Improve recorded reference voice similarity

README.md

@@ -21,7 +21,7 @@ models:
 ## 模式
 
 - **內建語者**：預設為內建語者 A；Speaker B 使用模型庫 `checkpoints/speaker_b_embedding.pt`，固定 seed 56789 與 projector scale 1.125。
-- **參考音色**：30 秒內整段提取一次 speaker embedding，固定 centroid/projector scale 1.25 與 seed 1337。
+- **參考音色**：以 3 秒窗、1.5 秒 hop（最多 4 個窗）平均成單一 speaker embedding，固定 centroid/projector scale 1.25 與 seed 1337；每個文字分段仍只生成一次。
 - **長文**：以最多 80 個字元快速分段，各分段單次生成後加入短靜音串接。
 
 TTS model snapshot 固定為

app.py

@@ -250,10 +250,13 @@ QUALITY_FAST_EXACT_MAX_BOUNDARY_SPEAKER_DROP = 0.08
 QUALITY_MAX_SEQUENCE_PATHS = 3
 SHORT_AUDIO_SPEAKER_GATE_SECONDS = 1.50
 REFERENCE_MIN_DURATION_SECONDS = 3.0
-REFERENCE_WINDOW_SECONDS = 30.0
-REFERENCE_HOP_SECONDS = 30.0
-REFERENCE_MAX_WINDOWS = 1
-REFERENCE_FULL_CLIP_MAX_SECONDS = 30.0
+# A short overlapping-window centroid is measurably closer to a recorded
+# reference than one whole-clip embedding.  Keep the cap small so extraction
+# remains bounded and generation itself is still exactly single-pass.
+REFERENCE_WINDOW_SECONDS = 3.0
+REFERENCE_HOP_SECONDS = 1.5
+REFERENCE_MAX_WINDOWS = 4
+REFERENCE_FULL_CLIP_MAX_SECONDS = 0.0
 SPEAKER_GENERATION_SEED = 56_789
 SPEAKER_B_LABEL = "內建語者 B"
 SPEAKER_B_PROJECTOR_SCALE = 1.125
@@ -3944,7 +3947,10 @@ with gr.Blocks(title="BlueMagpie-TTS Demo", theme=gr.themes.Soft()) as demo:
         )
 
     with gr.Tab("參考音色"):
-        gr.Markdown("請使用至少 3 秒、單一語者且背景乾淨的授權參考音檔。")
+        gr.Markdown(
+            "建議錄製 6–12 秒、單一語者且背景乾淨的授權參考音檔；"
+            "至少需要 3 秒。"
+        )
         with gr.Row():
             with gr.Column():
                 reference_text = gr.Textbox(label="文字", lines=4, max_lines=8)

tests/test_latency_first_mode.py

@@ -207,10 +207,10 @@ def test_reference_conditioning_uses_one_fixed_embedding():
     source = APP_PATH.read_text(encoding="utf-8")
     reference_source = ast.get_source_segment(source, _function_node("tts_reference"))
 
-    assert "REFERENCE_WINDOW_SECONDS = 30.0" in source
-    assert "REFERENCE_HOP_SECONDS = 30.0" in source
-    assert "REFERENCE_MAX_WINDOWS = 1" in source
-    assert "REFERENCE_FULL_CLIP_MAX_SECONDS = 30.0" in source
+    assert "REFERENCE_WINDOW_SECONDS = 3.0" in source
+    assert "REFERENCE_HOP_SECONDS = 1.5" in source
+    assert "REFERENCE_MAX_WINDOWS = 4" in source
+    assert "REFERENCE_FULL_CLIP_MAX_SECONDS = 0.0" in source
     assert "REFERENCE_CONDITIONING_SCALE = 1.25" in source
     assert "REFERENCE_GENERATION_SEED = 1_337" in source
     assert "REFERENCE_PROJECTOR_SCALE = 1.25" in source