my-wispr-npm

Offline speech-to-text for Node.js. Pipe any PCM audio stream into whisper.cpp — no cloud, no API key.

Audio format: 16-bit signed LE PCM, 16 kHz, mono.

1. Install

npm install my-wispr-npm
npx my-wispr-setup

Setup clones whisper.cpp, builds it, and downloads the base.en model (needs git, make, and a C++ compiler).

Already have whisper.cpp? Skip setup and set:

export WISPR_BIN=/path/to/whisper-cli
export WISPR_MODEL=/path/to/ggml-base.en.bin

2. Transcribe a stream

import { PassThrough } from 'stream';
import { createWisprStream } from 'my-wispr-npm';

const pcm = new PassThrough();
const wispr = createWisprStream({ stream: pcm, model: 'base.en' });

wispr.on('transcription', (text) => console.log(text));
wispr.start();

// Write raw Int16 PCM buffers into `pcm`
pcm.write(int16Buffer);

3. Or run a WebSocket server

npm install ws

import { createWisprServer } from 'my-wispr-npm/server';

const server = createWisprServer({ port: 8080 });

server.on('transcription', ({ text }) => console.log(text));
await server.listen();
// Clients send raw PCM; server replies: { "type": "transcription", "text": "..." }

Options

createWisprStream(opts)

| Option | Default | Notes | |---|---|---| | stream | required | Readable of raw PCM | | model | 'base.en' | Whisper model name | | windowChunks | 2 | Window size (20ms chunks) | | stepChunks | 1 | Slide step (20ms chunks) | | whisperBin | auto | Override binary path | | modelPath | auto | Override model path |

Events: transcription, error, start, stop
Methods: start(), stop()

createWisprServer(opts)

Same whisper options as above, plus:

| Option | Default | Notes | |---|---|---| | port | 8080 | Listen port | | host | '0.0.0.0' | Listen host | | path | — | Optional WS path | | server | — | Reuse an HTTP server | | autoReply | true | Send JSON to clients |

Methods: listen(), close()

License

MIT