DecodingOptions

[pop2pop3]

Since PyTorch CPU does not assist fp16, so I change DecodingOptions fp16 parameter to False, this will slow down the decoding process since CPU is reasoned through the use of fp16 combined precision calculation. At least this happened on arm64 CPU.

[Strangemother]

Thank you - very helpful.

As an example I applied it directly to the transcribe method because I'm lazy:

import whisper
model = whisper.load_model('base')
result = model.transcribe("audio.mp3", fp16=False)
print(result["text"])

[pop2pop3]

Sure, or you can change from dtype = torch.float16 to dtype = torch.float32 directly in transcribe() function, so you don't need to add another argument