parse the .wav file.
// You can read the "parseWav" in [example](./examples/audio_transcription.ts).
const { format, samples } = await parseWav(
  path.resolve(__dirname, './PATH_TO_YOUR_WAV_FILE.wav')
);

// Convert the samples into float array which is required by whisper-rs.
let audio = convertIntegerToFloatAudio(samples);

if (format.channels === 2) {
  // Convert into mono audio which is also required by whisper-rs.
  audio = convertStereoToMonoAudio(audio);
} else if (format.channels !== 1) {
  throw new Error('>2 channels unsupported');
}

if (format.sampleRate !== 16000) {
  throw new Error('sample rate must be 16KHz');
}

// Set config in FullParams which would be used in our WhisperState created before.
const fullParams = newGreedyFullParams(0);
fullParams.setNThreads(1);
fullParams.setTranslate(true);
fullParams.setLanguage('en');
fullParams.setPrintProgress(false);
fullParams.setPrintRealtime(false);
fullParams.setPrintSpecial(false);
fullParams.setPrintTimestamps(false);

// Run the model.
state.full(fullParams, audio);

// Get the result and print them.
for (let i = 0; i < state.fullNSegment(); i += 1) {
  console.log(
    `[${state.fullGetSegmentT0(i)}-${state.fullGetSegmentT1(
      i
    )}] ${state.fullGetSegmentText(i)}`
  );
}

See examples/audio_transcription.ts for a detail example.

LICENSE

MIT