Stabilize very short utterances

README.md

@@ -37,6 +37,7 @@ 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 |
 | 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 |
 | Endpoint cue | append terminal punctuation for model input when missing, except very short text |
 | Hard stop | native-pace target steps, independent of playback pace |
@@ -53,7 +54,8 @@ 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，
-並在預估進度 65% 後逐步降低 stop threshold。首段不會在無標點處硬切。進階參數可供研究比較，但 release
+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。
+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。
 
 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

@@ -20,6 +20,7 @@ from production import (
     StopHysteresisController,
     apply_loudness_floor,
     count_speech_units,
+    effective_generation_cfg,
     endpoint_generation_plan,
     estimate_step_seconds,
     extract_windowed_speaker_embedding,
@@ -56,6 +57,8 @@ TARGET_CPS = 4.0
 GENERATION_CPS = 5.2
 ASCII_GENERATION_CPS = 4.6
 MIN_ENDPOINT_CUE_UNITS = 6
+SHORT_TEXT_CFG_MIN = 3.0
+SHORT_TEXT_CFG_UNITS = 6
 CHUNK_CHARS = 80
 ONSET_CLAUSE_SEARCH_CHARS = 40
 MIN_CHUNK_CHARS = 12
@@ -188,11 +191,17 @@ def _generate_chunk(
     # 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
+    generation_cfg = effective_generation_cfg(
+        text,
+        cfg,
+        short_text_unit_threshold=SHORT_TEXT_CFG_UNITS,
+        short_text_min_cfg=SHORT_TEXT_CFG_MIN,
+    )
     set_generation_seed(request_seed)
     kwargs = {
         "target_text": model_text,
         "speaker_centroid": centroid,
-        "cfg_value": float(cfg),
+        "cfg_value": generation_cfg,
         "inference_timesteps": int(steps),
         "min_len": min_len,
         "max_len": hard_stop_steps,
@@ -222,7 +231,7 @@ def _generate_chunk(
             "[BlueMagpie] endpoint "
             f"expected_steps={expected_steps} hard_stop_steps={hard_stop_steps} "
             f"generated_steps={_STOP_CONTROLLER.last_generated_steps} "
-            f"reason={_STOP_CONTROLLER.last_stop_reason}"
+            f"reason={_STOP_CONTROLLER.last_stop_reason} cfg={generation_cfg:.2f}"
         )
     audio = audio.detach().float().cpu().numpy().reshape(-1)
     pace_speed = target_pace_speed(
@@ -367,7 +376,7 @@ 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`、
+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、
 只在自然標點切開首段、每 80 字切段，且不使用 retry 或 rerank。
 """

production.py

@@ -273,6 +273,22 @@ def count_speech_units(text: str) -> int:
     return units
 
 
+def effective_generation_cfg(
+    text: str,
+    requested_cfg: float,
+    *,
+    short_text_unit_threshold: int = 6,
+    short_text_min_cfg: float = 3.0,
+) -> float:
+    """Use stronger acoustic guidance only for empirically unstable short text."""
+
+    cfg = float(requested_cfg)
+    units = count_speech_units(text)
+    if 0 < units < max(1, int(short_text_unit_threshold)):
+        return max(cfg, float(short_text_min_cfg))
+    return cfg
+
+
 def estimate_step_seconds(model, sample_rate: int) -> float | None:
     patch_size = int(
         getattr(model, "patch_size", 0)

tests/test_production.py

@@ -8,6 +8,7 @@ from production import (
     StopHysteresisController,
     count_speech_units,
     duration_hard_stop_steps,
+    effective_generation_cfg,
     endpoint_generation_plan,
     ensure_terminal_punctuation,
     finish_audio,
@@ -166,6 +167,13 @@ def test_ascii_text_receives_a_wider_generation_window():
     assert select_generation_cps("AI TTS 測試", cjk_cps=5.2, ascii_cps=4.6) == 4.6
 
 
+def test_short_text_uses_stronger_cfg_without_changing_long_text():
+    assert effective_generation_cfg("你好", 2.0) == 3.0
+    assert effective_generation_cfg("測試完成", 2.5) == 3.0
+    assert effective_generation_cfg("這是一段完整測試", 2.0) == 2.0
+    assert effective_generation_cfg("你好", 3.5) == 3.5
+
+
 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