app.py

import gradio as gr
from transformers import pipeline
import torch

# Gerät auswählen (GPU, falls verfügbar, sonst CPU)
device = "cuda" if torch.cuda.is_available() else "cpu"

# Lade das kleinere Modell
pipe = pipeline("text-generation", model="google/flan-t5-xxl", device=device)

def chat(message, history):
  prompt = f"User: {message}\nAssistant:"
  sequences = pipe(prompt, max_length=256)
  response = sequences[0]['generated_text'].split("Assistant:")[1].strip()
  history.append((message, response))
  return history, history

def transcribe_and_send(audio, history):
  if audio is None:
    return history, "Keine Audioaufnahme erhalten."

  try:
      asr = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
      transcription = asr(audio)["text"]
      return chat(transcription, history)
  except Exception as e:
      print(f"Fehler bei der Transkription: {e}")
      return history, "Fehler bei der Audioverarbeitung."

with gr.Blocks() as demo:
    chatbot = gr.Chatbot()
    state = gr.State([])

    with gr.Row():
        audio_input = gr.Audio(type="filepath", label="Sprachaufnahme")
        text_input = gr.Textbox(placeholder="Nachricht eingeben...")

    send_button = gr.Button("Senden (Text)")
    clear_button = gr.Button("Chat löschen")

    send_button.click(chat, [text_input, state], [state, chatbot])
    audio_input.submit(transcribe_and_send, [audio_input, state], [state, chatbot])
    clear_button.click(lambda: [], outputs=[chatbot])
    text_input.submit(chat, [text_input, state], [state, chatbot])

if __name__ == "__main__":
    demo.launch()