Faster Whisper

1.2.1 · active · verified Thu Apr 09

Faster Whisper is a re-implementation of OpenAI's Whisper model using CTranslate2, which allows for faster inference and reduced memory usage. It is highly optimized for CPU and GPU, supporting various compute types. The current version is 1.2.1, with an active release cadence, frequently adding new features, model support, and performance improvements.

Warnings

breaking Version 1.0.0 upgraded CTranslate2 to v4.0, which added support for CUDA 12. Users on older CUDA versions (e.g., CUDA 11.x) might face compatibility issues and need to downgrade CTranslate2 or use a compatible `faster-whisper` version.
Fix: Ensure your CUDA toolkit and CTranslate2 version are compatible. If on older CUDA, consider installing a specific CTranslate2 version (e.g., `pip install ctranslate2<4.0`) or using a `faster-whisper` version prior to 1.0.0.
breaking In version 1.1.0, some Voice Activity Detection (VAD) parameters were renamed. However, this change was reverted in version 1.1.1. If you implemented VAD parameter tuning with v1.1.0, your code might break when upgrading to v1.1.1 or later due to the reversion to original names.
Fix: If using v1.1.0, review VAD parameter names when upgrading. For versions 1.1.1 and later, refer to the documentation for the established VAD parameter names, which were restored to their pre-1.1.0 state.
gotcha Older versions (prior to 1.1.1) and certain VAD configurations could lead to high RAM usage and Out-Of-Memory (OOM) errors, particularly with longer audio files or larger batch sizes.
Fix: Upgrade to `faster-whisper` v1.1.1 or newer, which includes fixes for VAD-related OOM errors. Monitor memory usage, especially when enabling VAD or using batched inference, and adjust VAD parameters or batch sizes if necessary.
gotcha When using batched inference, specific issues regarding `clip_timestamps` and the `<|nocaptions|>` token were fixed in version 1.2.1. In earlier versions, these features might not have behaved as expected in batched mode, potentially leading to incorrect timestamp merging or token suppression.
Fix: Upgrade to `faster-whisper` v1.2.1 or newer to ensure correct behavior of `clip_timestamps` and `suppress_tokens` (including `<|nocaptions|>`) during batched inference. Always test batched inference with your specific use case.

Install

pip install faster-whisper Basic installation
pip install faster-whisper[vad,audio] With VAD and audio file support (PyAV)

Imports

WhisperModel
```
from faster_whisper import WhisperModel
```

Quickstart

Demonstrates loading a Whisper model and transcribing an audio file. The model will automatically download from Hugging Face Hub if not already cached. Uses CPU by default for broad compatibility; change `device` and `compute_type` for GPU acceleration.

from faster_whisper import WhisperModel
import os

# Ensure you have an audio file named 'audio.mp3' in the current directory
# For example, download a short audio clip or record one.
# Example: https://www.soundhelix.com/examples/mp3/SoundHelix-Song-1.mp3

model_size = os.environ.get('WHISPER_MODEL_SIZE', 'tiny.en') # e.g., 'large-v3', 'medium', 'tiny.en'

# Run on CPU with INT8 compute type for general compatibility
# For GPU, change device='cuda' and compute_type='float16' if supported
model = WhisperModel(model_size, device='cpu', compute_type='int8')

# Transcribe the audio file
# Replace 'audio.mp3' with the path to your audio file
segments, info = model.transcribe("audio.mp3", beam_size=5)

print(f"Detected language '{info.language}' with probability {info.language_probability:.2f}")

for segment in segments:
    print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")

view raw JSON →