Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
|
@@ -70,38 +70,26 @@ def resolve_audio_prompt(language_id: str, provided_path: str | None) -> str | N
|
|
| 70 |
return LANGUAGE_CONFIG.get(language_id, {}).get("audio")
|
| 71 |
|
| 72 |
|
| 73 |
-
#
|
| 74 |
-
#
|
| 75 |
-
#
|
| 76 |
|
| 77 |
-
def
|
| 78 |
"""
|
| 79 |
-
|
| 80 |
-
-
|
| 81 |
-
-
|
|
|
|
| 82 |
"""
|
| 83 |
-
#
|
| 84 |
-
text = re.sub(r"([\.!\?])\s+", r"\1 ... ", text)
|
| 85 |
-
|
| 86 |
-
# Nghỉ ngắn sau dấu phẩy
|
| 87 |
-
text = re.sub(r"(,)\s+", r"\1 … ", text)
|
| 88 |
-
|
| 89 |
-
return text.strip()
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
# ===========================================================
|
| 93 |
-
# SMART CHUNKING
|
| 94 |
-
# ===========================================================
|
| 95 |
-
|
| 96 |
-
def smart_chunk_text(text: str, max_chars: int = 500) -> list[str]:
|
| 97 |
-
|
| 98 |
text = re.sub(r"\s+", " ", text.strip())
|
| 99 |
if not text:
|
| 100 |
return []
|
| 101 |
if len(text) <= max_chars:
|
| 102 |
return [text]
|
| 103 |
|
| 104 |
-
|
|
|
|
| 105 |
|
| 106 |
chunks: list[str] = []
|
| 107 |
current = ""
|
|
@@ -111,10 +99,26 @@ def smart_chunk_text(text: str, max_chars: int = 500) -> list[str]:
|
|
| 111 |
if not sent:
|
| 112 |
continue
|
| 113 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 114 |
if len(current) + len(sent) + 1 <= max_chars:
|
| 115 |
current += sent + " "
|
| 116 |
else:
|
| 117 |
-
|
|
|
|
| 118 |
current = sent + " "
|
| 119 |
|
| 120 |
if current:
|
|
@@ -123,32 +127,45 @@ def smart_chunk_text(text: str, max_chars: int = 500) -> list[str]:
|
|
| 123 |
return [c for c in chunks if c]
|
| 124 |
|
| 125 |
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
def concat_audio_with_pause(chunks, pause_ms: int = 220, sr: int = 24000):
|
| 131 |
"""
|
| 132 |
-
|
| 133 |
"""
|
| 134 |
if not chunks:
|
| 135 |
return torch.empty(0)
|
| 136 |
|
| 137 |
-
|
| 138 |
-
|
|
|
|
|
|
|
|
|
|
| 139 |
|
| 140 |
-
for i
|
| 141 |
-
output
|
| 142 |
-
|
| 143 |
-
if i < len(chunks) - 1:
|
| 144 |
-
output.append(silence.clone())
|
| 145 |
|
| 146 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 147 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 148 |
|
| 149 |
-
# ===========================================================
|
| 150 |
-
# TTS GENERATE
|
| 151 |
-
# ===========================================================
|
| 152 |
|
| 153 |
@spaces.GPU
|
| 154 |
def generate_tts_audio(
|
|
@@ -165,46 +182,55 @@ def generate_tts_audio(
|
|
| 165 |
if current_model is None:
|
| 166 |
raise RuntimeError("TTS model not loaded.")
