app.py

import gradio as gr
from transformers import pipeline, WhisperProcessor, WhisperForConditionalGeneration

processor = WhisperProcessor.from_pretrained("facebook/whisper-large")
model = WhisperForConditionalGeneration.from_pretrained("facebook/whisper-large")

def transcrire_audio(audio, prompt):
    input_features = processor(audio, return_tensors="pt").input_features

    output_without_prompt = model.generate(input_features)
    transcription_sans_prompt = processor.batch_decode(output_without_prompt, skip_special_tokens=True)[0]

    prompt_ids = processor.get_prompt_ids(prompt)
    output_with_prompt = model.generate(input_features, prompt_ids=prompt_ids)
    transcription_avec_prompt = processor.batch_decode(output_with_prompt, skip_special_tokens=True)[0]

    return {
        "Transcription sans prompt": transcription_sans_prompt,
        "Transcription avec prompt": transcription_avec_prompt
    }

iface = gr.Interface(
    fn=transcrire_audio,
    inputs=["audio", "text"],
    outputs=["text", "text"],
    live=True
)

iface.launch()