app.py

import os
import gradio as gr
import torch
import numpy as np
from transformers import pipeline
from diffusers import DiffusionPipeline
from pyannote.audio import Pipeline as PyannotePipeline
from dia.model import Dia
from dac.utils import load_model as load_dac_model
from accelerate import init_empty_weights, load_checkpoint_and_dispatch

#-- Configuration
HF_TOKEN   = os.environ["HF_TOKEN"]                 # Gated model access[2]
device_map = "auto"                                 # Distribute models on 4×L4 GPUs[3]

#-- 1. Descript Audio Codec (RVQ)
rvq = load_dac_model(tag="latest", model_type="44khz")  # RVQ encoder/decoder[4]
rvq.eval()
if torch.cuda.is_available(): rvq = rvq.to("cuda")

#-- 2. Voice Activity Detection via Pyannote
vad_pipe = PyannotePipeline.from_pretrained(
    "pyannote/voice-activity-detection",
    use_auth_token=HF_TOKEN
)                                                   # Proper gated VAD load[2]

#-- 3. Ultravox ASR+LLM Pipeline
ultravox_pipe = pipeline(
    model="fixie-ai/ultravox-v0_4",
    trust_remote_code=True,
    device_map=device_map,
    torch_dtype=torch.float16
)                                                   # Custom speech pipeline[2]

#-- 4. Audio Diffusion Model (Prosody)
diff_pipe = DiffusionPipeline.from_pretrained(
    "teticio/audio-diffusion-instrumental-hiphop-256",
    torch_dtype=torch.float16
).to("cuda")                                        # Diffusers-based load[2]

#-- 5. Dia TTS Model Sharded Across GPUs
dia = Dia.from_pretrained(
    "nari-labs/Dia-1.6B",
    device_map=device_map,
    torch_dtype=torch.float16,
    trust_remote_code=True
)                                                   # Auto-sharding in Transformers[2]

#-- Inference Function
def process_audio(audio):
    sr, arr = audio
    arr = arr.numpy() if torch.is_tensor(arr) else arr

    # VAD segmentation
    _ = vad_pipe({"waveform": torch.tensor(arr).unsqueeze(0), "sample_rate": sr})

    # RVQ encode/decode
    x = torch.tensor(arr).unsqueeze(0).to("cuda")
    codes = rvq.encode(x)
    decoded = rvq.decode(codes).squeeze().cpu().numpy()

    # Ultravox ASR → text
    ultra_out = ultravox_pipe({"array": decoded, "sampling_rate": sr})
    text = ultra_out.get("text", "")

    # Diffusion-based prosody enhancement
    pros = diff_pipe(raw_audio=decoded)["audios"][0]

    # Dia TTS synthesis
    tts = dia.generate(f"[emotion:neutral] {text}")
    tts_np = tts.squeeze().cpu().numpy()
    tts_np = tts_np / np.max(np.abs(tts_np)) * 0.95 if tts_np.size else tts_np

    return (sr, tts_np), text

#-- Gradio UI
with gr.Blocks(title="Maya AI 📈") as demo:
    gr.Markdown("## Maya-AI: Supernatural Conversational Agent")
    audio_in  = gr.Audio(source="microphone", type="numpy", label="Your Voice")
    send_btn  = gr.Button("Send")
    audio_out = gr.Audio(label="AI Response")
    text_out  = gr.Textbox(label="Generated Text")
    send_btn.click(process_audio, inputs=audio_in, outputs=[audio_out, text_out])

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