Stop earlier without clipping final speech

README.md

@@ -36,8 +36,8 @@ Demo 預設採用目前通過長文穩定性評估的推論設定：
 | CFG | 2.0 |
 | NFE steps | 10 |
 | Target pace | 4.0 speech units/sec |
-| Stop policy | 0.65 → 0.35 near endpoint, 2 consecutive hits |
-| Hard stop | predicted target steps |
+| Stop policy | 0.65 → 0.20 from 65% to 90% predicted progress, 2 consecutive hits |
+| Hard stop | predicted target steps + 1 step |
 | Pace correction | pitch-preserving stretch after natural completion; first chunk maximum 24 chars |
 | Maximum chunk | 80 chars |
 | Minimum chunk | 12 chars |
@@ -50,8 +50,9 @@ Demo 預設採用目前通過長文穩定性評估的推論設定：
 並依逗號、分號或句末標點插入不同長度的停頓。輸出最後會套用保守的 RMS floor 與 peak limit。
 
 自然 stop 在目前 checkpoint 上仍可能過快或錯過句尾。Demo 不再用最低生成長度強迫模型繼續
-發聲；模型自然完成文字後才做保音高語速校正，接近預期 endpoint 時降低 stop threshold，並以
-預估長度上界阻止多餘尾音。進階參數可供研究比較，但 release profile 是 CFG 2.0、NFE 10、後處理語速 1.0。
+發聲；模型自然完成文字後才做保音高語速校正，並在預估進度 65% 後逐步降低 stop threshold。
+最後保留一個 latent step 的收尾空間，再以硬上界阻止多餘尾音。進階參數可供研究比較，但 release
+profile 是 CFG 2.0、NFE 10、後處理語速 1.0。
 
 請只使用已取得授權的參考音檔。合成語音僅供研究與評估展示，正式使用前請人工檢視。
 

app.py

@@ -58,10 +58,12 @@ CROSSFADE_MS = 80.0
 CHUNK_EDGE_FADE_MS = 80.0
 CHUNK_RMS_MATCH_DB = 4.0
 STOP_THRESHOLD = 0.65
-STOP_LATE_THRESHOLD = 0.35
+STOP_LATE_THRESHOLD = 0.20
+STOP_LATE_START_RATIO = 0.65
+STOP_LATE_FULL_RATIO = 0.90
 STOP_CONSECUTIVE = 2
 HARD_STOP_RATIO = 1.0
-HARD_STOP_MARGIN_STEPS = 0
+HARD_STOP_MARGIN_STEPS = 1
 MIN_PACE_SPEED = 0.80
 MAX_TEXT_CHARS = 360
 
@@ -116,6 +118,8 @@ if not _NATIVE_STOP_POLICY:
         threshold=STOP_THRESHOLD,
         late_threshold=STOP_LATE_THRESHOLD,
         consecutive=STOP_CONSECUTIVE,
+        late_start_ratio=STOP_LATE_START_RATIO,
+        late_full_ratio=STOP_LATE_FULL_RATIO,
     )
     model.stop_head = _STOP_CONTROLLER
 
@@ -187,7 +191,7 @@ def _generate_chunk(
         "retry_badcase_ratio_threshold": 6.0,
     }
     if _NATIVE_STOP_POLICY:
-        kwargs["stop_threshold"] = STOP_LATE_THRESHOLD
+        kwargs["stop_threshold"] = STOP_THRESHOLD
         kwargs["stop_consecutive"] = STOP_CONSECUTIVE
     if "generation_seed" in _GENERATE_PARAMETERS:
         kwargs["generation_seed"] = request_seed
@@ -347,7 +351,7 @@ HEADER = f"""
 台灣華語與中英混合文字轉語音。模型版本：`{CHECKPOINT}`。
 
 目前預設採用穩定推論設定：`CFG 2.0`、`NFE 10`、目標語速 `4.0 字/秒`、
-生成完成後校正至目標語速、動態 stop hysteresis、target-length hard stop、每 80 字切段，且不使用 retry 或 rerank。
+生成完成後校正至目標語速、提前降低 stop threshold、target + 1 step hard stop、每 80 字切段，且不使用 retry 或 rerank。
 """
 
 

production.py

@@ -42,12 +42,19 @@ class StopHysteresisController(nn.Module):
         threshold: float = 0.65,
         late_threshold: float = 0.50,
         consecutive: int = 2,
+        late_start_ratio: float = 0.80,
+        late_full_ratio: float = 1.00,
     ) -> None:
         super().__init__()
         self.stop_head = stop_head
         self.threshold = float(threshold)
         self.late_threshold = float(late_threshold)
         self.consecutive = max(1, int(consecutive))
+        self.late_start_ratio = max(0.0, float(late_start_ratio))
+        self.late_full_ratio = max(
+            self.late_start_ratio + 1.0e-6,
+            float(late_full_ratio),
+        )
         self._active = False
         self._min_len = 2
         self._expected_steps = 0
@@ -82,8 +89,15 @@ class StopHysteresisController(nn.Module):
         threshold = self.threshold
         if self._expected_steps > 0:
             progress = generated_steps / float(self._expected_steps)
-            if progress >= 0.8:
-                blend = min(1.0, max(0.0, (progress - 0.8) / 0.2))
+            if progress >= self.late_start_ratio:
+                blend = min(
+                    1.0,
+                    max(
+                        0.0,
+                        (progress - self.late_start_ratio)
+                        / (self.late_full_ratio - self.late_start_ratio),
+                    ),
+                )
                 threshold = self.threshold + blend * (self.late_threshold - self.threshold)
         eligible = self._step > self._min_len
         if eligible and probability >= threshold:

tests/test_production.py

@@ -60,6 +60,20 @@ def test_stop_controller_relaxes_near_endpoint_and_enforces_hard_stop():
     assert decisions == [0, 0, 0, 1]
 
 
+def test_stop_controller_can_relax_before_the_text_pace_cap():
+    controller = StopHysteresisController(
+        _SequenceStopHead([0.4] * 9),
+        threshold=0.65,
+        late_threshold=0.20,
+        consecutive=2,
+        late_start_ratio=0.65,
+        late_full_ratio=0.90,
+    )
+    controller.begin(min_len=2, expected_steps=10, hard_stop_steps=11)
+    decisions = [controller(torch.zeros(1, 2)).argmax(dim=-1).item() for _ in range(9)]
+    assert decisions == [0, 0, 0, 0, 0, 0, 0, 0, 1]
+
+
 def test_split_preserves_punctuation_and_minimum_chunk_size():
     text = (
         "第一句很短。第二句也不長。第三句需要再多一些文字，才能測試切段是否正確。"
@@ -82,7 +96,7 @@ def test_duration_units_and_target_pace_min_len():
         stop_consecutive=2,
     ) == 5
     assert target_cps_steps("一二三四五六七八", target_cps=4.0, step_seconds=0.25) == 8
-    assert duration_hard_stop_steps(55, ratio=1.0, margin_steps=0) == 55
+    assert duration_hard_stop_steps(55, ratio=1.0, margin_steps=1) == 56
     assert duration_hard_stop_steps(0, ratio=1.08, margin_steps=3) == 2000