Install faster-whisper (CTranslate2 Whisper), pick compute types, and fix cuDNN load errors, int8 quality questions and VAD filter behavior.
pip install faster-whisper. GPU needs cuBLAS + cuDNN 9 for CUDA 12 — the pip wheels of nvidia-cublas-cu12 and nvidia-cudnn-cu12 are the easy path.
from faster_whisper import WhisperModel; m = WhisperModel("large-v3", compute_type="int8_float16")
vad_filter=True skips silence — big win on meeting audio.
Same Whisper-class accuracy, no install, no model downloads, no GPU questions. Speaker labels, word timestamps and every export (TXT, SRT, VTT, DOCX) included. Free instant preview; credits from $2 and they never expire.
Install the CUDA 12 wheels: pip install nvidia-cublas-cu12 nvidia-cudnn-cu12 and export their lib dirs to LD_LIBRARY_PATH — or use compute_type="int8" on CPU.
On clean speech, no measurable WER change; on noisy/accented audio expect a small hit. int8_float16 on GPU is the best size/quality trade.
The default min_silence_duration is aggressive for fast talkers — pass vad_parameters=dict(min_silence_duration_ms=1000).
Use BatchedInferencePipeline — 3-4× throughput on GPU for many files; it's the engine most hosted services (including ours) build on.
More on faster-whisper: the full faster-whisper page · faster-whisper alternatives · all transcription tools