🎤 SoulX-Singer
+
+ Official inference code for
+ SoulX-Singer: Towards High-Quality Zero-Shot Singing Voice Synthesis
+
+ 
+
+ 
+ Official inference code for
+ SoulX-Singer: Towards High-Quality Zero-Shot Singing Voice Synthesis
+
+
+
+
+ 
+
+
+
+
+
+ 
+
+
+
Refer to Edit Tutorial for metadata editing
', + zh='metadata 编辑请参考 编辑教程
', + ), + prompt_wav_label=dict(en="Prompt WAV (reference)", zh="Prompt WAV（参考音色）"), + generated_audio_label=dict(en="Generated merged audio", zh="合成结果音频"), + prompt_lyric_lang_label=dict(en="Prompt lyric language", zh="Prompt 歌词语种"), + target_lyric_lang_label=dict(en="Target lyric language", zh="Target 歌词语种"), + lyric_lang_mandarin=dict(en="Mandarin", zh="普通话"), + lyric_lang_cantonese=dict(en="Cantonese", zh="粤语"), + lyric_lang_english=dict(en="English", zh="英语"), + warn_missing_synthesis=dict(en="Please provide prompt WAV, prompt metadata, and target metadata", zh="请提供 Prompt WAV、Prompt metadata 与 Target metadata"), + prompt_vocal_sep_label=dict(en="Prompt vocal separation", zh="Prompt人声分离"), + target_vocal_sep_label=dict(en="Target vocal separation", zh="Target人声分离"), + auto_shift_label=dict(en="Auto pitch shift", zh="自动变调"), + pitch_shift_label=dict(en="Pitch shift (semitones)", zh="指定变调（半音）"), + control_type_label=dict(en="Control type", zh="控制类型"), + examples_label=dict(en="Reference examples (click to load)", zh="参考样例（点击加载）"), + example_choice_0=dict(en="—", zh="—"), + example_choice_1=dict(en="Example 1: Mandarin → Mandarin (melody), Start singing synthesis!", zh="样例 1: 普通话 → 普通话 (melody), 开始歌声合成吧!"), + example_choice_2=dict(en="Example 2: Mandarin → Cantonese (melody), Start singing synthesis!", zh="样例 2: 普通话 → 粤语 (melody), 开始歌声合成吧!"), + example_choice_3=dict(en="Example 3: Mandarin → English (melody), Start singing synthesis!", zh="样例 3: 普通话 → 英语 (melody), 开始歌声合成吧!"), + example_choice_4=dict(en="Example 4: Mandarin → Music (score), Start singing synthesis!", zh="样例 4: 普通话 → 音乐 (score), 开始歌声合成吧!"), + warn_missing_audio=dict( + en="Please upload both prompt audio and target audio", + zh="请上传 Prompt 音频与 Target 音频", + ), + # Instruction panel (workflow description) + instruction_title=dict(en="Usage", zh="使用说明"), + instruction_p1=dict( + en="After uploading prompt and target audio and clicking **Run singing transcription**, the system generates two metadata files (prompt and target).", + zh="上传 Prompt 与 Target 音频并点击「开始歌声转录」后，将生成 Prompt 与 Target 两份 metadata 文件。", + ), + instruction_p2=dict( + en="Auto-transcribed lyrics and notes are often misaligned. For better results, import the generated metadata into the **MIDI Editor** for manual adjustment: [SoulX-Singer-Midi-Editor](https://huggingface.co/spaces/Soul-AILab/SoulX-Singer-Midi-Editor).", + zh="自动转录的歌词与音高对齐效果通常不理想，建议将生成的 metadata 导入 **MIDI 编辑器** 进行手动调整：[SoulX-Singer-Midi-Editor](https://huggingface.co/spaces/Soul-AILab/SoulX-Singer-Midi-Editor)。", + ), + instruction_p3=dict( + en="Re-upload the adjusted metadata to the corresponding Prompt / Target Meta fields, then click **Run singing synthesis** to generate the final audio.", + zh="将调整后的 metadata 重新上传至对应的 Prompt / Target Meta 位置后，点击「歌声合成」开始最终生成。", + ), +) + +def _detect_initial_lang() -> Literal["zh", "en"]: + """Detect initial UI language from server locale (browser language applied later via JS).""" + try: + import locale + loc = (locale.getdefaultlocale()[0] or os.environ.get("LANG", "") or "").lower() + return "en" if loc.startswith("en") else "zh" + except Exception: + return "zh" + + +global_lang: Literal["zh", "en"] = _detect_initial_lang() + + +def i18n(key: str) -> str: + return _i18n_key2lang_dict[key][global_lang] + + +def get_lyric_lang_choices(): + """Lyric language dropdown (display, value) for current UI language.""" + return [ + (i18n("lyric_lang_mandarin"), "Mandarin"), + (i18n("lyric_lang_cantonese"), "Cantonese"), + (i18n("lyric_lang_english"), "English"), + ] + + +def _resolve_file_path(x): + """Gradio file input can be path string or (path, None) tuple.""" + if x is None: + return None + if isinstance(x, tuple): + x = x[0] + return x if (x and os.path.isfile(x)) else None + + +def transcription_function( + prompt_audio, + target_audio, + prompt_metadata, + target_metadata, + prompt_lyric_lang: str, + target_lyric_lang: str, + prompt_vocal_sep: