app.py

from transformers import pipeline
import gradio as gr
import torch

device = 'cuda' if torch.cuda.is_available else 'cpu'

model = pipeline(
    "automatic-speech-recognition", 
    device=device, 
    model='openai/whisper-large-v3',
    chunk_length_s=30,
    generate_kwargs={"task": "transcribe"}
)

def transcribe_audio(mic=None, file=None, return_timestamps=False):
    if mic is not None:
        audio = mic
    elif file is not None:
        audio = file
    else:
        return "You must either provide a mic recording or a file"
        
    gr.Info(f"Device: {device}") 
    with torch.no_grad():
        result = model(audio, return_timestamps=return_timestamps, batch_size=8)
        if return_timestamps:
            return result['chunks']
        else:
            return result['text']
        

gr.Interface(
    fn=transcribe_audio,
    inputs=[
        gr.Audio(sources=["microphone"], type="filepath"),
        gr.Audio(sources=["upload"], type="filepath"),
        gr.Checkbox(label="Add timestamps?")
    ],
    outputs="text",
).launch()