Update app.py

app.py

@@ -3,31 +3,29 @@ import gradio as gr
 import torch
 import numpy as np
 from transformers import pipeline
+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
 
-# Retrieve your HF token from the Space secrets
+# Environment token
 HF_TOKEN = os.environ["HF_TOKEN"]
 
-# Automatically shard across 4× L4 GPUs
+# Shard large models across 4× L4 GPUs
 device_map = "auto"
 
-# 1. Load RVQ codec
+# 1. RVQ codec (Descript Audio Codec)
 rvq = load_dac_model(tag="latest", model_type="44khz")
 rvq.eval()
-if torch.cuda.is_available():
-    rvq = rvq.to("cuda")
-
-# 2. Load VAD via Hugging Face pipeline (no segmentation mismatch)
-vad_pipe = pipeline(
-    "voice-activity-detection",
-    model="pyannote/voice-activity-detection",
-    use_auth_token=HF_TOKEN,
-    device=0 if torch.cuda.is_available() else -1
+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
 )
 
-# 3. Load Ultravox (speech-in → text+LLM)
+# 3. Ultravox pipeline (speech → text + LLM)
 ultravox_pipe = pipeline(
     model="fixie-ai/ultravox-v0_4",
     trust_remote_code=True,
@@ -35,7 +33,7 @@ ultravox_pipe = pipeline(
     torch_dtype=torch.float16
 )
 
-# 4. Load diffusion prosody model
+# 4. Diffusion-based prosody model
 diff_pipe = pipeline(
     "audio-to-audio",
     model="teticio/audio-diffusion-instrumental-hiphop-256",
@@ -44,7 +42,7 @@ diff_pipe = pipeline(
     torch_dtype=torch.float16
 )
 
-# 5. Load Dia TTS with multi-GPU dispatch
+# 5. Dia TTS loaded with multi-GPU dispatch
 with init_empty_weights():
     dia = Dia.from_pretrained(
         "nari-labs/Dia-1.6B",
@@ -58,40 +56,42 @@ dia = load_checkpoint_and_dispatch(
     dtype=torch.float16
 )
 
-# 6. Inference function
+# Inference function
 def process_audio(audio):
     sr, array = audio
-    array = array.numpy() if torch.is_tensor(array) else array
+    # Ensure numpy
+    if torch.is_tensor(array): array = array.numpy()
 
-    # Voice activity detection
-    speech = vad_pipe(array, sampling_rate=sr)[0]["chunks"]
+    # VAD: extract speech regions
+    chunks = vad_pipe(array, sampling_rate=sr)
 
     # RVQ encode/decode
     x = torch.tensor(array).unsqueeze(0).to("cuda")
     codes = rvq.encode(x)
     decoded = rvq.decode(codes).squeeze().cpu().numpy()
 
-    # Ultravox ASR → LLM
+    # Ultravox ASR + LLM
     out = ultravox_pipe({"array": decoded, "sampling_rate": sr})
     text = out.get("text", "")
 
-    # Diffusion-based prosody
-    pros = diff_pipe({"array": decoded, "sampling_rate": sr})["array"][0]
+    # Diffusion prosody enhancement
+    pros_audio = diff_pipe({"array": decoded, "sampling_rate": sr})["array"][0]
 
-    # Dia TTS synth
-    tts = dia.generate(f"[emotion:neutral] {text}").squeeze().cpu().numpy()
-    tts = tts / np.max(np.abs(tts)) * 0.95
+    # 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
 
-    return (sr, tts), text
+    return (sr, tts_np), text
 
-# 7. Gradio UI
-with gr.Blocks() as demo:
+# Gradio UI
+with gr.Blocks(title="Maya AI 📈", theme=None) as demo:
     gr.Markdown("## Maya-AI: Supernatural Conversational Agent")
     audio_in = gr.Audio(source="microphone", type="numpy", label="Your Voice")
-    send = gr.Button("Send")
+    send_btn = gr.Button("Send")
     audio_out = gr.Audio(label="AI’s Response")
     text_out = gr.Textbox(label="Generated Text")
-    send.click(process_audio, inputs=audio_in, outputs=[audio_out, text_out])
+    send_btn.click(process_audio, inputs=audio_in, outputs=[audio_out, text_out])
 
 if __name__ == "__main__":
     demo.launch()