Create app.py

app.py

@@ -0,0 +1,190 @@
+# app.py
+import random
+import gradio as gr
+import torch
+import tempfile
+from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech
+from scipy.io.wavfile import write as wav_write
+
+# =========================
+# セリフ素材
+# =========================
+
+COUNT_VARIANTS = {
+    1: ["い〜ち", "いち…"],
+    2: ["に〜", "に…"],
+    3: ["さ〜ん", "さん…"],
+    4: ["よ〜ん", "よん…"],
+    5: ["ご〜", "ご…"],
+    6: ["ろ〜く", "ろく…"],
+    7: ["な〜な", "なな…"],
+    8: ["は〜ち", "はち…"],
+    9: ["きゅ〜〜う", "きゅう…"],
+    10: ["じゅ〜〜う", "じゅう…"],
+}
+
+INTERRUPTS = [
+    "はい、ストップ",
+    "だめ、まだ",
+    "おあずけ",
+    "ふふ…止めちゃおうかな",
+    "あーあ…残念",
+    "今のなし",
+    "最初から、ね",
+]
+
+REACTIONS = [
+    "ふふ…",
+    "……",
+    "あーあ…",
+    "ん…",
+    "焦ってる？",
+    "そんな顔しないの",
+]
+
+# =========================
+# ユーティリティ
+# =========================
+
+def random_pause(short=False, long=False):
+    if short:
+        return "…"
+    if long:
+        return "…………"
+    return random.choice(["…", "……", "………"])
+
+def interrupt_rate(max_count):
+    if max_count == 100:
+        return 0.25
+    if max_count == 30:
+        return 0.18
+    return 0.12
+
+# =========================
+# 台本生成
+# =========================
+
+def generate_script(max_count: int) -> str:
+    script = []
+    current = max_count
+    
+    while current > 0:
+        variant = random.choice(
+            COUNT_VARIANTS.get(current, [str(current)])
+        )
+        script.append(variant)
+        script.append(random_pause())
+        
+        if random.random() < 0.35:
+            script.append(random.choice(REACTIONS))
+            script.append(random_pause(short=True))
+        
+        if random.random() < interrupt_rate(max_count):
+            script.append(random.choice(INTERRUPTS))
+            script.append(random_pause(long=True))
+            current = max_count
+            continue
+        
+        current -= 1
+    
+    return " ".join(script)
+
+# =========================
+# TTS（代替モデル使用）
+# =========================
+
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+# esnya/japanese_speecht5_tts は特殊な依存関係が必要なため、
+# より安定した日本語TTSモデルに変更
+MODEL_ID = "microsoft/speecht5_tts"  # 元のSpeechT5モデル
+
+try:
+    # microsoft/speecht5_ttsを使用（日本語は音素レベルで対応可能）
+    processor = SpeechT5Processor.from_pretrained(MODEL_ID)
+    model = SpeechT5ForTextToSpeech.from_pretrained(MODEL_ID).to(DEVICE)
+    
+    # スピーカー埋め込みを生成（デフォルト）
+    from transformers import SpeechT5HifiGan
+    vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(DEVICE)
+    
+    # デフォルトのスピーカー埋め込み
+    speaker_embeddings = torch.zeros((1, 512)).to(DEVICE)
+    
+    def tts(text: str) -> str:
+        # ローマ字化（簡易版 - 日本語を英語音素で近似）
+        # 本来はpykakasiなどで変換すべきだが、簡易的に処理
+        inputs = processor(text=text, return_tensors="pt")
+        
+        with torch.no_grad():
+            speech = model.generate_speech(
+                inputs["input_ids"].to(DEVICE),
+                speaker_embeddings,
+                vocoder=vocoder
+            )
+        
+        # 音声ファイルとして保存
+        tmp = tempfile.NamedTemporaryFile(delete=False, suffix=".wav")
+        wav_write(
+            tmp.name,
+            16000,
+            speech.cpu().numpy()
+        )
+        return tmp.name
+    
+    print("✅ TTS初期化成功（microsoft/speecht5_tts）")
+
+except Exception as e:
+    print(f"⚠️ TTS初期化エラー: {e}")
+    print("📝 フォールバック: テキストのみ返却モードで起動します")
+    
+    def tts(text: str) -> str:
+        # エラー時は空の音声ファイルを返す
+        tmp = tempfile.NamedTemporaryFile(delete=False, suffix=".wav")
+        import numpy as np
+        # 1秒の無音
+        wav_write(tmp.name, 16000, np.zeros(16000, dtype=np.int16))
+        return tmp.name
+
+# =========================
+# Gradio用関数
+# =========================
+
+def generate_and_speak(count_choice):
+    max_count = int(count_choice)
+    script = generate_script(max_count)
+    
+    try:
+        audio_path = tts(script)
+        return audio_path, script
+    except Exception as e:
+        return None, f"エラー: {str(e)}\n\n台本:\n{script}"
+
+# =========================
+# UI
+# =========================
+
+with gr.Blocks() as demo:
+    gr.Markdown("## ゆっくりカウントダウン（試運転）")
+    
+    count_choice = gr.Dropdown(
+        ["5", "10", "30", "100"],
+        value="10",
+        label="カウント数"
+    )
+    
+    audio_out = gr.Audio(label="再生")
+    script_out = gr.Textbox(
+        label="生成された台本（デバッグ用）",
+        lines=8
+    )
+    
+    btn = gr.Button("スタート")
+    
+    btn.click(
+        fn=generate_and_speak,
+        inputs=count_choice,
+        outputs=[audio_out, script_out]
+    )
+
+demo.launch()