Update app.py

app.py

@@ -1,4 +1,3 @@
-import spaces
 import os
 from huggingface_hub import login
 import gradio as gr
@@ -40,11 +39,10 @@ vocoder = load_vocoder()
 model = load_model(
     DiT,
     dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4),
-    ckpt_path=str(cached_path("hf://hynt/F5-TTS-Vietnamese-100h/model_500000.pt")),
-    vocab_file=str(cached_path("hf://hynt/F5-TTS-Vietnamese-100h/vocab.txt")),
+    ckpt_path=str(cached_path("hf://nam194/F5-TTS-Vietnamese/model_350000.safetensors")),
+    vocab_file=str(cached_path("hf://nam194/F5-TTS-Vietnamese/vocab.txt")),
 )
 
-@spaces.GPU
 def infer_tts(ref_audio_orig: str, gen_text: str, speed: float = 1.0, request: gr.Request = None):
 
     if not ref_audio_orig:
@@ -68,11 +66,12 @@ def infer_tts(ref_audio_orig: str, gen_text: str, speed: float = 1.0, request: g
         raise gr.Error(f"Error generating voice: {e}")
 
 # Gradio UI
-with gr.Blocks(theme=gr.themes.Soft()) as demo:
+with gr.Blocks() as demo:
     gr.Markdown("""
     # 🎤 F5-TTS: Vietnamese Text-to-Speech Synthesis.
-    # The model was trained for 500.000 steps with approximately 150 hours of data on an RTX 3090 GPU. 
+    # The model was trained for 350.000 steps with approximately 1000 hours of data on an RTX 3090 GPU. 
     Enter text and upload a sample voice to generate natural speech.
+    CPU inference time may take minutes.
     """)
     
     with gr.Row():
@@ -90,7 +89,7 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
         value="""1. This model may not perform well with numerical characters, dates, special characters, etc. => A text normalization module is needed.
 2. The rhythm of some generated audios may be inconsistent or choppy => It is recommended to select clearly pronounced sample audios with minimal pauses for better synthesis quality.
 3. Default, reference audio text uses the whisper-large-v3-turbo model, which may not always accurately recognize Vietnamese, resulting in poor voice synthesis quality.
-4. Checkpoint is stopped at step 500.000, trained with 150 hours of public data => Voice cloning for non-native voices may not be perfectly accurate.
+4. Checkpoint is stopped at step 350.000, trained with 1000 hours of public data => Voice cloning for non-native voices may not be perfectly accurate.
 5. Inference with overly long paragraphs may produce poor results.""", 
         label="❗ Model Limitations",
         lines=5,