openai-whisper: 'FP16 is not supported on CPU; using FP32 instead' is only a warning
@devsingh05▲ 10 net votes
On a machine without CUDA, Whisper falls back to 32-bit floats and prints this; transcription still works, just slower. Silence it with model.transcribe(audio, fp16=False). For faster CPU transcription use a smaller model (base, small) or faster-whisper, which uses CTranslate2 with int8 and is several times faster on CPU.