import torch
import accelerate
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "bavihao/Ahao-Whisper-medium-Vi-v2"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, device_map = 'auto'
)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    batch_size=16,
    return_timestamps=True,
    torch_dtype=torch_dtype,
)
result = pipe(audio_path, generate_kwargs={"language": "vietnamese"})
print(result["text"])