faster-whisper

Benchmarks show up to 4x faster transcription than OpenAI's Whisper with similar accuracy, thanks to the optimized CTranslate2 engine, reducing time from minutes to seconds for batch processing.

Supports 8-bit integer quantization for both CPU and GPU, cutting VRAM usage by nearly half in tests (e.g., from 4525MB to 2926MB for large-v2 on GPU) without significant accuracy loss.

The BatchedInferencePipeline enables parallel transcription of multiple audio segments, increasing throughput dramatically—benchmarks show a 17s transcription time with batch_size=8 versus 1m03s without.

Includes built-in Silero VAD filtering to remove non-speech segments and word-level timestamps for detailed analysis, enhancing quality out-of-the-box without extra dependencies.