更新 3, 2, 1 的声音

tools/gen_voice.py

@@ -112,7 +112,69 @@ def to_mp3(wav_path: Path, mp3_path: Path) -> None:
     )
 
 
+def normalize_clip(wav: np.ndarray, sr: int) -> np.ndarray:
+    wav = trim_edges(np.asarray(wav, dtype=np.float32).reshape(-1), sr)
+    peak = float(np.max(np.abs(wav)))
+    if peak < 0.05:
+        raise RuntimeError(f"peak too low ({peak:.4f})")
+    wav = wav / peak * PEAK_NORM
+    dur = len(wav) / sr
+    if dur < 0.35:
+        wav = np.concatenate([wav, np.zeros(int((0.35 - dur) * sr), dtype=np.float32)])
+    return wav
+
+
+def write_mp3(name: str, wav: np.ndarray, sr: int) -> None:
+    wav_path = STAGING / f"{name}.wav"
+    mp3_path = OUT_DIR / f"{name}.mp3"
+    sf.write(str(wav_path), wav, sr)
+    to_mp3(wav_path, mp3_path)
+    dur = len(wav) / sr
+    print(f"[ok] {name}.mp3  {dur:.2f}s  {mp3_path.stat().st_size} bytes", flush=True)
+
+
+def gen_countdown(model) -> tuple[np.ndarray, int]:
+    """倒计时单独按字生成再拼接，避免「三、二、一」整句被读成播报腔、与提示音气质漂移。
+
+    与提示音共用同一 speaker + instruct；句间留约 0.35s，总长约 2.5s。
+    """
+    digits = ["三", "二", "一"]
+    # 同批带上一条短提示作音色锚点，丢弃锚点音频
+    texts = ["保持不动", *digits]
+    wavs, sr = model.generate_custom_voice(
+        text=texts,
+        language=["Chinese"] * len(texts),
+        speaker=[SPEAKER] * len(texts),
+        instruct=[INSTRUCT] * len(texts),
+        max_new_tokens=2048,
+    )
+    gap = np.zeros(int(0.35 * sr), dtype=np.float32)
+    parts: list[np.ndarray] = []
+    for i, digit in enumerate(digits):
+        clip = normalize_clip(wavs[i + 1], sr)
+        parts.append(clip)
+        if i < len(digits) - 1:
+            parts.append(gap)
+        print(f"  countdown digit {digit}: {len(clip) / sr:.2f}s", flush=True)
+    return np.concatenate(parts), sr
+
+
 def main() -> int:
+    import argparse
+
+    ap = argparse.ArgumentParser()
+    ap.add_argument(
+        "--only",
+        default="",
+        help="逗号分隔的 id，只生成这些；空则全量",
+    )
+    args = ap.parse_args()
+    only = {x.strip() for x in args.only.split(",") if x.strip()}
+    todo = [(n, t) for n, t in LINES if not only or n in only]
+    if only and not todo:
+        print(f"unknown --only ids: {sorted(only)}", flush=True)
+        return 1
+
     OUT_DIR.mkdir(parents=True, exist_ok=True)
     STAGING.mkdir(parents=True, exist_ok=True)
 
@@ -123,46 +185,46 @@ def main() -> int:
         dtype=torch.bfloat16,
     )
 
-    names = [n for n, _ in LINES]
-    texts = [t for _, t in LINES]
-    print(f"generating {len(names)} lines with speaker={SPEAKER} ...", flush=True)
-    wavs, sr = model.generate_custom_voice(
-        text=texts,
-        language=["Chinese"] * len(names),
-        speaker=[SPEAKER] * len(names),
-        instruct=[INSTRUCT] * len(names),
-        max_new_tokens=2048,
-    )
+    # 倒计时走专用拼接路径，其余整句生成
+    normal = [(n, t) for n, t in todo if n != "countdown-321"]
+    do_countdown = any(n == "countdown-321" for n, _ in todo)
 
-    manifest_ids: list[str] = []
-    for name, wav in zip(names, wavs):
-        wav = np.asarray(wav, dtype=np.float32).reshape(-1)
-        wav = trim_edges(wav, sr)
+    if normal:
+        names = [n for n, _ in normal]
+        texts = [t for _, t in normal]
+        print(f"generating {len(names)} lines with speaker={SPEAKER} ...", flush=True)
+        wavs, sr = model.generate_custom_voice(
+            text=texts,
+            language=["Chinese"] * len(names),
+            speaker=[SPEAKER] * len(names),
+            instruct=[INSTRUCT] * len(names),
+            max_new_tokens=2048,
+        )
+        for name, wav in zip(names, wavs):
+            try:
+                write_mp3(name, normalize_clip(wav, sr), sr)
+            except RuntimeError as e:
+                print(f"[FAIL] {name}: {e}", flush=True)
+                return 1
+
+    if do_countdown:
+        print(f"generating countdown-321 (digit stitch, speaker={SPEAKER}) ...", flush=True)
+        wav, sr = gen_countdown(model)
+        # 整段再峰值归一一次，保证与提示音响度一致
         peak = float(np.max(np.abs(wav)))
-        if peak < 0.05:
-            print(f"[FAIL] {name}: peak too low ({peak:.4f})", flush=True)
-            return 1
         wav = wav / peak * PEAK_NORM
-        dur = len(wav) / sr
-        if dur < 0.35:
-            pad = np.zeros(int((0.35 - dur) * sr), dtype=np.float32)
-            wav = np.concatenate([wav, pad])
-            dur = len(wav) / sr
+        write_mp3("countdown-321", wav, sr)
 
-        wav_path = STAGING / f"{name}.wav"
-        mp3_path = OUT_DIR / f"{name}.mp3"
-        sf.write(str(wav_path), wav, sr)
-        to_mp3(wav_path, mp3_path)
-        size = mp3_path.stat().st_size
-        print(f"[ok] {name}.mp3  {dur:.2f}s  {size} bytes", flush=True)
-        manifest_ids.append(name)
-
-    manifest = {"files": manifest_ids}
+    # 清单始终覆盖文档全量已落地文件（缺文件则跳过写入名）
+    existing = [
+        n for n, _ in LINES if (OUT_DIR / f"{n}.mp3").is_file()
+    ]
+    manifest = {"files": existing}
     (OUT_DIR / "manifest.json").write_text(
         json.dumps(manifest, ensure_ascii=False, indent=2) + "\n",
         encoding="utf-8",
     )
-    print(f"wrote {OUT_DIR / 'manifest.json'} ({len(manifest_ids)} files)", flush=True)
+    print(f"wrote {OUT_DIR / 'manifest.json'} ({len(existing)} files)", flush=True)
     return 0