import random import re import numpy as np import torch import torchaudio from src.chatterbox.mtl_tts import ChatterboxMultilingualTTS, SUPPORTED_LANGUAGES import gradio as gr import spaces DEVICE = "cuda" if torch.cuda.is_available() else "cpu" print(f"🚀 Running on device: {DEVICE}") MODEL = None LANGUAGE_CONFIG = { "ar": {"audio": "https://storage.googleapis.com/chatterbox-demo-samples/mtl_prompts/ar_f/ar_prompts2.flac", "text": "في الشهر الماضي، وصلنا إلى معلم جديد بمليارين من المشاهدات على قناتنا على يوتيوب."}, "en": {"audio": "https://storage.googleapis.com/chatterbox-demo-samples/mtl_prompts/en_f1.flac", "text": "Last month, we reached a new milestone with two billion views on our YouTube channel."}, "fr": {"audio": "https://storage.googleapis.com/chatterbox-demo-samples/mtl_prompts/fr_f1.flac", "text": "Le mois dernier, nous avons atteint un nouveau jalon avec deux milliards de vues sur notre chaîne YouTube."}, "hi": {"audio": "https://storage.googleapis.com/chatterbox-demo-samples/mtl_prompts/hi_f1.flac", "text": "पिछले महीने हमने एक नया मील का पत्थर छुआ: हमारे YouTube चैनल पर दो अरब व्यूज़।"}, "tr": {"audio": "https://storage.googleapis.com/chatterbox-demo-samples/mtl_prompts/tr_m.flac", "text": "Geçen ay YouTube kanalımızda iki milyar görüntüleme ile yeni bir dönüm noktasına ulaştık."}, "zh": {"audio": "https://storage.googleapis.com/chatterbox-demo-samples/mtl_prompts/zh_f2.flac", "text": "上个月，我们达到了一个新的里程碑。 我们的YouTube频道观看次数达到了二十亿次，这绝对令人难以置信。"}, } def default_audio_for_ui(lang: str) -> str | None: return LANGUAGE_CONFIG.get(lang, {}).get("audio") def default_text_for_ui(lang: str) -> str: return LANGUAGE_CONFIG.get(lang, {}).get("text", "") def get_supported_languages_display() -> str: items = [f"**{name}** (`{code}`)" for code, name in sorted(SUPPORTED_LANGUAGES.items())] mid = len(items)//2 return ( f"### 🌍 Supported Languages ({len(SUPPORTED_LANGUAGES)} total)\n" f"{' • '.join(items[:mid])}\n\n{' • '.join(items[mid:])}" ) def get_or_load_model(): global MODEL if MODEL is None: print("Model not loaded, initializing...") MODEL = ChatterboxMultilingualTTS.from_pretrained(DEVICE) if hasattr(MODEL, "to"): MODEL.to(DEVICE) print(f"✅ Model loaded successfully on {DEVICE}") return MODEL try: get_or_load_model() except Exception as e: print(f"CRITICAL: Failed to load model. Error: {e}") def set_seed(seed: int): torch.manual_seed(seed) if DEVICE == "cuda": torch.cuda.manual_seed(seed) torch.cuda.manual_seed_all(seed) random.seed(seed) np.random.seed(seed) def resolve_audio_prompt(language_id: str, provided_path: str | None) -> str | None: if provided_path and str(provided_path).strip(): return provided_path return LANGUAGE_CONFIG.get(language_id, {}).get("audio") # ============================ # SMART CHUNKING (TỐI ƯU) # ============================ def smart_chunk_text(text: str, max_chars: int = 500) -> list[str]: """ Chia text thành các đoạn (chunk) ngắn: - Ưu tiên tách theo câu. - Nếu câu quá dài thì tách tiếp theo từ. - Gộp nhiều câu nhỏ vào 1 chunk để giảm số lần gọi model. """ # Normalize khoảng trắng text = re.sub(r"\s+", " ", text.strip()) if not text: return [] if len(text) <= max_chars: return [text] # Hỗ trợ nhiều dấu câu đa ngôn ngữ: . ! ? … ؟ ، : ؛ । sentences = re.split(r'(?<=[\.!\?…؟،:؛।])\s+', text) chunks: list[str] = [] current = "" for sent in sentences: sent = sent.strip() if not sent: continue # Nếu bản thân câu đã dài hơn max_chars -> chia mềm theo từ if len(sent) > max_chars: words = sent.split() temp = "" for w in words: if len(temp) + len(w) + 1 > max_chars: if temp: chunks.append(temp.strip()) temp = "" temp += w + " " if temp: chunks.append(temp.strip()) continue # Nếu gộp thêm câu mà vẫn không vượt max_chars -> gộp chung if len(current) + len(sent) + 1 <= max_chars: current += sent + " " else: if current: chunks.append(current.strip()) current = sent + " " if current: chunks.append(current.strip()) return [c for c in chunks if c] def concat_audio_torch(chunks: list[torch.Tensor], crossfade_ms: int = 10, sr: int = 24000) -> torch.Tensor: """ Nối nhiều đoạn audio (1D tensor) bằng crossfade nhẹ để tránh tiếng "click". """ if not chunks: return torch.empty(0) if len(chunks) == 1 or crossfade_ms <= 0: return torch.cat(chunks, dim=-1) output = chunks[0] crossfade = int(crossfade_ms * sr / 1000) for i in range(1, len(chunks)): a = output b = chunks[i] # Đảm bảo crossfade không lớn hơn độ dài đoạn cf = min(crossfade, a.shape[-1], b.shape[-1]) if cf <= 0: output = torch.cat([a, b], dim=-1) continue fade_out = torch.linspace(1.0, 0.0, steps=cf, device=a.device, dtype=a.dtype) fade_in = torch.linspace(0.0, 1.0, steps=cf, device=b.device, dtype=b.dtype) a_tail = a[..., -cf:] * fade_out b_head = b[..., :cf] * fade_in mixed = a_tail + b_head a_main = a[..., :-cf] b_rest = b[..., cf:] output = torch.cat([a_main, mixed, b_rest], dim=-1) return output @spaces.GPU def generate_tts_audio( text_input: str, language_id: str, audio_prompt_path_input: str = None, exaggeration_input: float = 0.5, temperature_input: float = 0.8, seed_num_input: int = 0, cfgw_input: float = 0.5 ): current_model = get_or_load_model() if current_model is None: raise RuntimeError("TTS model not loaded.") # --- SEED LOGIC --- if seed_num_input == 0: seed_num_input = random.randint(1, 2**32 - 1) print(f"🌱 Random seed generated: {seed_num_input}") else: print(f"🌱 Using provided seed: {seed_num_input}") set_seed(int(seed_num_input)) chosen_prompt = audio_prompt_path_input or default_audio_for_ui(language_id) generate_kwargs = { "exaggeration": exaggeration_input, "temperature": temperature_input, "cfg_weight": cfgw_input, } if chosen_prompt: generate_kwargs["audio_prompt_path"] = chosen_prompt # 💡 DÙNG SMART CHUNKING TỐI ƯU chunks = smart_chunk_text(text_input, max_chars=500) print(f"📚 Total chunks: {len(chunks)}") all_audio: list[torch.Tensor] = [] for idx, chunk in enumerate(chunks, start=1): print(f"🎧 Rendering chunk {idx}/{len(chunks)} (len={len(chunk)} chars)") wav = current_model.generate(chunk, language_id=language_id, **generate_kwargs) all_audio.append(wav.squeeze(0).cpu()) # 🔗 NỐI AUDIO VỚI CROSSFADE NHẸ final_audio = concat_audio_torch( all_audio, crossfade_ms=12, sr=current_model.sr ) # RETURN AUDIO + SEED return (current_model.sr, final_audio.numpy()), str(seed_num_input) # ============================ # GRADIO UI # ============================ with gr.Blocks() as demo: gr.Markdown(""" # 🎙️ Multi Language Realistic Voice Cloner Generate long-form multilingual speech with reference audio styling and smart chunking (crossfaded). """) gr.Markdown(get_supported_languages_display()) with gr.Row(): with gr.Column(): initial_lang = "en" text = gr.Textbox( value=default_text_for_ui(initial_lang), label="Text to synthesize", lines=8 ) language_id = gr.Dropdown( choices=list(ChatterboxMultilingualTTS.get_supported_languages().keys()), value=initial_lang, label="Language" ) ref_wav = gr.Audio( sources=["upload", "microphone"], type="filepath", label="Reference Audio (Optional)", value=default_audio_for_ui(initial_lang) ) exaggeration = gr.Slider(0.25, 2, step=.05, label="Exaggeration", value=.5) cfg_weight = gr.Slider(0.2, 1, step=.05, label="CFG Weight", value=0.5) with gr.Accordion("Advanced", open=False): seed_num = gr.Number(value=0, label="Random Seed (0=random)") temp = gr.Slider(0.05, 5, step=.05, label="Temperature", value=.8) run_btn = gr.Button("Generate", variant="primary") # OUTPUT COLUMN with gr.Column(): audio_output = gr.Audio(label="Output Audio") seed_output = gr.Textbox(label="Seed Used", interactive=False) def on_lang_change(lang, current_ref, current_text): return default_audio_for_ui(lang), default_text_for_ui(lang) language_id.change( fn=on_lang_change, inputs=[language_id, ref_wav, text], outputs=[ref_wav, text], show_progress=False ) # CONNECT BUTTON run_btn.click( fn=generate_tts_audio, inputs=[text, language_id, ref_wav, exaggeration, temp, seed_num, cfg_weight], outputs=[audio_output, seed_output], ) demo.launch(mcp_server=True, share=True)