The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's attention_mask to obtain reliable results.

[djj067]

I Met the issue in title when im trying to use whisper large to do the recognition.

Here is my setup:

model_id = "openai/whisper-large-v3"
whisper_model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True)
processor = AutoProcessor.from_pretrained(model_id)
pipe = pipeline(
"automatic-speech-recognition",
model=whisper_model,
tokenizer=processor.tokenizer,
feature_extractor=processor.feature_extractor,
chunk_length_s=30,
batch_size=8,
torch_dtype=torch_dtype,
device=device,
)
result = pipe(audio_data, generate_kwargs={"language": "chinese"}, return_timestamps=True)

and i do see some problems, for example, about 20% for the output files doesnt have the timestamp,

also sometimes, the model recognize background music(no lyrics, just melody) as the speech. and sometimes it has some werid generetion.

May i know if those problems are due to some wrong setting(like the attention mask one) or just sometimes the model cant do it correctly,