追加: OJT 音素のバリデーション

test/tts_pipeline/test_text_analyzer.py

@@ -1,6 +1,13 @@
 from unittest import TestCase
 
-from voicevox_engine.model import AccentPhrase, Mora
+import pytest
+
+from voicevox_engine.model import (
+    AccentPhrase,
+    Mora,
+    NonOjtPhonemeError,
+    OjtUnknownPhonemeError,
+)
 from voicevox_engine.tts_pipeline.text_analyzer import (
     AccentPhraseLabel,
     BreathGroupLabel,
@@ -402,22 +409,24 @@ def stub_unknown_features_koxx(_: str) -> list[str]:
     ]
 
 
+def test_label_non_ojt_phoneme():
+    """`Label` は OJT で想定されない音素をパース失敗する"""
+    non_ojt_phoneme = "G"
+    non_ojt_feature = f".^.-{non_ojt_phoneme}+.=./A:.+2+./B:.-._./C:._.+./D:.+._./E:._.!._.-./F:2_1#0_.@1_.|._./G:._.%._._./H:._./I:.-.@1+.&.-.|.+./J:._./K:.+.-."  # noqa: B950
+    with pytest.raises(NonOjtPhonemeError):
+        unknown_label = Label.from_feature(non_ojt_feature)
+        unknown_label.phoneme
+
+
+def test_label_unknown_phoneme():
+    """`Label` は unknown 音素 `xx` をパース失敗する"""
+    unknown_feature = stub_unknown_features_koxx("dummy")[3]
+    with pytest.raises(OjtUnknownPhonemeError):
+        unknown_label = Label.from_feature(unknown_feature)
+        unknown_label.phoneme
+
+
 def test_text_to_accent_phrases_unknown():
-    """`text_to_accent_phrases` は unknown 音素を含む features をパースする"""
-    # Expects
-    true_accent_phrases = [
-        AccentPhrase(
-            moras=[
-                _gen_mora("コ", "k", "o"),
-                _gen_mora("xx", None, "xx"),
-            ],
-            accent=1,
-            pause_mora=None,
-        ),
-    ]
-    # Outputs
-    accent_phrases = text_to_accent_phrases(
-        "dummy", text_to_features=stub_unknown_features_koxx
-    )
-    # Tests
-    assert accent_phrases == true_accent_phrases
+    """`text_to_accent_phrases` は unknown 音素を含む features をパース失敗する"""
+    with pytest.raises(OjtUnknownPhonemeError):
+        text_to_accent_phrases("dummy", text_to_features=stub_unknown_features_koxx)

test/tts_pipeline/test_tts_engine.py

@@ -3,7 +3,6 @@
 from unittest.mock import Mock
 
 import numpy as np
-import pytest
 from numpy.typing import NDArray
 from syrupy.assertion import SnapshotAssertion
 
@@ -17,16 +16,13 @@
     Note,
     Score,
 )
-from voicevox_engine.tts_pipeline.text_analyzer import text_to_accent_phrases
 from voicevox_engine.tts_pipeline.tts_engine import (
     TTSEngine,
     apply_interrogative_upspeak,
     to_flatten_moras,
     to_flatten_phonemes,
 )
 
-from .test_text_analyzer import stub_unknown_features_koxx
-
 
 def yukarin_s_mock(
     length: int, phoneme_list: NDArray[np.int64], style_id: NDArray[np.int64]
@@ -296,19 +292,6 @@ def test_update_pitch(self):
         np.testing.assert_array_equal(end_accent_phrase_list, true_phrase_ends)
 
 
-def test_create_accent_phrases_toward_unknown():
-    """`TTSEngine.create_accent_phrases()` は unknown 音素の Phoneme 化に失敗する"""
-    engine = TTSEngine(MockCoreWrapper())
-
-    # NOTE: TTSEngine.create_accent_phrases() のコールで unknown feature を得ることが難しいため、疑似再現
-    accent_phrases = text_to_accent_phrases(
-        "dummy", text_to_features=stub_unknown_features_koxx
-    )
-    with pytest.raises(ValueError) as e:
-        accent_phrases = engine.update_length_and_pitch(accent_phrases, StyleId(0))
-    assert str(e.value) == "tuple.index(x): x not in tuple"
-
-
 def test_mocked_update_length_output(snapshot_json: SnapshotAssertion) -> None:
     """モックされた `TTSEngine.update_length()` の出力スナップショットが一定である"""
     # Inputs

voicevox_engine/model.py

@@ -149,6 +149,16 @@ def __init__(self, err: ParseKanaError):
         super().__init__(text=err.text, error_name=err.errname, error_args=err.kwargs)
 
 
+class NonOjtPhonemeError(Exception):
+    def __init__(self, **kwargs: Any) -> None:
+        self.text = "OpenJTalk で想定されていない音素が生成されたため処理できません。"
+
+
+class OjtUnknownPhonemeError(Exception):
+    def __init__(self, **kwargs: Any) -> None:
+        self.text = "OpenJTalk の unknown 音素 `xx` は非対応です。"
+
+
 class MorphableTargetInfo(BaseModel):
     is_morphable: bool = Field(title="指定した話者に対してモーフィングの可否")
     # FIXME: add reason property

voicevox_engine/tts_pipeline/text_analyzer.py

@@ -5,13 +5,28 @@
 
 import pyopenjtalk
 
-from ..model import AccentPhrase, Mora
+from ..model import AccentPhrase, Mora, NonOjtPhonemeError, OjtUnknownPhonemeError
 from .mora_mapping import mora_phonemes_to_mora_kana
-
-OjtVowel = Literal[
-    "A", "E", "I", "N", "O", "U", "a", "cl", "e", "i", "o", "pau", "sil", "u"
-]
-OjtConsonant = Literal[
+from .phoneme import Consonant, Vowel
+
+# OpenJTalk が出力する音素の一覧。
+_OJT_VOWELS = (
+    "A",
+    "E",
+    "I",
+    "N",
+    "O",
+    "U",
+    "a",
+    "cl",
+    "e",
+    "i",
+    "o",
+    "pau",
+    "sil",
+    "u",
+)
+_OJT_CONSONANTS = (
     "b",
     "by",
     "ch",
@@ -44,9 +59,9 @@
     "w",
     "y",
     "z",
-]
-OjtUnknown = Literal["xx"]
-OjtPhoneme = OjtVowel | OjtConsonant | OjtUnknown
+)
+_OJT_UNKNOWNS = ("xx",)
+_OJT_PHONEMES = _OJT_VOWELS + _OJT_CONSONANTS + _OJT_UNKNOWNS
 
 
 @dataclass
@@ -82,10 +97,16 @@ def from_feature(cls, feature: str) -> Self:
         return cls(contexts=contexts)
 
     @property
-    def phoneme(self) -> OjtPhoneme:
+    def phoneme(self) -> Vowel | Consonant | Literal["sil"]:
         """このラベルに含まれる音素。子音 or 母音 (無音含む)。"""
-        # FIXME: バリデーションする
-        return self.contexts["p3"]  # type: ignore
+        p = self.contexts["p3"]
+        if p not in _OJT_PHONEMES:
+            raise NonOjtPhonemeError()
+        elif p == "xx":
+            raise OjtUnknownPhonemeError()
+        else:
+            # NOTE: mypy が型推論に失敗。pyright の推論した型が返り値型と一致することをマニュアル確認済み @2024-01-10 tarepan
+            return p  # type: ignore
 
     @property
     def mora_index(self) -> int: