修正: 話速によって前後の無音時間が変わらないようにする

test/unit/tts_pipeline/test_wave_synthesizer.py

@@ -258,23 +258,26 @@ def test_query_to_decoder_feature() -> None:
     # Expects
     # frame_per_phoneme
     #                        Pre k  o  N pau h  i  h  O Pst
-    true_frame_per_phoneme = [1, 1, 2, 2, 2, 1, 2, 2, 1, 3]
+    true_frame_per_phoneme = [2, 1, 2, 2, 2, 1, 2, 2, 1, 6]
     n_frame = sum(true_frame_per_phoneme)
     # phoneme
-    #                     Pr  k   o   o  N  N pau pau h   i   i   h   h  O Pt Pt Pt
-    frame_phoneme_idxs = [0, 23, 30, 30, 4, 4, 0, 0, 19, 21, 21, 19, 19, 5, 0, 0, 0]
+    #                     Pr Pr  k   o   o  N  N pau pau]
+    frame_phoneme_idxs = [0, 0, 23, 30, 30, 4, 4, 0, 0]
+    #                       h   i   i   h   h  O Pt Pt Pt Pt Pt Pt
+    frame_phoneme_idxs += [19, 21, 21, 19, 19, 5, 0, 0, 0, 0, 0, 0]
+
     true_phoneme = np.zeros([n_frame, TRUE_NUM_PHONEME], dtype=np.float32)
     for frame_idx, phoneme_idx in enumerate(frame_phoneme_idxs):
         true_phoneme[frame_idx, phoneme_idx] = 1.0
     # Pitch
     #                   paw ko  N pau hi hO paw
-    # frame_per_vowel = [1, 3,  2, 1, 3, 3, 3]
-    #           pau   ko    ko    ko     N     N
-    true1_f0 = [0.0, 22.0, 22.0, 22.0, 22.0, 22.0]
-    #           pau  pau  hi    hi    hi
+    # frame_per_vowel = [2, 3,  2, 2, 3, 3, 6]
+    #           pau  pau   ko    ko    ko    N     N
+    true1_f0 = [0.0, 0.0, 22.0, 22.0, 22.0, 22.0, 22.0]
+    #           pau  pau   hi    hi    hi
     true2_f0 = [0.0, 0.0, 28.0, 28.0, 28.0]
-    #           hO   hO   hO   paw  paw  paw
-    true3_f0 = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
+    #           hO   hO   hO   paw  paw  paw  paw  paw  paw
+    true3_f0 = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
     true_f0 = np.array(true1_f0 + true2_f0 + true3_f0, dtype=np.float32)
 
     # Outputs

voicevox_engine/tts_pipeline/tts_engine.py

@@ -228,10 +228,10 @@ def _query_to_decoder_feature(
     moras = to_flatten_moras(query.accent_phrases)
 
     # 設定を適用する
-    moras = _apply_prepost_silence(moras, query)
     moras = _apply_pause_length(moras, query)
     moras = _apply_pause_length_scale(moras, query)
     moras = _apply_speed_scale(moras, query)
+    moras = _apply_prepost_silence(moras, query)
     moras = _apply_pitch_scale(moras, query)
     moras = _apply_intonation_scale(moras, query)