app.py

import numpy as np

@app.get("/speak")
def speak(text: str = Query(..., description="Text to convert to speech")):
    # Prepare input
    inputs = processor(text=text, return_tensors="pt")

    # Generate speech
    speech = model.generate_speech(inputs["input_ids"], speaker_embeddings, vocoder=vocoder)
    audio = speech.numpy().astype(np.float32)

    # --- Normalize ---
    peak = np.max(np.abs(audio))
    if peak > 0:
        audio = (audio / peak) * 0.1  # match your NORMALIZATION_LEVEL

    # --- Smooth (moving average) ---
    window_size = 3  # like SMOOTHING_WINDOW
    if window_size > 1:
        cumsum = np.cumsum(np.insert(audio, 0, 0)) 
        audio = (cumsum[window_size:] - cumsum[:-window_size]) / window_size
        # pad to original length
        audio = np.pad(audio, (window_size//2, window_size-1-window_size//2), mode='edge')

    # Write WAV as 32-bit float
    buf = io.BytesIO()
    sf.write(buf, audio, samplerate=16000, format="WAV", subtype="FLOAT")
    buf.seek(0)

    return StreamingResponse(buf, media_type="audio/wav")