# %% Utilisation de Whisper pour la transcription de podcasts en français
import numpy as np
import torch
import torchaudio
import tqdm
from transformers import (
    AutoModelForCausalLM,
    AutoModelForSpeechSeq2Seq,
    AutoProcessor,
    pipeline,
)

# %% File paths
audio_paths = ["METTRE LES LIENS DES FICHIERS MP3 OU WAV ICI"]

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

# %% Load model
model_name_or_path = "bofenghuang/whisper-large-v3-french"
processor = AutoProcessor.from_pretrained(model_name_or_path)
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_name_or_path,
    torch_dtype=torch_dtype,
    low_cpu_mem_usage=True,
)
model.to(device)

# %% Load draft model
assistant_model_name_or_path = "bofenghuang/whisper-large-v3-french-distil-dec2"
assistant_model = AutoModelForCausalLM.from_pretrained(
    assistant_model_name_or_path,
    torch_dtype=torch_dtype,
    low_cpu_mem_usage=True,
)
assistant_model.to(device)

# %% Init pipeline
pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    feature_extractor=processor.feature_extractor,
    tokenizer=processor.tokenizer,
    torch_dtype=torch_dtype,
    device=device,
    generate_kwargs={"assistant_model": assistant_model},
    max_new_tokens=128,
)


# %% Transcript function
def transcript(audio_path):
    # Load audio
    model_sr = 16000
    speech, sr = torchaudio.load(audio_path)
    speech_16000 = torchaudio.functional.resample(speech, orig_freq=sr, new_freq=model_sr)
    speech_16000 = speech_16000.squeeze()

    # Run pipeline
    result = pipe(np.array(speech_16000))

    # Save text result to file
    transcript_path = f'whisper-large/{audio_path.replace(".mp3", "_transcript_whisper.txt")}'
    with open(transcript_path, "w") as f:
        f.write(result["text"])
    return None


# %% Transcription loop
for audio_path in tqdm.tqdm(audio_paths):
    transcript(audio_path)