faster-whisper

faster-whisper is an open source Python library for developers adding speech transcription to their software. It uses CTranslate2 to run OpenAI's Whisper models, with faster transcription and lower memory use than the original implementation in the project's benchmark comparisons.

• Batch transcription: Process audio in batches to increase throughput, with higher memory use in the benchmark comparisons.
• Reduced memory use: Optional eight-bit quantization reduces memory needs on both CPU and GPU.
• Silence filtering: The integrated Silero VAD filter skips parts of audio without speech; its conservative default removes silences longer than two seconds.
• Model compatibility: Use Whisper models such as large-v3, compatible Distil-Whisper checkpoints including distil-large-v3, or converted fine-tuned Whisper models that work with Transformers.

The library uses the MIT license and runs on a CPU or an NVIDIA GPU. GPU execution requires NVIDIA CUDA libraries. It's a library for integration into Python software, rather than a standalone transcription app.

Transcription runs on your hardware with compatible converted models loaded from a local directory. Loading a model by name downloads it from Hugging Face. Offline use requires the model and its associated files to be stored locally, including tokenizer files that may otherwise download when the model loads.