|
| 167 |
|
|
|
|
| 168 |
if seed_num_input == 0:
|
| 169 |
seed_num_input = random.randint(1, 2**32 - 1)
|
| 170 |
-
|
|
|
|
|
|
|
| 171 |
|
| 172 |
-
|
| 173 |
-
text_input = natural_pause_text(text_input)
|
| 174 |
|
| 175 |
chosen_prompt = audio_prompt_path_input or default_audio_for_ui(language_id)
|
| 176 |
generate_kwargs = {
|
| 177 |
"exaggeration": exaggeration_input,
|
| 178 |
"temperature": temperature_input,
|
| 179 |
"cfg_weight": cfgw_input,
|
| 180 |
-
"audio_prompt_path": chosen_prompt
|
| 181 |
}
|
|
|
|
|
|
|
| 182 |
|
|
|
|
| 183 |
chunks = smart_chunk_text(text_input, max_chars=500)
|
|
|
|
|
|
|
| 184 |
all_audio: list[torch.Tensor] = []
|
| 185 |
|
| 186 |
-
for chunk in chunks:
|
|
|
|
| 187 |
wav = current_model.generate(chunk, language_id=language_id, **generate_kwargs)
|
| 188 |
all_audio.append(wav.squeeze(0).cpu())
|
| 189 |
|
| 190 |
-
#
|
| 191 |
-
final_audio =
|
| 192 |
all_audio,
|
| 193 |
-
|
| 194 |
sr=current_model.sr
|
| 195 |
)
|
| 196 |
|
|
|
|
| 197 |
return (current_model.sr, final_audio.numpy()), str(seed_num_input)
|
| 198 |
|
| 199 |
|
| 200 |
-
|
| 201 |
-
#
|
| 202 |
-
#
|
|
|
|
| 203 |
|
| 204 |
with gr.Blocks() as demo:
|
| 205 |
gr.Markdown("""
|
| 206 |
-
# 🎙️ Multi Language Realistic Voice Cloner
|
| 207 |
-
|
| 208 |
""")
|
| 209 |
|
| 210 |
gr.Markdown(get_supported_languages_display())
|
|
@@ -212,16 +238,22 @@ with gr.Blocks() as demo:
|
|
| 212 |
with gr.Row():
|
| 213 |
with gr.Column():
|
| 214 |
initial_lang = "en"
|
| 215 |
-
text = gr.Textbox(
|
|
|
|
|
|
|
|
|
|
|
|
|
| 216 |
language_id = gr.Dropdown(
|
| 217 |
choices=list(ChatterboxMultilingualTTS.get_supported_languages().keys()),
|
| 218 |
value=initial_lang,
|
| 219 |
label="Language"
|
| 220 |
)
|
| 221 |
-
ref_wav = gr.Audio(
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
|
|
|
|
|
|
| 225 |
exaggeration = gr.Slider(0.25, 2, step=.05, label="Exaggeration", value=.5)
|
| 226 |
cfg_weight = gr.Slider(0.2, 1, step=.05, label="CFG Weight", value=0.5)
|
| 227 |
|
|
@@ -231,6 +263,7 @@ with gr.Blocks() as demo:
|
|
| 231 |
|
| 232 |
run_btn = gr.Button("Generate", variant="primary")
|
| 233 |
|
|
|
|
| 234 |
with gr.Column():
|
| 235 |
audio_output = gr.Audio(label="Output Audio")
|
| 236 |
seed_output = gr.Textbox(label="Seed Used", interactive=False)
|
|
@@ -241,13 +274,15 @@ with gr.Blocks() as demo:
|
|
| 241 |
language_id.change(
|
| 242 |
fn=on_lang_change,
|
| 243 |
inputs=[language_id, ref_wav, text],
|
| 244 |
-
outputs=[ref_wav, text]
|
|
|
|
| 245 |
)
|
| 246 |
|
|
|
|
| 247 |
run_btn.click(
|
| 248 |
fn=generate_tts_audio,
|
| 249 |
inputs=[text, language_id, ref_wav, exaggeration, temp, seed_num, cfg_weight],
|
| 250 |
-
outputs=[audio_output, seed_output]
|
| 251 |
)
|
| 252 |
|
| 253 |
demo.launch(mcp_server=True, share=True)
|
|
|
|
| 70 |
return LANGUAGE_CONFIG.get(language_id, {}).get("audio")
|
| 71 |
|
| 72 |
|
| 73 |
+
# ============================
|
| 74 |
+
# SMART CHUNKING (TỐI ƯU)
|
| 75 |
+
# ============================
|
| 76 |
|
| 77 |
+
def smart_chunk_text(text: str, max_chars: int = 500) -> list[str]:
|
| 78 |
"""
|
| 79 |
+
Chia text thành các đoạn (chunk) ngắn:
|
| 80 |
+
- Ưu tiên tách theo câu.
|
| 81 |
+
- Nếu câu quá dài thì tách tiếp theo từ.
|
| 82 |
+
- Gộp nhiều câu nhỏ vào 1 chunk để giảm số lần gọi model.
