Update app.py

app.py

@@ -1,5 +1,6 @@
 import spaces
 import os
+import numpy as np
 from huggingface_hub import login
 import gradio as gr
 from cached_path import cached_path
@@ -18,23 +19,47 @@ from f5_tts.infer.utils_infer import (
 # Retrieve token from secrets
 hf_token = os.getenv("HUGGINGFACEHUB_API_TOKEN")
 
-
 # Log in to Hugging Face
 if hf_token:
     login(token=hf_token)
 
-def post_process(text):
+def post_process(text: str):
+    """
+    Chuẩn hóa text trước khi synthesize.
+    """
     text = " " + text + " "
     text = text.replace(" . . ", " . ")
-    text = " " + text + " "
     text = text.replace(" .. ", " . ")
-    text = " " + text + " "
     text = text.replace(" , , ", " , ")
-    text = " " + text + " "
     text = text.replace(" ,, ", " , ")
-    text = " " + text + " "
     text = text.replace('"', "")
-    return " ".join(text.split())
+    return " ".join(text.split()).strip()
+
+def synthesize_with_pauses(ref_audio, ref_text, text, model, vocoder, speed=1.0, volume=1.0, pause_duration=1.0):
+    """
+    Chia text theo dấu chấm, synthesize từng câu và ghép lại với khoảng im lặng.
+    """
+    processed_text = post_process(TTSnorm(text)).lower()
+    sentences = [s.strip() for s in processed_text.split(".") if s.strip()]
+
+    all_waves = []
+    sr = 22050  # sample rate mặc định (cập nhật sau từ infer_process)
+
+    for idx, sentence in enumerate(sentences):
+        wave, sr, _ = infer_process(ref_audio, ref_text.lower(), sentence, model, vocoder, speed=speed)
+        wave = np.clip(wave * volume, -1.0, 1.0)
+        all_waves.append(wave)
+
+        # Thêm im lặng giữa các câu (trừ câu cuối)
+        if idx < len(sentences) - 1:
+            silence = np.zeros(int(sr * pause_duration), dtype=np.float32)
+            all_waves.append(silence)
+
+    if all_waves:
+        final_wave = np.concatenate(all_waves)
+    else:
+        final_wave = np.array([], dtype=np.float32)
+    return final_wave, sr
 
 # Load models
 vocoder = load_vocoder()
@@ -46,20 +71,26 @@ model = load_model(
 )
 
 @spaces.GPU
-def infer_tts(ref_audio_orig: str, gen_text: str, speed: float = 1.0, request: gr.Request = None):
-
+def infer_tts(ref_audio_orig: str, gen_text: str, speed: float = 1.0, volume: float = 1.0, pause: float = 1.0, request: gr.Request = None):
     if not ref_audio_orig:
         raise gr.Error("Please upload a sample audio file.")
     if not gen_text.strip():
         raise gr.Error("Please enter the text content to generate voice.")
     if len(gen_text.split()) > 1000:
         raise gr.Error("Please enter text content with less than 1000 words.")
-    
+
     try:
+        # Tiền xử lý sample voice
         ref_audio, ref_text = preprocess_ref_audio_text(ref_audio_orig, "")
-        final_wave, final_sample_rate, spectrogram = infer_process(
-            ref_audio, ref_text.lower(), post_process(TTSnorm(gen_text)).lower(), model, vocoder, speed=speed
+
+        # Synthesize với ngắt nghỉ
+        final_wave, final_sample_rate = synthesize_with_pauses(
+            ref_audio, ref_text, gen_text, model, vocoder, speed=speed, volume=volume, pause_duration=pause
         )
+
+        # Tạo spectrogram (dùng đoạn text đầy đủ để hiển thị, nhưng không tái synthesize)
+        _, _, spectrogram = infer_process(ref_audio, ref_text.lower(), post_process(TTSnorm(gen_text)).lower(), model, vocoder, speed=speed)
+
         with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tmp_spectrogram:
             spectrogram_path = tmp_spectrogram.name
             save_spectrogram(spectrogram, spectrogram_path)
@@ -71,16 +102,17 @@ def infer_tts(ref_audio_orig: str, gen_text: str, speed: float = 1.0, request: g
 # Gradio UI
 with gr.Blocks(theme=gr.themes.Soft()) as demo:
     gr.Markdown("""
-    # 🎤 F5-TTS: Vietnamese Text-to-Speech Synthesis.
-    # The model was trained with approximately 1000 hours of data on a RTX 3090 GPU. 
-    Enter text and upload a sample voice to generate natural speech.
+    # 🎤 Chương trình chuyển đổi text thành giọng nói.
     """)
     
     with gr.Row():
         ref_audio = gr.Audio(label="🔊 Sample Voice", type="filepath")
         gen_text = gr.Textbox(label="📝 Text", placeholder="Enter the text to generate voice...", lines=3)
     
-    speed = gr.Slider(0.3, 2.0, value=1.0, step=0.1, label="⚡ Speed")
+    speed = gr.Slider(0.3, 2.0, value=0.95, step=0.01, label="⚡ Speed")
+    volume = gr.Slider(0.1, 2.0, value=1.0, step=0.01, label="🔊 Volume")
+    pause = gr.Slider(0.0, 3.0, value=1.1, step=0.01, label="⏸ Pause between sentences (seconds)")
+    
     btn_synthesize = gr.Button("🔥 Generate Voice")
     
     with gr.Row():
@@ -97,7 +129,11 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
         interactive=False
     )
 
-    btn_synthesize.click(infer_tts, inputs=[ref_audio, gen_text, speed], outputs=[output_audio, output_spectrogram])
+    btn_synthesize.click(
+        infer_tts,
+        inputs=[ref_audio, gen_text, speed, volume, pause],
+        outputs=[output_audio, output_spectrogram]
+    )
 
-# Run Gradio with share=True to get a gradio.live link
+# Run Gradio
 demo.queue().launch()