Stop forcing generated tail speech

README.md

@@ -37,8 +37,8 @@ Demo 預設採用目前通過長文穩定性評估的推論設定：
 | 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 + 1 step |
-| Onset pacing | first chunk at 0.90×, maximum 24 chars |
+| Hard stop | predicted target steps |
+| Pace correction | pitch-preserving stretch after natural completion; first chunk maximum 24 chars |
 | Maximum chunk | 80 chars |
 | Minimum chunk | 12 chars |
 | Crossfade / internal edge fade | 80 ms / 80 ms |
@@ -49,9 +49,9 @@ Demo 預設採用目前通過長文穩定性評估的推論設定：
 每個請求會取得新的隨機 seed；同一請求內的所有長文切段會重用該 seed。切段保留標點，
 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，因此 Demo 會依文字 speech units 設定
-最低生成長度，接近預期 endpoint 時降低 stop threshold，並以有限上界阻止多餘尾音。首段會做
-保音高的輕量緩速。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。
+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。
 
 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

@@ -30,7 +30,7 @@ from production import (
     punctuation_pause_seconds,
     set_generation_seed,
     split_text_for_tts,
-    target_cps_min_len,
+    target_pace_speed,
     target_cps_steps,
 )
 
@@ -61,8 +61,8 @@ STOP_THRESHOLD = 0.65
 STOP_LATE_THRESHOLD = 0.35
 STOP_CONSECUTIVE = 2
 HARD_STOP_RATIO = 1.0
-HARD_STOP_MARGIN_STEPS = 1
-ONSET_SPEED = 0.90
+HARD_STOP_MARGIN_STEPS = 0
+MIN_PACE_SPEED = 0.80
 MAX_TEXT_CHARS = 360
 
 
@@ -166,13 +166,9 @@ def _generate_chunk(
     request_seed: int,
 ) -> np.ndarray:
     expected_steps = target_cps_steps(text, TARGET_CPS, STEP_SECONDS)
-    min_len = target_cps_min_len(
-        text,
-        target_cps=TARGET_CPS,
-        step_seconds=STEP_SECONDS,
-        base_min_len=2,
-        stop_consecutive=STOP_CONSECUTIVE,
-    )
+    # Do not hold generation open to enforce pace. The model can finish the
+    # requested text early; extending its latent sequence creates tail speech.
+    min_len = 2
     hard_stop_steps = duration_hard_stop_steps(
         expected_steps,
         ratio=HARD_STOP_RATIO,
@@ -207,7 +203,15 @@ def _generate_chunk(
     finally:
         if _STOP_CONTROLLER is not None:
             _STOP_CONTROLLER.end()
-    return audio.detach().float().cpu().numpy().reshape(-1)
+    audio = audio.detach().float().cpu().numpy().reshape(-1)
+    pace_speed = target_pace_speed(
+        audio.size,
+        SR,
+        text,
+        target_cps=TARGET_CPS,
+        min_speed=MIN_PACE_SPEED,
+    )
+    return _apply_speed(audio, pace_speed)
 
 
 def _synthesize(
@@ -244,8 +248,6 @@ def _synthesize(
                 steps=steps,
                 request_seed=request_seed,
             )
-            if index == 0:
-                audio = _apply_speed(audio, ONSET_SPEED)
             if audio_chunks:
                 audio = match_chunk_rms(audio_chunks[0], audio, max_adjust_db=CHUNK_RMS_MATCH_DB)
             audio_chunks.append(audio)
@@ -345,7 +347,7 @@ 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`、
-首段 `0.90×` 緩速、動態 stop hysteresis、target + 1 step 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、動態 stop hysteresis、target-length hard stop、每 80 字切段，且不使用 retry 或 rerank。
 """
 
 

production.py

@@ -267,6 +267,26 @@ def target_cps_steps(text: str, target_cps: float, step_seconds: float | None) -
     return max(1, math.ceil((units / target_cps) / step_seconds))
 
 
+def target_pace_speed(
+    audio_samples: int,
+    sample_rate: int,
+    text: str,
+    *,
+    target_cps: float,
+    min_speed: float = 0.80,
+) -> float:
+    """Return a pitch-preserving stretch rate without extending generation."""
+
+    units = count_speech_units(text)
+    if audio_samples <= 0 or sample_rate <= 0 or units <= 0 or target_cps <= 0.0:
+        return 1.0
+    actual_seconds = float(audio_samples) / float(sample_rate)
+    target_seconds = float(units) / float(target_cps)
+    if actual_seconds >= target_seconds:
+        return 1.0
+    return min(1.0, max(float(min_speed), actual_seconds / target_seconds))
+
+
 def duration_hard_stop_steps(
     expected_steps: int,
     *,

tests/test_production.py

@@ -12,6 +12,7 @@ from production import (
     split_text_for_tts,
     target_cps_min_len,
     target_cps_steps,
+    target_pace_speed,
 )
 
 
@@ -81,10 +82,17 @@ 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=1) == 56
+    assert duration_hard_stop_steps(55, ratio=1.0, margin_steps=0) == 55
     assert duration_hard_stop_steps(0, ratio=1.08, margin_steps=3) == 2000
 
 
+def test_target_pace_speed_slows_completed_audio_without_extending_generation():
+    text = "一二三四五六七八"
+    assert target_pace_speed(1800, 1000, text, target_cps=4.0) == 0.9
+    assert target_pace_speed(1000, 1000, text, target_cps=4.0) == 0.8
+    assert target_pace_speed(2500, 1000, text, target_cps=4.0) == 1.0
+
+
 def test_finish_audio_fades_endpoint_and_appends_silence():
     output = finish_audio(
         np.ones(10, dtype=np.float32),