Improve: switch pyworld to pyworld-prebuilt and enable it by default

Prebuilt wheels for almost every OS/architecture (except musl) are now available on PyPI, eliminating the need for a build environment.
ref: https://pypi.org/project/pyworld-prebuilt/

pyproject.toml

@@ -36,7 +36,7 @@ dependencies = [
     'pydantic>=2.0',
     'pyopenjtalk-dict',
     'pypinyin',
-    # 'pyworld',
+    'pyworld-prebuilt',
     'safetensors',
     'scipy',
     'torch>=2.1,<2.2',

requirements.txt

@@ -19,7 +19,7 @@ pyloudnorm
 # pyopenjtalk-prebuilt  # Should be manually uninstalled
 pyopenjtalk-dict
 pypinyin
-# pyworld  # Not supported on Windows without Cython...
+pyworld-prebuilt
 PyYAML
 requests
 safetensors

style_bert_vits2/voice.py

@@ -1,6 +1,7 @@
 from typing import Any
 
 import numpy as np
+import pyworld
 from numpy.typing import NDArray
 
 
@@ -10,29 +11,34 @@ def adjust_voice(
     pitch_scale: float = 1.0,
     intonation_scale: float = 1.0,
 ) -> tuple[int, NDArray[Any]]:
+    """
+    音声のピッチとイントネーションを調整する。
+    変更すると若干音質が劣化するので、どちらも初期値のままならそのまま返す。
+
+    Args:
+        fs (int): 音声のサンプリング周波数
+        wave (NDArray[Any]): 音声データ
+        pitch_scale (float, optional): ピッチの高さ. Defaults to 1.0.
+        intonation_scale (float, optional): イントネーションの高さ. Defaults to 1.0.
+
+    Returns:
+        tuple[int, NDArray[Any]]: 調整後の音声データのサンプリング周波数と音声データ
+    """
 
     if pitch_scale == 1.0 and intonation_scale == 1.0:
         # 初期値の場合は、音質劣化を避けるためにそのまま返す
         return fs, wave
 
-    try:
-        import pyworld
-    except ImportError:
-        raise ImportError(
-            "pyworld is not installed. Please install it by `pip install pyworld`"
-        )
-
     # pyworld で f0 を加工して合成
     # pyworld よりもよいのがあるかもしれないが……
-    ## pyworld は Cython で書かれているが、スタブファイルがないため型補完が全く効かない…
 
     wave = wave.astype(np.double)
 
     # 質が高そうだしとりあえずharvestにしておく
-    f0, t = pyworld.harvest(wave, fs)  # type: ignore
+    f0, t = pyworld.harvest(wave, fs)
 
-    sp = pyworld.cheaptrick(wave, f0, t, fs)  # type: ignore
-    ap = pyworld.d4c(wave, f0, t, fs)  # type: ignore
+    sp = pyworld.cheaptrick(wave, f0, t, fs)
+    ap = pyworld.d4c(wave, f0, t, fs)
 
     non_zero_f0 = [f for f in f0 if f != 0]
     f0_mean = sum(non_zero_f0) / len(non_zero_f0)
@@ -42,5 +48,5 @@ def adjust_voice(
             continue
         f0[i] = pitch_scale * f0_mean + intonation_scale * (f - f0_mean)
 
-    wave = pyworld.synthesize(f0, sp, ap, fs)  # type: ignore
+    wave = pyworld.synthesize(f0, sp, ap, fs)
     return fs, wave

webui/inference.py

@@ -294,7 +294,6 @@ def create_inference_app(model_holder: TTSModelHolder) -> gr.Blocks:
                     value=1,
                     step=0.05,
                     label="音程(1以外では音質劣化)",
-                    visible=False,  # pyworldが必要
                 )
                 intonation_scale = gr.Slider(
                     minimum=0,
@@ -302,7 +301,6 @@ def create_inference_app(model_holder: TTSModelHolder) -> gr.Blocks:
                     value=1,
                     step=0.1,
                     label="抑揚(1以外では音質劣化)",
-                    visible=False,  # pyworldが必要
                 )
 
                 line_split = gr.Checkbox(