Quiet Crab - Run Whisper Models in Rust Natively

Quiet Crab is a project to let developers run Whisper transcription models against various audio files in pure Rust.

I was working on a flashcard app but couldn't get Whisper.cpp Rust bindings to build on NixOS so it was easier to rebuilt it.

Installation

At the moment this project is not publishing a crate.

To install, add the following to your Cargo.toml

[dependencies]
quiet-crab = { version = "0.1.0"}
# Use wgpu for GPU support, and ndarray for CPU-only
burn = { version = "0.20.1", features = ["wgpu", "ndarray"] }

Quick Start

Grab A Model And Tokenizer

Get yourself a tokenizer.json from https://huggingface.co/openai/whisper-large-v3/blob/main/tokenizer.json

Also grab a medium/small model from huggingface (large is currently unsupported):

Run A Transcription

use quiet_crab::{
    inference::Transcriber,
    model::{config::WhisperConfig, whisper::WhisperModel},
    tokenizer::whisper_tokenizer::WhisperTokenizer,
};

// For CPU users
// use burn::backend::{ndarray::NdArrayDevice, NdArray};
// type Backend = NdArray<f32>;

// For GPU users
use burn::backend::Wgpu;
type Backend = Wgpu;

fn main() {
    let device = Default::default();
    let config = WhisperConfig::medium();

    println!("Loading model weights...");
    let model =
        WhisperModel::<Backend>::from_safetensors(&config, "whisper-medium.safetensors", &device)
            .unwrap();

    println!("Loading tokenizer...");
    let tokenizer = WhisperTokenizer::from_file("tokenizer.json").unwrap();

    let transcriber = Transcriber::new(model, tokenizer, config, device);

    println!("Transcribing...");
    let text = transcriber.transcribe("sample.mp3", Some("en")).unwrap();

    println!("{text}");
}

Model Support