Update app.py

app.py

@@ -9,25 +9,28 @@ 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 HF_TOKEN from Secrets
+# Environment token from HF Secrets
 HF_TOKEN = os.environ["HF_TOKEN"]
-
-# Automatically shard across 4× L4 GPUs
 device_map = "auto"
 
-# 1. Load Descript Audio Codec (RVQ)
+print("Loading models...")
+
+# 1. RVQ Codec (Descript Audio Codec)
+print("Loading RVQ Codec...")
 rvq = load_dac_model(tag="latest", model_type="44khz")
 rvq.eval()
 if torch.cuda.is_available():
     rvq = rvq.to("cuda")
 
-# 2. Load Voice Activity Detection via Pyannote
+# 2. Voice Activity Detection
+print("Loading VAD...")
 vad_pipe = PyannotePipeline.from_pretrained(
     "pyannote/voice-activity-detection",
     use_auth_token=HF_TOKEN
 )
 
-# 3. Load Ultravox (speech-to-text + LLM) via Transformers
+# 3. Ultravox ASR+LLM
+print("Loading Ultravox...")
 ultravox_pipe = pipeline(
     model="fixie-ai/ultravox-v0_4",
     trust_remote_code=True,
@@ -35,14 +38,18 @@ ultravox_pipe = pipeline(
     torch_dtype=torch.float16
 )
 
-# 4. Load Audio Diffusion model via Diffusers
+# 4. Audio Diffusion Model
+print("Loading Audio Diffusion...")
 diff_pipe = DiffusionPipeline.from_pretrained(
-    "teticio/audio-diffusion-instrumental-hiphop-256"
+    "teticio/audio-diffusion-instrumental-hiphop-256",
+    torch_dtype=torch.float16
 ).to("cuda")
 
-# 5. Load Dia TTS with meta-weight initialization and multi-GPU dispatch
+# 5. Dia TTS Model
+print("Loading Dia TTS...")
 with init_empty_weights():
     dia = Dia.from_pretrained("nari-labs/Dia-1.6B")
+
 dia = load_checkpoint_and_dispatch(
     dia,
     "nari-labs/Dia-1.6B",
@@ -50,41 +57,92 @@ dia = load_checkpoint_and_dispatch(
     dtype=torch.float16
 )
 
-# Inference function
-def process_audio(audio):
-    sr, array = audio
-    array = array.numpy() if torch.is_tensor(array) else array
-
-    # 2.1 VAD: segment speech regions (not used further here)
-    _ = vad_pipe(array, sampling_rate=sr)
+print("All models loaded successfully!")
 
-    # 1.1 RVQ encode/decode for discrete audio tokens
-    x = torch.tensor(array).unsqueeze(0).to("cuda")
-    codes = rvq.encode(x)
-    decoded = rvq.decode(codes).squeeze().cpu().numpy()
-
-    # 3. Ultravox ASR + LLM to generate response text
-    ultra_out = ultravox_pipe({"array": decoded, "sampling_rate": sr})
-    text = ultra_out.get("text", "")
-
-    # 4. Diffusion-based prosody enhancement
-    pros = diff_pipe(raw_audio=decoded)["audios"][0]
-
-    # 5. Dia TTS synthesis with neutral emotion tag
-    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_np), text
-
-# Gradio UI
+# Audio processing function
+def process_audio(audio):
+    try:
+        if audio is None:
+            return None, "No audio input provided"
+        
+        sr, array = audio
+        
+        # Ensure numpy array
+        if torch.is_tensor(array):
+            array = array.numpy()
+        
+        # Voice Activity Detection
+        vad_result = vad_pipe({"waveform": torch.tensor(array).unsqueeze(0), "sample_rate": sr})
+        
+        # RVQ encode/decode for audio compression
+        audio_tensor = torch.tensor(array).unsqueeze(0)
+        if torch.cuda.is_available():
+            audio_tensor = audio_tensor.to("cuda")
+        codes = rvq.encode(audio_tensor)
+        decoded = rvq.decode(codes).squeeze().cpu().numpy()
+        
+        # Ultravox ASR + LLM processing
+        ultra_out = ultravox_pipe({"array": decoded, "sampling_rate": sr})
+        text = ultra_out.get("text", "I understand your audio input.")
+        
+        # Audio diffusion for prosody enhancement
+        try:
+            prosody_result = diff_pipe(raw_audio=decoded)
+            if "audios" in prosody_result:
+                prosody_audio = prosody_result["audios"][0]
+            else:
+                prosody_audio = decoded
+        except Exception as e:
+            print(f"Diffusion processing error: {e}")
+            prosody_audio = decoded
+        
+        # Dia TTS generation
+        tts_output = dia.generate(f"[emotion:neutral] {text}")
+        
+        # Convert to numpy and normalize
+        if torch.is_tensor(tts_output):
+            tts_np = tts_output.squeeze().cpu().numpy()
+        else:
+            tts_np = tts_output
+        
+        # Normalize audio output
+        if len(tts_np) > 0:
+            tts_np = tts_np / np.max(np.abs(tts_np)) * 0.95
+        
+        return (sr, tts_np), text
+        
+    except Exception as e:
+        print(f"Error in process_audio: {e}")
+        return None, f"Processing error: {str(e)}"
+
+# Gradio Interface
 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’s Response")
-    text_out   = gr.Textbox(label="Generated Text")
-    send_btn.click(process_audio, inputs=audio_in, outputs=[audio_out, text_out])
+    gr.Markdown("# Maya-AI: Supernatural Conversational Agent")
+    gr.Markdown("Record audio to interact with the AI agent that understands emotions and responds naturally.")
+    
+    with gr.Row():
+        with gr.Column():
+            audio_in = gr.Audio(
+                source="microphone", 
+                type="numpy", 
+                label="Record Your Voice"
+            )
+            send_btn = gr.Button("Send", variant="primary")
+        
+        with gr.Column():
+            audio_out = gr.Audio(label="AI Response")
+            text_out = gr.Textbox(
+                label="Generated Text", 
+                lines=3,
+                placeholder="AI response will appear here..."
+            )
+    
+    # Event handler
+    send_btn.click(
+        fn=process_audio,
+        inputs=audio_in,
+        outputs=[audio_out, text_out]
+    )
 
 if __name__ == "__main__":
     demo.launch()