整理: 前後空白挿入の統合

test/test_synthesis_engine.py

@@ -17,6 +17,7 @@
     calc_frame_phoneme,
     calc_frame_pitch,
     mora_phoneme_list,
+    pad_with_silence,
     pre_process,
     split_mora,
     to_flatten_moras,
@@ -168,24 +169,43 @@ def _gen_mora(
     )
 
 
+def test_pad_with_silence():
+    """Test `pad_with_silence`."""
+    # Inputs
+    query = _gen_query(prePhonemeLength=2 * 0.01067, postPhonemeLength=6 * 0.01067)
+    moras = [
+        _gen_mora("ヒ", "h", 2 * 0.01067, "i", 4 * 0.01067, 100.0),
+    ]
+
+    # Expects
+    true_moras_with_silence = [
+        _gen_mora("　", None, None, "sil", 2 * 0.01067, 0.0),
+        _gen_mora("ヒ", "h", 2 * 0.01067, "i", 4 * 0.01067, 100.0),
+        _gen_mora("　", None, None, "sil", 6 * 0.01067, 0.0),
+    ]
+
+    # Outputs
+    moras_with_silence = pad_with_silence(moras, query)
+
+    assert moras_with_silence == true_moras_with_silence
+
+
 def test_calc_frame_per_phoneme():
     """Test `calc_frame_per_phoneme`."""
     # Inputs
-    query = _gen_query(
-        speedScale=2.0,
-        prePhonemeLength=2 * 0.01067,  # 0.01067 [sec/frame]
-        postPhonemeLength=6 * 0.01067,
-    )
+    query = _gen_query(speedScale=2.0)
     moras = [
+        _gen_mora("　", None, None, "　", 2 * 0.01067, 0.0),  # 0.01067 [sec/frame]
         _gen_mora("コ", "k", 2 * 0.01067, "o", 4 * 0.01067, 0.0),
         _gen_mora("ン", None, None, "N", 4 * 0.01067, 0.0),
         _gen_mora("、", None, None, "pau", 2 * 0.01067, 0.0),
         _gen_mora("ヒ", "h", 2 * 0.01067, "i", 4 * 0.01067, 0.0),
         _gen_mora("ホ", "h", 4 * 0.01067, "O", 2 * 0.01067, 0.0),
+        _gen_mora("　", None, None, "　", 6 * 0.01067, 0.0),
     ]
 
     # Expects
-    #                      Pre k  o  N pau h  i  h  O Pst
+    #                        Pre k  o  N pau h  i  h  O Pst
     true_frame_per_phoneme = [1, 1, 2, 2, 1, 1, 2, 2, 1, 3]
     true_frame_per_phoneme = numpy.array(true_frame_per_phoneme, dtype=numpy.int32)
 
@@ -200,15 +220,17 @@ def test_calc_frame_pitch():
     # Inputs
     query = _gen_query(pitchScale=2.0, intonationScale=0.5)
     moras = [
+        _gen_mora("　", None, None, "　", 0.0, 0.0),
         _gen_mora("コ", "k", 0.0, "o", 0.0, 50.0),
         _gen_mora("ン", None, None, "N", 0.0, 50.0),
         _gen_mora("、", None, None, "pau", 0.0, 0.0),
         _gen_mora("ヒ", "h", 0.0, "i", 0.0, 125.0),
         _gen_mora("ホ", "h", 0.0, "O", 0.0, 0.0),
+        _gen_mora("　", None, None, "　", 0.0, 0.0),
     ]
     phoneme_str = "pau k o N pau h i h O pau"
     phonemes = [OjtPhoneme(p, 0, 0) for p in phoneme_str.split()]
-    #              Pre k  o  N pau h  i  h  O Pst
+    #                   Pre k  o  N pau h  i  h  O Pst
     frame_per_phoneme = [1, 1, 2, 2, 1, 1, 2, 2, 1, 3]
     frame_per_phoneme = numpy.array(frame_per_phoneme, dtype=numpy.int32)
 
@@ -299,6 +321,7 @@ def test_feat_to_framescale():
     assert true_frame_per_phoneme.shape[0] == len(phoneme_data_list), "Prerequisites"
 
