Catch weak stop signals near sentence end

README.md

@@ -38,7 +38,7 @@ Demo 預設採用目前通過長文穩定性評估的推論設定：
 | Target pace | 4.0 speech units/sec |
 | Generation safety pace | 5.2 CJK / 4.6 ASCII-mixed units/sec + 1 latent step |
 | Short-text guidance | minimum CFG 3.0 below 6 speech units |
-| Stop policy | 0.50 → 0.20 from 65% to 90% predicted progress, 1 hit |
+| Stop policy | 0.50 → 0.05 from 75% to 95% predicted progress, 1 hit |
 | Endpoint cue | append terminal punctuation for model input when missing, except very short text |
 | Hard stop | native-pace target steps, independent of playback pace |
 | Pace correction | pitch-preserving stretch after completion; onset split only at punctuation |
@@ -54,7 +54,7 @@ Demo 預設採用目前通過長文穩定性評估的推論設定：
 
 自然 stop 在目前 checkpoint 上仍可能過快或錯過句尾。Demo 不再用播放目標語速決定生成長度；
 模型使用原生語速的安全上限，完成後才做保音高語速校正。缺少句末標點時只在模型輸入補上句號，
-少於 6 個 speech units 的短句會使用最低 CFG 3.0，並在預估進度 65% 後逐步降低 stop threshold。
+少於 6 個 speech units 的短句會使用最低 CFG 3.0；stop threshold 只在預估進度 75% 後逐步降低，並於 95% 進度降至 0.05，以捕捉句末弱 stop 訊號而不影響前段內容。
 首段不會在無標點處硬切。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。
 

app.py

@@ -66,9 +66,9 @@ CROSSFADE_MS = 80.0
 CHUNK_EDGE_FADE_MS = 80.0
 CHUNK_RMS_MATCH_DB = 4.0
 STOP_THRESHOLD = 0.50
-STOP_LATE_THRESHOLD = 0.20
-STOP_LATE_START_RATIO = 0.65
-STOP_LATE_FULL_RATIO = 0.90
+STOP_LATE_THRESHOLD = 0.05
+STOP_LATE_START_RATIO = 0.75
+STOP_LATE_FULL_RATIO = 0.95
 STOP_CONSECUTIVE = 1
 HARD_STOP_MARGIN_STEPS = 1
 MIN_PACE_SPEED = 0.80
@@ -377,7 +377,7 @@ 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`（極短句最低 `3.0`）、`NFE 10`、目標語速 `4.0 字/秒`、
-生成上限依模型原生語速計算、生成完成後校正至目標語速、補齊句末提示、單次高信心 stop、
+生成上限依模型原生語速計算、生成完成後校正至目標語速、補齊句末提示、尾端 weak-stop 保護、
 只在自然標點切開首段、每 80 字切段，且不使用 retry 或 rerank。
 """
 

tests/test_production.py

@@ -81,16 +81,30 @@ def test_stop_controller_relaxes_near_endpoint_and_enforces_hard_stop():
 
 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,
+        _SequenceStopHead([0.04] * 9 + [0.06]),
+        threshold=0.50,
+        late_threshold=0.05,
+        consecutive=1,
+        late_start_ratio=0.75,
+        late_full_ratio=0.95,
+    )
+    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(10)]
+    assert decisions == [0, 0, 0, 0, 0, 0, 0, 0, 0, 1]
+
+
+def test_stop_controller_does_not_accept_weak_signal_before_late_window():
+    controller = StopHysteresisController(
+        _SequenceStopHead([0.10] * 8),
+        threshold=0.50,
+        late_threshold=0.05,
+        consecutive=1,
+        late_start_ratio=0.75,
+        late_full_ratio=0.95,
     )
     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]
+    decisions = [controller(torch.zeros(1, 2)).argmax(dim=-1).item() for _ in range(8)]
+    assert decisions == [0, 0, 0, 0, 0, 0, 0, 0]
 
 
 def test_split_preserves_punctuation_and_minimum_chunk_size():