app.py

import gradio as gr
from huggingface_hub import InferenceClient
import edge_tts
import tempfile
import asyncio

# Client Hugging Face
client = InferenceClient("google/gemma-1.1-2b-it")

# Fonction de synthèse vocale (TTS)
async def text_to_speech(text, voice="fr-FR-DeniseNeural", rate=0, pitch=0):
    if not text.strip():
        return None, "Veuillez entrer du texte."

    rate_str = f"{rate:+d}%"
    pitch_str = f"{pitch:+d}Hz"
    communicate = edge_tts.Communicate(text, voice, rate=rate_str, pitch=pitch_str)

    # Sauvegarde en fichier temporaire
    with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as tmp_file:
        tmp_path = tmp_file.name
        await communicate.save(tmp_path)
    
    return tmp_path

# Modèle rapide (Fast)
def models(query):
    messages = [{"role": "user", "content": f"[SYSTEM] You are a fast assistant. [USER] {query}"}]
    response = ""

    for message in client.chat_completion(messages, max_tokens=2048, stream=True):
        token = message.choices[0].delta.content
        response += token

    # Convertir en audio
    tts_path = asyncio.run(text_to_speech(response))
    return response, tts_path

# Modèle critique (Critical Thinker)
def nemo(query):
    budget = 3
    message = f"""[INST] [SYSTEM] You are a deep-thinking assistant.
<count> {budget} </count> <step> Analyzing question... </step> <count> {budget-1} </count>
<answer> Here is your answer: {query} </answer> [/INST]"""

    stream = client.text_generation(message, max_new_tokens=4096, stream=True)
    output = "".join([response.token.text for response in stream])

    # Convertir en audio
    tts_path = asyncio.run(text_to_speech(output))
    return output, tts_path

# Interface Gradio
description = "# Light ChatBox\n### Enter a question and get a response with voice!"

with gr.Blocks() as demo1:
    gr.Interface(fn=models, inputs="text", outputs=["text", "audio"], description=description)

with gr.Blocks() as demo2:
    gr.Interface(fn=nemo, inputs="text", outputs=["text", "audio"], description="Critical Thinker")

with gr.Blocks() as demo:
    gr.TabbedInterface([demo1, demo2], ["Fast", "Critical"])

demo.queue()
demo.launch()