Whisper Experiments

A Rust implementation for experimenting with the Whisper speech recognition model using the Candle framework.

Dependencies

• Mac with GPU and Metal
• Huggingface CLI
• MP3 to WAV converter, e.g. FFMpeg

Installation

1. Clone the repository:

git clone https://github.com/corybuecker/whisper-experiments.git
cd whisper-experiments

2. Install the Huggingface CLI if you haven't already.

brew install huggingface-cli
huggingface-cli login

3. Run the model download script:

./download_models.sh

4. Create the Mel spectrogram filter bank with Python:

pip install librosa

import numpy as np
from librosa import filters

filters.mel(sr=16000, n_fft=400, n_mels=128, dtype=np.float32).flatten().tofile("melfilter.bytes")

Usage

1. Convert your MP3 file to the required format:
   • Mono, single channel
   • 16-bit depth
   • 16kHz sample rate
   • WAV format

ffmpeg -i input.mp3 -ac 1 -acodec pcm_s16le -ar 16000 -map_metadata -1 output.wav

2. Run the project:

cargo run --release --features metal -- --file output.wav

Once the transcription process completes, it will write a JSON file with the full text and each parsed segment with its timestamps.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgements

• Candle framework and examples
• Huggingface