bool, + target_vocal_sep: bool, +): + """Step 1: Run transcription only; output (prompt_meta_path, target_meta_path).""" + try: + if isinstance(prompt_audio, tuple): + prompt_audio = prompt_audio[0] + if isinstance(target_audio, tuple): + target_audio = target_audio[0] + if prompt_audio is None or target_audio is None: + gr.Warning(message=i18n("warn_missing_audio")) + return None, None + prompt_meta_resolved = _resolve_file_path(prompt_metadata) + target_meta_resolved = _resolve_file_path(target_metadata) + use_input_metadata = prompt_meta_resolved is not None and target_meta_resolved is not None + + session_base = _session_dir_from_target(target_audio) + audio_dir = session_base / "audio" + audio_dir.mkdir(parents=True, exist_ok=True) + transfer_prompt_path = audio_dir / "prompt.wav" + transfer_target_path = audio_dir / "target.wav" + SR = 44100 + PROMPT_MAX_SEC = 30 + TARGET_MAX_SEC = 60 + prompt_audio_data, _ = librosa.load(prompt_audio, sr=SR, mono=True) + target_audio_data, _ = librosa.load(target_audio, sr=SR, mono=True) + prompt_audio_data = prompt_audio_data[: PROMPT_MAX_SEC * SR] + target_audio_data = target_audio_data[: TARGET_MAX_SEC * SR] + sf.write(transfer_prompt_path, prompt_audio_data, SR) + sf.write(transfer_target_path, target_audio_data, SR) + + prompt_meta_path = session_base / "transcriptions" / "prompt" / "metadata.json" + target_meta_path = session_base / "transcriptions" / "target" / "metadata.json" + if use_input_metadata: + (session_base / "transcriptions" / "prompt").mkdir(parents=True, exist_ok=True) + (session_base / "transcriptions" / "target").mkdir(parents=True, exist_ok=True) + shutil.copy2(prompt_meta_resolved, prompt_meta_path) + shutil.copy2(target_meta_resolved, target_meta_path) + else: + ok, msg = APP_STATE.run_preprocess( + transfer_prompt_path, + transfer_target_path, + session_base, + prompt_vocal_sep=prompt_vocal_sep, + target_vocal_sep=target_vocal_sep, + prompt_lyric_lang=prompt_lyric_lang or "Mandarin", + target_lyric_lang=target_lyric_lang or "Mandarin", + ) + if not ok: + print(msg, file=sys.stderr, flush=True) + return None, None + + prompt_meta_file = str(prompt_meta_path) if prompt_meta_path.exists() else None + target_meta_file = str(target_meta_path) if target_meta_path.exists() else None + return prompt_meta_file, target_meta_file + except Exception: + print(traceback.format_exc(), file=sys.stderr, flush=True) + return None, None + + +def synthesis_function( + prompt_audio, + prompt_metadata, + target_metadata, + control: str, + auto_shift: bool, + pitch_shift, + seed: int, +): + """Step 2: Run SVS from top prompt_audio + prompt_metadata + target_metadata.""" + try: + if isinstance(prompt_audio, tuple): + prompt_audio = prompt_audio[0] + prompt_wav_path = prompt_audio + prompt_meta_path = _resolve_file_path(prompt_metadata) + target_meta_path = _resolve_file_path(target_metadata) + if not prompt_wav_path or not os.path.isfile(prompt_wav_path): + gr.Warning(message=i18n("warn_missing_synthesis")) + return None + if not prompt_meta_path or not os.path.isfile(prompt_meta_path): + gr.Warning(message=i18n("warn_missing_synthesis")) + return None + if not target_meta_path or not os.path.isfile(target_meta_path): + gr.Warning(message=i18n("warn_missing_synthesis")) + return None + if control not in ("melody", "score"): + control = "score" + seed = int(seed) + torch.manual_seed(seed) + np.random.seed(seed) + random.seed(seed) + ok, msg, merged = APP_STATE.run_svs_from_paths( + prompt_wav_path=prompt_wav_path, + prompt_metadata_path=prompt_meta_path, + target_metadata_path=target_meta_path, + control=control, + auto_shift=auto_shift, + pitch_shift=int(pitch_shift), + ) + if not ok or merged is None: + print(msg or "synthesis failed", file=sys.stderr, flush=True) + return None + return str(merged) + except Exception: + print(traceback.format_exc(), file=sys.stderr, flush=True) + return None + + +def _instruction_md() -> str: + """Markdown content for the instruction panel (supports links).""" + return "\n\n".join([ + f"**1.** {i18n('instruction_p1')}", + f"**2.** {i18n('instruction_p2')}", + f"**3.** {i18n('instruction_p3')}", + ]) + + +def render_interface() -> gr.Blocks: + with gr.Blocks(title="SoulX-Singer 歌声合成Demo", theme=gr.themes.Default()) as page: + gr.HTML( + '