import gradio as gr import torch import torchaudio import os import tempfile import spaces from datetime import datetime from omnivoice import OmniVoice # ─── Model ─── print("モデルを読み込み中...") device = "cuda" if torch.cuda.is_available() else "cpu" dtype = torch.float16 if device == "cuda" else torch.float32 model = OmniVoice.from_pretrained("drbaph/OmniVoice-bf16", device_map=device, dtype=dtype) print(f"モデル読み込み完了（{device}）") def _build_instruct(gender, age, pitch, style): parts = [] if gender and gender != "Auto": parts.append(gender.lower()) if age and age != "Auto": parts.append(age.lower()) if pitch and pitch != "Auto": parts.append(f"{pitch.lower()} pitch") if style and style != "Auto": parts.append(style.lower()) return ", ".join(parts) if parts else None # ─── Voice Design / Auto ─── @spaces.GPU def generate_design(text, mode, language, gender, age, pitch, style, speed, duration, num_step, guidance_scale, denoise, postprocess): if not text or not text.strip(): return None, "テキストを入力してください。" kwargs = dict(num_step=int(num_step), guidance_scale=guidance_scale, denoise=denoise) if language and language != "Auto": kwargs["language"] = language if mode == "Voice Design": instruct = _build_instruct(gender, age, pitch, style) if instruct: kwargs["instruct"] = instruct if duration and duration > 0: kwargs["duration"] = duration else: kwargs["speed"] = speed if postprocess: kwargs["postprocess_output"] = True try: audio = model.generate(text=text, **kwargs) with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f: torchaudio.save(f.name, audio[0], 24000) return f.name, f"生成完了（{audio[0].shape[1]/24000:.1f}秒）" except Exception as e: return None, f"エラー: {e}" # ─── Voice Clone ─── @spaces.GPU def generate_clone(text, ref_audio, ref_text, language, speed, duration, num_step, guidance_scale, denoise, postprocess): if not text or not text.strip(): return None, "テキストを入力してください。" if ref_audio is None: return None, "リファレンス音声をアップロードしてください。" kwargs = dict(num_step=int(num_step), guidance_scale=guidance_scale, denoise=denoise) if language and language != "Auto": kwargs["language"] = language if duration and duration > 0: kwargs["duration"] = duration else: kwargs["speed"] = speed if postprocess: kwargs["postprocess_output"] = True try: audio = model.generate( text=text, ref_audio=ref_audio, ref_text=ref_text if ref_text and ref_text.strip() else None, **kwargs, ) with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f: torchaudio.save(f.name, audio[0], 24000) return f.name, f"生成完了（{audio[0].shape[1]/24000:.1f}秒）" except Exception as e: return None, f"エラー: {e}" # ─── UI ─── CSS = """ .main-title { text-align: center; font-size: 1.8em; font-weight: 800; margin-bottom: 0; } .subtitle { text-align: center; color: #888; font-size: 0.9em; margin-bottom: 1em; } footer { display: none !important; } """ with gr.Blocks(title="OmniVoice") as app: gr.HTML("

OmniVoice

") gr.HTML("

AI Voice Generator — Personal

") with gr.Tabs(): # ── Voice Design / Auto ── with gr.Tab("Voice Design"): with gr.Row(): with gr.Column(scale=1): d_text = gr.Textbox(label="読み上げテキスト", lines=4, placeholder="テキストを入力...") d_mode = gr.Radio(["Auto", "Voice Design"], value="Auto", label="モード") d_lang = gr.Dropdown(["Auto", "Japanese", "English", "Korean"], value="Auto", label="言語") with gr.Group(visible=False) as d_voice_opts: with gr.Row(): d_gender = gr.Dropdown(["Auto", "Female", "Male"], value="Auto", label="性別") d_age = gr.Dropdown(["Auto", "Child", "Young", "Middle-aged", "Elderly"], value="Auto", label="年齢") with gr.Row(): d_pitch = gr.Dropdown( ["Auto", "Very low", "Low", "Moderate", "High", "Very high"], value="Auto", label="ピッチ") d_style = gr.Dropdown(["Auto", "Whisper"], value="Auto", label="スタイル") d_speed = gr.Slider(0.5, 2.0, value=1.0, step=0.1, label="速度") with gr.Accordion("詳細設定", open=False): d_duration = gr.Number(value=0, label="Duration（秒）", info="0で自動。設定するとSpeedは無視") d_steps = gr.Slider(4, 64, value=32, step=1, label="Inference Steps") d_cfg = gr.Slider(0.5, 5.0, value=2.0, step=0.1, label="Guidance Scale") d_denoise = gr.Checkbox(value=True, label="Denoise") d_postprocess = gr.Checkbox(value=True, label="Postprocess（無音除去）") d_btn = gr.Button("音声を生成", variant="primary", size="lg") with gr.Column(scale=1): d_audio = gr.Audio(label="生成結果", type="filepath") d_status = gr.Textbox(label="ステータス", interactive=False) d_mode.change( fn=lambda m: gr.update(visible=m == "Voice Design"), inputs=d_mode, outputs=d_voice_opts, ) d_btn.click( fn=generate_design, inputs=[d_text, d_mode, d_lang, d_gender, d_age, d_pitch, d_style, d_speed, d_duration, d_steps, d_cfg, d_denoise, d_postprocess], outputs=[d_audio, d_status], ) # ── Voice Clone ── with gr.Tab("Voice Clone"): with gr.Row(): with gr.Column(scale=1): c_text = gr.Textbox(label="読み上げテキスト", lines=4, placeholder="この声で読み上げたいテキスト...") c_ref = gr.Audio(label="リファレンス音声（3〜15秒）", type="filepath") c_ref_text = gr.Textbox(label="書き起こし（任意）", lines=2, placeholder="省略すると自動書き起こし") c_lang = gr.Dropdown(["Auto", "Japanese", "English", "Korean"], value="Auto", label="言語") c_speed = gr.Slider(0.5, 2.0, value=1.0, step=0.1, label="速度") with gr.Accordion("詳細設定", open=False): c_duration = gr.Number(value=0, label="Duration（秒）") c_steps = gr.Slider(4, 64, value=32, step=1, label="Inference Steps") c_cfg = gr.Slider(0.5, 5.0, value=2.0, step=0.1, label="Guidance Scale") c_denoise = gr.Checkbox(value=True, label="Denoise") c_postprocess = gr.Checkbox(value=True, label="Postprocess（無音除去）") c_btn = gr.Button("音声を生成", variant="primary", size="lg") with gr.Column(scale=1): c_audio = gr.Audio(label="生成結果", type="filepath") c_status = gr.Textbox(label="ステータス", interactive=False) c_btn.click( fn=generate_clone, inputs=[c_text, c_ref, c_ref_text, c_lang, c_speed, c_duration, c_steps, c_cfg, c_denoise, c_postprocess], outputs=[c_audio, c_status], ) if __name__ == "__main__": app.launch()