     # Outputs
+    flatten_moras = pad_with_silence(flatten_moras, query)
     frame_per_phoneme = calc_frame_per_phoneme(query, flatten_moras)
     f0 = calc_frame_pitch(query, flatten_moras, phoneme_data_list, frame_per_phoneme)
     frame_phoneme = calc_frame_phoneme(phoneme_data_list, frame_per_phoneme)

voicevox_engine/synthesis_engine/synthesis_engine.py

@@ -89,9 +89,9 @@ def pre_process(
     Returns
     -------
     flatten_moras : List[Mora]
-        AccentPhraseモデルのリスト内に含まれるすべてのMoraをリスト化したものを返す
+        モーラ列（前後の無音含まない）
     phoneme_data_list : List[OjtPhoneme]
-        flatten_morasから取り出したすべてのPhonemeをOjtPhonemeに変換したものを返す
+        音素列（前後の無音含む）
     """
     flatten_moras = to_flatten_moras(accent_phrases)
 
@@ -110,6 +110,30 @@ def pre_process(
     return flatten_moras, phoneme_data_list
 
 
+def generate_silence_mora(length: float) -> Mora:
+    """無音モーラの生成"""
+    return Mora(text="　", vowel="sil", vowel_length=length, pitch=0.0)
+
+
+def pad_with_silence(moras: list[Mora], query: AudioQuery) -> list[Mora]:
+    """モーラ列の先頭/最後尾へqueryに基づいた無音モーラを追加
+    Parameters
+    ----------
+    moras : List[Mora]
+        モーラ時系列
+    query : AudioQuery
+        音声合成クエリ
+    Returns
+    -------
+    moras : List[Mora]
+        前後無音が付加されたモーラ時系列
+    """
+    pre_silence_moras = [generate_silence_mora(query.prePhonemeLength)]
+    post_silence_moras = [generate_silence_mora(query.postPhonemeLength)]
+    moras = pre_silence_moras + moras + post_silence_moras
+    return moras
+
+
 def calc_frame_per_phoneme(query: AudioQuery, moras: List[Mora]):
     """
     音素あたりのフレーム長を算出
@@ -122,27 +146,25 @@ def calc_frame_per_phoneme(query: AudioQuery, moras: List[Mora]):
     Returns
     -------
     frame_per_phoneme : NDArray[]
-        音素（前後の無音含む）あたりのフレーム長。端数丸め。
+        音素あたりのフレーム長。端数丸め。
     """
-    # 音素（前後の無音含む）あたりの継続長
+    # 音素あたりの継続長
     sec_per_phoneme = numpy.array(
-        [query.prePhonemeLength]
-        + [
+        [
             length
             for mora in moras
             for length in (
                 [mora.consonant_length] if mora.consonant is not None else []
             )
             + [mora.vowel_length]
-        ]
-        + [query.postPhonemeLength],
+        ],
         dtype=numpy.float32,
     )
 
     # 話速による継続長の補正
     sec_per_phoneme /= query.speedScale
 
-    # 音素（前後の無音含む）あたりのフレーム長。端数丸め。
+    # 音素あたりのフレーム長。端数丸め。
     framerate = 24000 / 256  # framerate 93.75 [frame/sec]
     frame_per_phoneme = numpy.round(sec_per_phoneme * framerate).astype(numpy.int32)
 
@@ -166,15 +188,15 @@ def calc_frame_pitch(
     phonemes : List[OjtPhoneme]
         音素列
     frame_per_phoneme: NDArray
-        音素（前後の無音含む）あたりのフレーム長。端数丸め。
+        音素あたりのフレーム長。端数丸め。
     Returns
     -------
     frame_f0 : NDArray[]
         フレームごとの基本周波数系列
     """
     # TODO: Better function name (c.f. VOICEVOX/voicevox_engine#790)
-    # モーラ（前後の無音含む）ごとの基本周波数
-    f0 = numpy.array([0] + [mora.pitch for mora in moras] + [0], dtype=numpy.float32)
+    # モーラごとの基本周波数
+    f0 = numpy.array([mora.pitch for mora in moras], dtype=numpy.float32)
 
     # 音高スケールによる補正
     f0 *= 2**query.pitchScale
@@ -482,6 +504,7 @@ def _synthesis_impl(self, query: AudioQuery, style_id: int):
         # AccentPhraseをすべてMoraおよびOjtPhonemeの形に分解し、処理可能な形にする
         flatten_moras, phoneme_data_list = pre_process(query.accent_phrases)
 
+        flatten_moras = pad_with_silence(flatten_moras, query)
         frame_per_phoneme = calc_frame_per_phoneme(query, flatten_moras)
         f0 = calc_frame_pitch(
             query, flatten_moras, phoneme_data_list, frame_per_phoneme