|
| 83 |
"""
|
| 84 |
+
# Normalize khoảng trắng
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 85 |
text = re.sub(r"\s+", " ", text.strip())
|
| 86 |
if not text:
|
| 87 |
return []
|
| 88 |
if len(text) <= max_chars:
|
| 89 |
return [text]
|
| 90 |
|
| 91 |
+
# Hỗ trợ nhiều dấu câu đa ngôn ngữ: . ! ? … ؟ ، : ؛ ।
|
| 92 |
+
sentences = re.split(r'(?<=[\.!\?…؟،:؛।])\s+', text)
|
| 93 |
|
| 94 |
chunks: list[str] = []
|
| 95 |
current = ""
|
|
|
|
| 99 |
if not sent:
|
| 100 |
continue
|
| 101 |
|
| 102 |
+
# Nếu bản thân câu đã dài hơn max_chars -> chia mềm theo từ
|
| 103 |
+
if len(sent) > max_chars:
|
| 104 |
+
words = sent.split()
|
| 105 |
+
temp = ""
|
| 106 |
+
for w in words:
|
| 107 |
+
if len(temp) + len(w) + 1 > max_chars:
|
| 108 |
+
if temp:
|
| 109 |
+
chunks.append(temp.strip())
|
| 110 |
+
temp = ""
|
| 111 |
+
temp += w + " "
|
| 112 |
+
if temp:
|
| 113 |
+
chunks.append(temp.strip())
|
| 114 |
+
continue
|
| 115 |
+
|
| 116 |
+
# Nếu gộp thêm câu mà vẫn không vượt max_chars -> gộp chung
|
| 117 |
if len(current) + len(sent) + 1 <= max_chars:
|
| 118 |
current += sent + " "
|
| 119 |
else:
|
| 120 |
+
if current:
|
| 121 |
+
chunks.append(current.strip())
|
| 122 |
current = sent + " "
|
| 123 |
|
| 124 |
if current:
|
|
|
|
| 127 |
return [c for c in chunks if c]
|
| 128 |
|
| 129 |
|
| 130 |
+
def concat_audio_torch(chunks: list[torch.Tensor],
|
| 131 |
+
crossfade_ms: int = 10,
|
| 132 |
+
sr: int = 24000) -> torch.Tensor:
|
|
|
|
|
|
|
| 133 |
"""
|
| 134 |
+
Nối nhiều đoạn audio (1D tensor) bằng crossfade nhẹ để tránh tiếng "click".
|
| 135 |
"""
|
| 136 |
if not chunks:
|
| 137 |
return torch.empty(0)
|
| 138 |
|
| 139 |
+
if len(chunks) == 1 or crossfade_ms <= 0:
|
| 140 |
+
return torch.cat(chunks, dim=-1)
|
| 141 |
+
|
| 142 |
+
output = chunks[0]
|
| 143 |
+
crossfade = int(crossfade_ms * sr / 1000)
|
| 144 |
|
| 145 |
+
for i in range(1, len(chunks)):
|
| 146 |
+
a = output
|
| 147 |
+
b = chunks[i]
|
|
|
|
|
|
|
| 148 |
|
| 149 |
+
# Đảm bảo crossfade không lớn hơn độ dài đoạn
|
| 150 |
+
cf = min(crossfade, a.shape[-1], b.shape[-1])
|
| 151 |
+
if cf <= 0:
|
| 152 |
+
output = torch.cat([a, b], dim=-1)
|
| 153 |
+
continue
|
| 154 |
+
|
| 155 |
+
fade_out = torch.linspace(1.0, 0.0, steps=cf, device=a.device, dtype=a.dtype)
|
| 156 |
+
fade_in = torch.linspace(0.0, 1.0, steps=cf, device=b.device, dtype=b.dtype)
|
| 157 |
+
|
| 158 |
+
a_tail = a[..., -cf:] * fade_out
|
| 159 |
+
b_head = b[..., :cf] * fade_in
|
| 160 |
|
| 161 |
+
mixed = a_tail + b_head
|
| 162 |
+
a_main = a[..., :-cf]
|
| 163 |
+
b_rest = b[..., cf:]
|
| 164 |
+
|
| 165 |
+
output = torch.cat([a_main, mixed, b_rest], dim=-1)
|
| 166 |
+
|
| 167 |
+
return output
|
| 168 |
|
|
|
|
|
|
|
|
|
|
| 169 |
|
| 170 |
@spaces.GPU
|
| 171 |
def generate_tts_audio(
|
|
|
|
| 182 |
if current_model is None:
|
| 183 |
raise RuntimeError("TTS model not loaded.")
