faster-whisper

faster-whisper is a reimplementation of OpenAI's Whisper speech-to-text model utilizing the CTranslate2 inference engine. It is designed to be more efficient than the original implementation, providing up to 4x faster transcription while consuming less memory. Efficiency can be further enhanced through 8-bit quantization on both CPU and GPU. Key capabilities include: - Support for various model sizes, including compatibility with Distil-Whisper checkpoints. - Batched transcription for increased throughput. - Word-level timestamps for precise audio alignment. - Integrated Silero VAD (Voice Activity Detection) filter to remove non-speech segments. - Support for multiple compute types (float16, int8_float16, int8) across CUDA and CPU devices. - A conversion tool to transform Transformers-compatible Whisper models into CTranslate2 format. Unlike the original Whisper, it uses PyAV for audio decoding, removing the requirement for a system-wide FFmpeg installation.