tools/audio_transcriber.py

import torch

from langchain_core.tools import tool
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline

from .data_helpers import get_file_path


@tool(parse_docstring=True)
def transcribe_audio_file(file_name: str) -> str:
    """
    Transcribes an audio file to text.

    Args:
        file_name:  The name of the audio file. This is simply the file name,
                    not the full path.

    Returns:
        The transcribed text.
    """
    # Specific setting for local run with GPU busy for the LLM (ollama)
    cuda_available = False
    device = "cuda:0" if cuda_available else "cpu"
    torch_dtype = torch.float16 if cuda_available else torch.float32

    model_id = "openai/whisper-large-v3-turbo"

    model = AutoModelForSpeechSeq2Seq.from_pretrained(
        model_id,
        torch_dtype=torch_dtype,
        low_cpu_mem_usage=True,
        use_safetensors=True
    )
    model.to(device)

    processor = AutoProcessor.from_pretrained(model_id)

    pipe = pipeline(
        "automatic-speech-recognition",
        model=model,
        tokenizer=processor.tokenizer,
        feature_extractor=processor.feature_extractor,
        torch_dtype=torch_dtype,
        device=device,
    )

    generate_kwargs = {
        "return_timestamps": True,
    }

    file_path = get_file_path(file_name)
    result = pipe(file_path, generate_kwargs=generate_kwargs)

    return result["text"]