""" Gradio frontend for multilingual voice cloning. Two tabs: 1. Clone Voice — single language 2. Language Tour — same reference, many languages The language dropdown is populated from `VoiceCloneClient.discover_languages()`, which queries the backend `/languages` endpoint and filters LANGUAGE_CATALOG. The language code is passed end-to-end into synthesis so the selection actually changes the output — fixing the bug shown in the original PDF notes. """ from __future__ import annotations import logging import os from pathlib import Path from typing import Any import gradio as gr from app.client import VoiceCloneClient from app.config import LANGUAGE_CATALOG, get_settings from app.errors import VoiceCloneError log = logging.getLogger(__name__) # Pre-baked demo sentences per language. DEMO_SENTENCES = { "yue": "大家好，歡迎體驗我哋嘅多語言聲音克隆系統，廣東話都支援到。", "en": "Hello everyone, welcome to our multilingual voice cloning demonstration.", "zh": "大家好，欢迎体验我们的多语言语音克隆系统。", "ja": "皆さん、こんにちは。多言語音声クローニングのデモへようこそ。", "ko": "안녕하세요, 다국어 음성 복제 데모에 오신 것을 환영합니다.", "fr": "Bonjour à tous, bienvenue dans notre démo de clonage vocal multilingue.", "es": "Hola a todos, bienvenidos a nuestra demostración de clonación de voz multilingüe.", } def _format_label(code: str, meta: dict[str, str]) -> str: # Make Cantonese visually prominent. if code == "yue": return f"🇭🇰 {meta['native']} — {meta['english']} (yue)" return f"{meta['native']} — {meta['english']} ({code})" def _build_label_map(client: VoiceCloneClient) -> tuple[list[str], dict[str, str]]: """Returns (dropdown_choices, {label: code}).""" langs = client.discover_languages() if not langs: # Fall back to full catalog so the UI still loads if the backend is cold. langs = dict(LANGUAGE_CATALOG) # Keep Cantonese pinned to the top regardless of dict order. ordered = [] if "yue" in langs: ordered.append(("yue", langs["yue"])) for code, meta in langs.items(): if code != "yue": ordered.append((code, meta)) choices = [_format_label(c, m) for c, m in ordered] label_to_code = {_format_label(c, m): c for c, m in ordered} return choices, label_to_code def _format_health(health_map: dict[str, Any]) -> str: lines = [] for name, st in health_map.items(): mark = "✅" if st.reachable else "⚠️" lines.append(f"{mark} **{name}** — {st.detail}") return "\n".join(lines) def build_app(client: VoiceCloneClient | None = None) -> gr.Blocks: client = client or VoiceCloneClient() settings = client.settings choices, label_to_code = _build_label_map(client) cantonese_present = "yue" in client.discover_languages() default_label = next((lbl for lbl in choices if "yue" in lbl), choices[0] if choices else "") # ---- callbacks ----------------------------------------------------- def on_clone(ref_audio, text, language_label, exaggeration, cfg_weight): if not ref_audio: return None, "❌ Please upload reference audio first." if not language_label: return None, "❌ Please select a language." code = label_to_code.get(language_label) if not code: return None, f"❌ Unknown language: {language_label}" try: res = client.clone_voice( reference_audio_path=Path(ref_audio), text=text, language_code=code, exaggeration=float(exaggeration), cfg_weight=float(cfg_weight), ) except VoiceCloneError as e: return None, f"❌ {e}" status = ( f"✅ Generated in {res.latency_seconds:.1f}s via **{res.provider_name}** " f"— language={code} ({language_label}), model={res.model}, " f"sr={res.sample_rate} Hz" ) return str(res.wav_path), status def on_tour(ref_audio, text, exaggeration, cfg_weight): if not ref_audio: return [None] * len(choices) + ["❌ Please upload reference audio first."] results = client.language_tour( reference_audio_path=Path(ref_audio), text=text, exaggeration=float(exaggeration), cfg_weight=float(cfg_weight), language_codes=[label_to_code[lbl] for lbl in choices], ) audio_outputs: list[str | None] = [] lines = [] for (code, res, err) in results: meta = LANGUAGE_CATALOG.get(code, {"english": code, "native": code}) if res: audio_outputs.append(str(res.wav_path)) lines.append(f"✅ {meta['english']} ({code}) — {res.latency_seconds:.1f}s via {res.provider_name}") else: audio_outputs.append(None) lines.append(f"❌ {meta['english']} ({code}) — {err}") return audio_outputs + ["\n".join(lines)] def on_health(): return _format_health(client.health()) def on_language_change(label: str) -> str: code = label_to_code.get(label or "", "en") return DEMO_SENTENCES.get(code, "") # ---- layout -------------------------------------------------------- with gr.Blocks( title="ZhouFenMianFan — Multilingual Voice Cloning (Cloud)", theme=gr.themes.Soft(), ) as demo: gr.Markdown( "# 🎤 粥粉麵飯 — Multilingual Voice Cloning\n" "Cross-lingual voice cloning for Hong Kong users. Upload a " "**3–30 second reference clip** and generate speech in 粵語 " "(Cantonese), English, Mandarin, Japanese, Korean, French, " "and Spanish. Powered by a Chatterbox-compatible cloud API " "with a MiniMax adapter for Cantonese.\n" ) if not cantonese_present: gr.Markdown( "> ⚠️ **Cantonese (yue) is currently unavailable.** " "Configure `MINIMAX_API_KEY` in your `.env` to enable it." ) with gr.Accordion("Backend status", open=False): health_md = gr.Markdown(_format_health(client.health())) refresh_btn = gr.Button("Refresh health", size="sm") refresh_btn.click(on_health, outputs=health_md) with gr.Tab("Clone Voice"): with gr.Row(): with gr.Column(): ref1 = gr.Audio( type="filepath", label=f"Reference voice (WAV, " f"{settings.min_reference_seconds:.0f}–" f"{settings.max_reference_seconds:.0f}s)", sources=["upload", "microphone"], ) lang1 = gr.Dropdown( choices=choices, value=default_label, label="Target language", info="Cantonese (粵語) is routed through the MiniMax adapter.", ) text1 = gr.Textbox( lines=3, value=DEMO_SENTENCES.get("yue", "") if cantonese_present else DEMO_SENTENCES.get("en", ""), label="Text to synthesise", ) with gr.Row(): exag1 = gr.Slider( settings.exaggeration_range[0], settings.exaggeration_range[1], value=settings.default_exaggeration, step=0.05, label="Exaggeration", ) cfg1 = gr.Slider( settings.cfg_weight_range[0], settings.cfg_weight_range[1], value=settings.default_cfg_weight, step=0.05, label="CFG weight", ) go1 = gr.Button("🎙️ Generate", variant="primary") with gr.Column(): audio1 = gr.Audio(type="filepath", label="Cloned output", interactive=False) status1 = gr.Markdown("") lang1.change(on_language_change, inputs=lang1, outputs=text1) go1.click( on_clone, inputs=[ref1, text1, lang1, exag1, cfg1], outputs=[audio1, status1], ) with gr.Tab("Language Tour"): gr.Markdown( "Clone the same reference voice across every supported language " "using one shared sentence." ) with gr.Row(): ref2 = gr.Audio(type="filepath", label="Reference voice (WAV, 3–30s)", sources=["upload", "microphone"]) text2 = gr.Textbox( lines=2, value="Hello, welcome to our AI voice cloning demonstration.", label="Shared text (model will synthesise this same " "sentence in every language)", ) with gr.Row(): exag2 = gr.Slider( settings.exaggeration_range[0], settings.exaggeration_range[1], value=settings.default_exaggeration, step=0.05, label="Exaggeration", ) cfg2 = gr.Slider( settings.cfg_weight_range[0], settings.cfg_weight_range[1], value=settings.default_cfg_weight, step=0.05, label="CFG weight", ) go2 = gr.Button("🌏 Generate language tour", variant="primary") audio_widgets: list[gr.Audio] = [] for lbl in choices: audio_widgets.append(gr.Audio(type="filepath", label=lbl, interactive=False)) status2 = gr.Markdown("") go2.click( on_tour, inputs=[ref2, text2, exag2, cfg2], outputs=audio_widgets + [status2], ) gr.Markdown( "---\n" "**Guidance**: 10–30 s of clean speech works best. Avoid music, " "background noise, and clipping. 24 kHz / 32 kHz mono WAV " "produces the cleanest clones. Long-form texts over 500 characters " "are split by the backend automatically.\n" ) return demo def main() -> None: logging.basicConfig( level=os.getenv("LOG_LEVEL", "INFO"), format="%(asctime)s %(levelname)s %(name)s: %(message)s", ) settings = get_settings() demo = build_app() demo.queue().launch( server_name=settings.gradio_server_name, server_port=settings.gradio_server_port, share=False, show_error=True, ) if __name__ == "__main__": main()