Fix inconsistent voice across TTS segments

Use the same manual_seed for every chunk and normalize per-segment peaks before concat so long voiceovers no longer sound like different speakers between segments.

Co-authored-by: Cursor <cursoragent@cursor.com>

config.py

@@ -142,6 +142,10 @@ TTS_TEMPERATURE = 0.3
 TTS_TOP_P = 0.7
 TTS_TOP_K = 20
 TTS_SPEED_PROMPT = "[speed_5]"
+# 多段拼接时各段必须使用同一随机种子，否则音色会像「换了个人」
+TTS_MANUAL_SEED = _env_int("TTS_MANUAL_SEED", 42)
+# 段间静音间隔（秒）
+TTS_SEGMENT_PAUSE_SEC = 0.35
 
 # 单段 TTS 最大字数（超长稿按句切分后逐段合成再拼接；8GB 显存建议 ≤200）
 TTS_MAX_CHARS_PER_CHUNK = _env_int("TTS_MAX_CHARS_PER_CHUNK", 200)

tts_service.py

@@ -33,9 +33,11 @@ from config import (
     SPEAKER_SAMPLE_MIN_SEC,
     TTS_MAX_CHARS_PER_CHUNK,
     TTS_ENABLE_CACHE,
+    TTS_MANUAL_SEED,
     TTS_MAX_NEW_TOKEN,
     TTS_MIN_NEW_TOKEN,
     TTS_SAMPLE_RATE,
+    TTS_SEGMENT_PAUSE_SEC,
     TTS_SPEED_PROMPT,
     TTS_TEMPERATURE,
     TTS_TOP_K,
@@ -615,10 +617,17 @@ def _run_chattts_infer(
     )
 
 
+def _normalize_segment_peak(wav: np.ndarray, target_peak: float = 0.95) -> np.ndarray:
+    """各段单独归一化峰值，避免拼接后某段偏响/偏轻像换了人声。"""
+    arr = np.asarray(wav, dtype=np.float32).flatten()
+    peak = float(np.max(np.abs(arr))) or 1.0
+    return arr / peak * target_peak
+
+
 def _concat_wavs(
     wavs: List[np.ndarray],
     sample_rate: int,
-    pause_sec: float = 0.35,
+    pause_sec: float = TTS_SEGMENT_PAUSE_SEC,
 ) -> np.ndarray:
     if not wavs:
         return np.array([], dtype=np.float32)
@@ -626,7 +635,7 @@ def _concat_wavs(
     pause = np.zeros(int(sample_rate * pause_sec), dtype=np.float32)
     segments: List[np.ndarray] = []
     for i, wav in enumerate(wavs):
-        segments.append(np.asarray(wav, dtype=np.float32).flatten())
+        segments.append(_normalize_segment_peak(wav))
         if i < len(wavs) - 1:
             segments.append(pause)
     return np.concatenate(segments)
@@ -694,18 +703,24 @@ def generate_voice(
         if speaker_warn:
             logger.warning(speaker_warn)
 
-        params_infer_code = ChatTTS.Chat.InferCodeParams(
+        chunk_temperature = (
-            prompt=TTS_SPEED_PROMPT,
+            min(TTS_TEMPERATURE, 0.2) if len(chunks) > 1 else TTS_TEMPERATURE
-            spk_emb=speaker_params.get("spk_emb"),
-            spk_smp=speaker_params.get("spk_smp"),
-            txt_smp=speaker_params.get("txt_smp"),
-            temperature=TTS_TEMPERATURE,
-            top_P=TTS_TOP_P,
-            top_K=TTS_TOP_K,
-            max_new_token=TTS_MAX_NEW_TOKEN,
-            min_new_token=TTS_MIN_NEW_TOKEN,
-            ensure_non_empty=False,
         )
+        infer_kwargs: Dict[str, Any] = {
+            "prompt": TTS_SPEED_PROMPT,
+            "spk_emb": speaker_params.get("spk_emb"),
+            "spk_smp": speaker_params.get("spk_smp"),
+            "txt_smp": speaker_params.get("txt_smp"),
+            "temperature": chunk_temperature,
+            "top_P": TTS_TOP_P,
+            "top_K": TTS_TOP_K,
+            "max_new_token": TTS_MAX_NEW_TOKEN,
+            "min_new_token": TTS_MIN_NEW_TOKEN,
+            "ensure_non_empty": False,
+        }
+        if "manual_seed" in inspect.signature(ChatTTS.Chat.InferCodeParams).parameters:
+            infer_kwargs["manual_seed"] = TTS_MANUAL_SEED
+        params_infer_code = ChatTTS.Chat.InferCodeParams(**infer_kwargs)
 
         params_refine_text = ChatTTS.Chat.RefineTextParams(
             prompt="[oral_2][laugh_0][break_4]",
@@ -725,7 +740,7 @@ def generate_voice(
             if not chunk or len(chunk) < 2:
                 continue
             release_cuda_cache()
-            chunk_infer = replace(params_infer_code, manual_seed=42 + idx)
+            chunk_infer = params_infer_code
             wavs = None
             last_exc: Optional[BaseException] = None
             for attempt in range(3):
@@ -736,8 +751,11 @@ def generate_voice(
                     break
                 except RecursionError as exc:
                     last_exc = exc
+                    # 重试时仍保持同一 manual_seed，避免段内/段间音色突变
+                    if "manual_seed" in infer_kwargs and attempt < 2:
                         chunk_infer = replace(
-                        chunk_infer, manual_seed=1000 + idx * 10 + attempt
+                            params_infer_code,
+                            manual_seed=TTS_MANUAL_SEED + attempt + 1,
                         )
                     release_cuda_cache()
                 except RuntimeError as exc:
@@ -755,8 +773,10 @@ def generate_voice(
                     chat, reload_err = get_chattts_instance()
                     if chat is None:
                         raise RuntimeError(reload_err or "ChatTTS 重载失败") from exc
+                    if "manual_seed" in infer_kwargs:
                         chunk_infer = replace(
-                        chunk_infer, manual_seed=2000 + idx * 10 + attempt
+                            params_infer_code,
+                            manual_seed=TTS_MANUAL_SEED + attempt + 1,
                         )
             if wavs is None:
                 return (