|
| 184 |
|
| 185 |
+
# --- SEED LOGIC ---
|
| 186 |
if seed_num_input == 0:
|
| 187 |
seed_num_input = random.randint(1, 2**32 - 1)
|
| 188 |
+
print(f"🌱 Random seed generated: {seed_num_input}")
|
| 189 |
+
else:
|
| 190 |
+
print(f"🌱 Using provided seed: {seed_num_input}")
|
| 191 |
|
| 192 |
+
set_seed(int(seed_num_input))
|
|
|
|
| 193 |
|
| 194 |
chosen_prompt = audio_prompt_path_input or default_audio_for_ui(language_id)
|
| 195 |
generate_kwargs = {
|
| 196 |
"exaggeration": exaggeration_input,
|
| 197 |
"temperature": temperature_input,
|
| 198 |
"cfg_weight": cfgw_input,
|
|
|
|
| 199 |
}
|
| 200 |
+
if chosen_prompt:
|
| 201 |
+
generate_kwargs["audio_prompt_path"] = chosen_prompt
|
| 202 |
|
| 203 |
+
# 💡 DÙNG SMART CHUNKING TỐI ƯU
|
| 204 |
chunks = smart_chunk_text(text_input, max_chars=500)
|
| 205 |
+
print(f"📚 Total chunks: {len(chunks)}")
|
| 206 |
+
|
| 207 |
all_audio: list[torch.Tensor] = []
|
| 208 |
|
| 209 |
+
for idx, chunk in enumerate(chunks, start=1):
|
| 210 |
+
print(f"🎧 Rendering chunk {idx}/{len(chunks)} (len={len(chunk)} chars)")
|
| 211 |
wav = current_model.generate(chunk, language_id=language_id, **generate_kwargs)
|
| 212 |
all_audio.append(wav.squeeze(0).cpu())
|
| 213 |
|
| 214 |
+
# 🔗 NỐI AUDIO VỚI CROSSFADE NHẸ
|
| 215 |
+
final_audio = concat_audio_torch(
|
| 216 |
all_audio,
|
| 217 |
+
crossfade_ms=12,
|
| 218 |
sr=current_model.sr
|
| 219 |
)
|
| 220 |
|
| 221 |
+
# RETURN AUDIO + SEED
|
| 222 |
return (current_model.sr, final_audio.numpy()), str(seed_num_input)
|
| 223 |
|
| 224 |
|
| 225 |
+
|
| 226 |
+
# ============================
|
| 227 |
+
# GRADIO UI
|
| 228 |
+
# ============================
|
| 229 |
|
| 230 |
with gr.Blocks() as demo:
|
| 231 |
gr.Markdown("""
|
| 232 |
+
# 🎙️ Multi Language Realistic Voice Cloner
|
| 233 |
+
Generate long-form multilingual speech with reference audio styling and smart chunking (crossfaded).
|
| 234 |
""")
|
| 235 |
|
| 236 |
gr.Markdown(get_supported_languages_display())
|
|
|
|
| 238 |
with gr.Row():
|
| 239 |
with gr.Column():
|
| 240 |
initial_lang = "en"
|
| 241 |
+
text = gr.Textbox(
|
| 242 |
+
value=default_text_for_ui(initial_lang),
|
| 243 |
+
label="Text to synthesize",
|
| 244 |
+
lines=8
|
| 245 |
+
)
|
| 246 |
language_id = gr.Dropdown(
|
| 247 |
choices=list(ChatterboxMultilingualTTS.get_supported_languages().keys()),
|
| 248 |
value=initial_lang,
|
| 249 |
label="Language"
|
| 250 |
)
|
| 251 |
+
ref_wav = gr.Audio(
|
| 252 |
+
sources=["upload", "microphone"],
|
| 253 |
+
type="filepath",
|
| 254 |
+
label="Reference Audio (Optional)",
|
| 255 |
+
value=default_audio_for_ui(initial_lang)
|
| 256 |
+
)
|
| 257 |
exaggeration = gr.Slider(0.25, 2, step=.05, label="Exaggeration", value=.5)
|
| 258 |
cfg_weight = gr.Slider(0.2, 1, step=.05, label="CFG Weight", value=0.5)
|
| 259 |
|
|
|
|
| 263 |
|
| 264 |
run_btn = gr.Button("Generate", variant="primary")
|
| 265 |
|
| 266 |
+
# OUTPUT COLUMN
|
| 267 |
with gr.Column():
|
| 268 |
audio_output = gr.Audio(label="Output Audio")
|
| 269 |
seed_output = gr.Textbox(label="Seed Used", interactive=False)
|
|
|
|
| 274 |
language_id.change(
|
| 275 |
fn=on_lang_change,
|
| 276 |
inputs=[language_id, ref_wav, text],
|
| 277 |
+
outputs=[ref_wav, text],
|
| 278 |
+
show_progress=False
|
| 279 |
)
|
| 280 |
|
| 281 |
+
# CONNECT BUTTON
|
| 282 |
run_btn.click(
|
| 283 |
fn=generate_tts_audio,
|
| 284 |
inputs=[text, language_id, ref_wav, exaggeration, temp, seed_num, cfg_weight],
|
| 285 |
+
outputs=[audio_output, seed_output],
|
| 286 |
)
|
| 287 |
|
| 288 |
demo.launch(mcp_server=True, share=True)
|