Upload 6 files

app.py

@@ -1,15 +1,17 @@
 import time
 import os
+import spaces
 import torch
 import gradio as gr
 from huggingface_hub import snapshot_download
-from mira.model import MiraTTS
-from mira.utils import split_text
 
 # Model config
 HF_MODEL_ID = "dolly-vn/Vira-TTS"
 MODEL_PATH = "model_pretrained"
 
+# Global model variable
+mira_tts = None
+
 def download_model_if_needed():
     """Download model from HuggingFace if not exists locally."""
     if not os.path.exists(MODEL_PATH) or not os.listdir(MODEL_PATH):
@@ -23,18 +25,28 @@ def download_model_if_needed():
     else:
         print(f"✅ Model found at: {MODEL_PATH}")
 
-# Download model if needed
+# Download model at startup (no GPU needed)
 download_model_if_needed()
 
-print("🔄 Loading Vira-TTS...")
-mira_tts = MiraTTS(MODEL_PATH)
-print("✅ Model loaded!")
-
 SAMPLE_RATE = 48000
 
 
+def get_model():
+    """Lazy load model when GPU is available."""
+    global mira_tts
+    if mira_tts is None:
+        from mira.model import MiraTTS
+        from mira.utils import split_text
+        print("🔄 Loading Vira-TTS...")
+        mira_tts = MiraTTS(MODEL_PATH)
+        print("✅ Model loaded!")
+    return mira_tts
+
+
+@spaces.GPU
 def generate_speech(text: str, reference_audio: str):
     """Generate speech from text using reference audio for voice cloning."""
+    from mira.utils import split_text
 
     if not text.strip():
         return None, "Vui lòng nhập văn bản."
@@ -43,8 +55,11 @@ def generate_speech(text: str, reference_audio: str):
         return None, "Vui lòng upload file audio tham chiếu."
 
     try:
+        # Get model (lazy load with GPU)
+        model = get_model()
+
         # Encode reference audio
-        context_tokens = mira_tts.encode_audio(reference_audio)
+        context_tokens = model.encode_audio(reference_audio)
 
         # Split text into sentences
         sentences = split_text(text)
@@ -53,11 +68,9 @@ def generate_speech(text: str, reference_audio: str):
         start_time = time.time()
 
         if len(sentences) == 1:
-            # Single sentence - use generate
-            audio = mira_tts.generate(sentences[0], context_tokens)
+            audio = model.generate(sentences[0], context_tokens)
         else:
-            # Multiple sentences - use batch_generate
-            audio = mira_tts.batch_generate(sentences, [context_tokens])
+            audio = model.batch_generate(sentences, [context_tokens])
 
         inference_time = time.time() - start_time
 
@@ -66,13 +79,13 @@ def generate_speech(text: str, reference_audio: str):
         audio_duration = len(audio_np) / SAMPLE_RATE
         rtf = inference_time / audio_duration
 
-        # Create stats message
         stats = f"📝 Số câu: {len(sentences)} | ⏱️ Inference: {inference_time:.2f}s | 🎵 Audio: {audio_duration:.2f}s | 📊 RTF: {rtf:.4f}"
 
         return (SAMPLE_RATE, audio_np), stats
 
     except Exception as e:
-        return None, f"Lỗi: {str(e)}"
+        import traceback
+        return None, f"Lỗi: {str(e)}\n{traceback.format_exc()}"
 
 
 # Create Gradio interface

requirements.txt

@@ -1,5 +1,6 @@
 --extra-index-url https://huggingface.github.io/lmdeploy-wheel-index/
 
+spaces
 lmdeploy
 librosa
 einops
@@ -8,8 +9,8 @@ gradio
 soundfile
 soe-vinorm
 huggingface_hub
-fastaudiosr @ https://github.com/ysharma3501/FlashSR/archive/refs/heads/master.zip
-ncodec @ https://github.com/ysharma3501/FastBiCodec/archive/refs/heads/master.zip
 torch
 torchaudio
-omegaconf
+omegaconf
+fastaudiosr @ https://github.com/ysharma3501/FlashSR/archive/refs/heads/master.zip
+ncodec @ https://github.com/ysharma3501/FastBiCodec/archive/refs/heads/master.zip