mutranscriber

Native Rust audio transcription using Qwen3-ASR, powered by Candle.

Features

• Pure Rust implementation with no Python dependencies
• Automatic model download from HuggingFace Hub
• GPU acceleration via CUDA or Metal
• Audio extraction from video files via GStreamer
• Both CLI tool and library API

Installation

From source

# Standard build (includes GStreamer support)
cargo install --path .

# With GPU support
cargo install --path . --features cuda    # NVIDIA
cargo install --path . --features metal   # macOS

# Without GStreamer (library-only, no file loading)
cargo install --path . --no-default-features

With Nix

# CPU-only build
nix build github:cfcosta/mutranscriber#mutranscriber-cpu

# CUDA-enabled build (NixOS with NVIDIA drivers)
nix build github:cfcosta/mutranscriber#mutranscriber-cuda

# Run directly without installing
nix run github:cfcosta/mutranscriber#mutranscriber-cpu -- audio.wav

The Nix packages include all dependencies (GStreamer plugins, CUDA libraries) and work out of the box.

Requirements

• Rust 1.70+
• GStreamer development libraries (enabled by default)
• For cuda feature: CUDA toolkit
• For metal feature: macOS with Metal support

Usage

CLI

# Transcribe an audio file
mutranscriber recording.wav

# Transcribe a video file
mutranscriber video.mp4

# Use the larger model
mutranscriber audio.wav --model large

# Force CPU mode
mutranscriber audio.wav --cpu

# Print to stdout only
mutranscriber audio.wav --stdout-only

# Custom output path
mutranscriber audio.wav --output transcript.txt

Library

use mutranscriber::{Transcriber, TranscriberConfig, ModelVariant};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Create transcriber with default settings
    let transcriber = Transcriber::from_env();

    // Or with custom configuration
    let config = TranscriberConfig {
        variant: ModelVariant::Small,
        use_gpu: true,
        sample_rate: 16000,
        output_dir: None,
    };
    let transcriber = Transcriber::with_config(config);

    // Preload the model
    transcriber.preload().await?;

    // Transcribe raw audio samples (16kHz, f32)
    let audio_samples: Vec<f32> = load_audio_somehow();
    let text = transcriber.transcribe_audio(&audio_samples).await?;

    println!("{}", text);
    Ok(())
}

Model Variants

Models are automatically downloaded from HuggingFace Hub on first use and cached locally.

Audio Requirements

• Sample rate: 16kHz
• Format: f32 mono samples
• The library handles padding to 30 seconds internally (matching WhisperFeatureExtractor)

When using the gstreamer feature, audio is automatically extracted and resampled from any format GStreamer supports.

Build Features

Development

# Enter dev environment (requires Nix)
nix develop

# Run tests
cargo test --lib

# Run integration tests (downloads model)
cargo test --test integration_test -- --ignored

# Format and lint
cargo fmt
cargo clippy

Architecture

Audio Input (16kHz f32)
    │
    ▼
Mel Spectrogram (128 bins, 30s padded)
    │
    ▼
Audio Encoder (Qwen3-AuT Transformer)
    │
    ▼
Audio Features (projected to LLM dim)
    │
    ▼
Qwen3 LLM Decoder (with audio embeddings)
    │
    ▼
Text Output

License

MIT OR Apache-2.0