SYSTRAN Packages faster-whisper-small.en for 4x Faster English Speech Recognition
SYSTRAN has published a CTranslate2 conversion of OpenAI's English-only Whisper small model to Hugging Face, ready for use with the faster-whisper library.
Takeaways
- SYSTRAN published faster-whisper-small.en, a CTranslate2 conversion of OpenAI's English-only Whisper small model.
- MIT-licensed, FP16 weights, already past 870k downloads on the Hugging Face Hub.
- Loads directly via
WhisperModel("small.en")in the faster-whisper library, no manual conversion needed. - faster-whisper runs up to 4x faster than openai/whisper at equal accuracy, with lower memory use.
compute_typelets you re-quantize to int8 on load for further CPU or GPU speedups.- Requires CUDA 12 and cuDNN 9 for GPU execution on current ctranslate2 releases.
SYSTRAN has published a ready-to-run checkpoint that converts OpenAI’s English-only Whisper small model into the CTranslate2 format. The package lets faster-whisper download and load the model directly, removing the need to install the Transformers conversion tools or prepare the weights locally.
The Hugging Face repository lists an MIT license and had recorded 172,845 downloads in the last month. It joins SYSTRAN’s other converted checkpoints in the model collection.
Same weights, CTranslate2 layout
The repository preserves the architecture and trained parameters of openai/whisper-small.en. CTranslate2 changes how those parameters are stored and executed, using an inference-oriented layout and optimized kernels. The upload includes no additional training, capability, or evaluation results.
| Property | Value |
|---|---|
| Source checkpoint | openai/whisper-small.en |
| Model size | About 244 million parameters |
| Language | English |
| Stored precision | FP16 |
| Runtime | CTranslate2 and faster-whisper |
| Task | Automatic speech recognition |
The model card records the conversion as a single command that copies the tokenizer and preprocessing configuration while storing the converted weights in float16:
ct2-transformers-converter \
--model openai/whisper-small.en \
--output_dir faster-whisper-small.en \
--copy_files tokenizer.json preprocessor_config.json \
--quantization float16