fluester – [ˈflʏstɐ]

Node.js bindings for OpenAI's Whisper. Hard-fork of whisper-node.

Features

• Output transcripts to JSON (also .txt .srt .vtt)
• Optimized for CPU (Including Apple Silicon ARM)
• Timestamp precision to single word

Installation

Requirements

• make and everything else listed as required to compile whisper.cpp
• Node.js >= 20

1. Add dependency to project

npm install @pr0gramm/fluester

2. Download whisper model of choice

npx --package @pr0gramm/fluester download-model

3. Compile whisper.cpp if you don't want to provide you own version:

npx --package @pr0gramm/fluester compile-whisper

Usage

Important: The API only supports WAV files (just like the original whisper.cpp). You need to convert any files to a supported format before. You can do this using ffmpeg (example taken from the whisper project):

ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav

Translation

import { createWhisperClient } from "@pr0gramm/fluester";

const client = createWhisperClient({
  modelName: "base",
});

const transcript = await client.translate("example/sample.wav");

console.log(transcript); // output: [ {start,end,speech} ]

Output (JSON)

[
  {
    "start": "00:00:14.310", // timestamp start
    "end": "00:00:16.480", // timestamp end
    "speech": "howdy" // transcription
  }
]

Language Detection

import { createWhisperClient } from "@pr0gramm/fluester";

const client = createWhisperClient({
  modelName: "base",
});

const result = await client.detectLanguage("example/sample.wav");
if(!result) {
  console.log(`Detected: ${result.language} with probability ${result.probability}`);
} else {
  console.log("Did not detect anything :(");
}

File Conversion

You cannot use any file for processing, as whisper needs a special encoded wav file. This package ships a helper function that invokes the required ffmpeg commands, so you can convert your files easily:

import { convertFileToProcessableFile } from "@pr0gramm/fluester";

const inputFile = "...";
const outputFile = "output.wav";
await convertFileToProcessableFile(inputFile, outputFile);
// ...
 await client.detectLanguage(outputFile);

Tricks

This library is designed to work well in dockerized environments.

We took time and made some steps independent from each other, so they can be used in a multi-stage docker build.

FROM node:latest as dependencies
    WORKDIR /app
    COPY package.json package-lock.json ./
    RUN npm ci

    RUN npx --package @pr0gramm/fluester compile-whisper
    RUN npx --package @pr0gramm/fluester download-model tiny

FROM node:latest
    WORKDIR /app
    COPY --from=dependencies /app/node_modules /app/node_modules
    COPY ./ ./

This includes the model in the image. If you want to keep your image small, you can also download the model in your entrypoint using the commands above.

Made with

• A lot of love by @ariym at whisper-node
• Whisper OpenAI (using C++ port by: ggerganov)

Roadmap

• Nothing ¯\_(ツ)_/¯