voicevox / voicevox_engine /morphing.py
GaenKoki's picture
Duplicate from 2ndelement/voicevox
5cda731
from copy import deepcopy
from dataclasses import dataclass
from itertools import chain
from typing import Dict, List, Tuple
import numpy as np
import pyworld as pw
from scipy.signal import resample
from .metas.Metas import Speaker, SpeakerSupportPermittedSynthesisMorphing, StyleInfo
from .metas.MetasStore import construct_lookup
from .model import AudioQuery, MorphableTargetInfo, SpeakerNotFoundError
from .synthesis_engine import SynthesisEngine
# FIXME: ndarray type hint, https://github.com/JeremyCCHsu/Python-Wrapper-for-World-Vocoder/blob/2b64f86197573497c685c785c6e0e743f407b63e/pyworld/pyworld.pyx#L398 # noqa
@dataclass(frozen=True)
class MorphingParameter:
fs: int
frame_period: float
base_f0: np.ndarray
base_aperiodicity: np.ndarray
base_spectrogram: np.ndarray
target_spectrogram: np.ndarray
def create_morphing_parameter(
base_wave: np.ndarray,
target_wave: np.ndarray,
fs: int,
) -> MorphingParameter:
frame_period = 1.0
base_f0, base_time_axis = pw.harvest(base_wave, fs, frame_period=frame_period)
base_spectrogram = pw.cheaptrick(base_wave, base_f0, base_time_axis, fs)
base_aperiodicity = pw.d4c(base_wave, base_f0, base_time_axis, fs)
target_f0, morph_time_axis = pw.harvest(target_wave, fs, frame_period=frame_period)
target_spectrogram = pw.cheaptrick(target_wave, target_f0, morph_time_axis, fs)
target_spectrogram.resize(base_spectrogram.shape)
return MorphingParameter(
fs=fs,
frame_period=frame_period,
base_f0=base_f0,
base_aperiodicity=base_aperiodicity,
base_spectrogram=base_spectrogram,
target_spectrogram=target_spectrogram,
)
def get_morphable_targets(
speakers: List[Speaker],
base_speakers: List[int],
) -> List[Dict[int, MorphableTargetInfo]]:
"""
speakers: 全話者の情報
base_speakers: モーフィング可能か判定したいベースの話者リスト(スタイルID)
"""
speaker_lookup = construct_lookup(speakers)
morphable_targets_arr = []
for base_speaker in base_speakers:
morphable_targets = dict()
for style in chain.from_iterable(speaker.styles for speaker in speakers):
morphable_targets[style.id] = MorphableTargetInfo(
is_morphable=is_synthesis_morphing_permitted(
speaker_lookup=speaker_lookup,
base_speaker=base_speaker,
target_speaker=style.id,
)
)
morphable_targets_arr.append(morphable_targets)
return morphable_targets_arr
def is_synthesis_morphing_permitted(
speaker_lookup: Dict[int, Tuple[Speaker, StyleInfo]],
base_speaker: int,
target_speaker: int,
) -> bool:
"""
指定されたspeakerがモーフィング可能かどうか返す
speakerが見つからない場合はSpeakerNotFoundErrorを送出する
"""
base_speaker_data = speaker_lookup[base_speaker]
target_speaker_data = speaker_lookup[target_speaker]
if base_speaker_data is None or target_speaker_data is None:
raise SpeakerNotFoundError(
base_speaker if base_speaker_data is None else target_speaker
)
base_speaker_info, _ = base_speaker_data
target_speaker_info, _ = target_speaker_data
base_speaker_uuid = base_speaker_info.speaker_uuid
target_speaker_uuid = target_speaker_info.speaker_uuid
base_speaker_morphing_info: SpeakerSupportPermittedSynthesisMorphing = (
base_speaker_info.supported_features.permitted_synthesis_morphing
)
target_speaker_morphing_info: SpeakerSupportPermittedSynthesisMorphing = (
target_speaker_info.supported_features.permitted_synthesis_morphing
)
# 禁止されている場合はFalse
if (
base_speaker_morphing_info == SpeakerSupportPermittedSynthesisMorphing.NOTHING
or target_speaker_morphing_info
== SpeakerSupportPermittedSynthesisMorphing.NOTHING
):
return False
# 同一話者のみの場合は同一話者判定
if (
base_speaker_morphing_info == SpeakerSupportPermittedSynthesisMorphing.SELF_ONLY
or target_speaker_morphing_info
== SpeakerSupportPermittedSynthesisMorphing.SELF_ONLY
):
return base_speaker_uuid == target_speaker_uuid
# 念のため許可されているかチェック
return (
base_speaker_morphing_info == SpeakerSupportPermittedSynthesisMorphing.ALL
and target_speaker_morphing_info == SpeakerSupportPermittedSynthesisMorphing.ALL
)
def synthesis_morphing_parameter(
engine: SynthesisEngine,
query: AudioQuery,
base_speaker: int,
target_speaker: int,
) -> MorphingParameter:
query = deepcopy(query)
# 不具合回避のためデフォルトのサンプリングレートでWORLDに掛けた後に指定のサンプリングレートに変換する
query.outputSamplingRate = engine.default_sampling_rate
# WORLDに掛けるため合成はモノラルで行う
query.outputStereo = False
base_wave = engine.synthesis(query=query, speaker_id=base_speaker).astype("float")
target_wave = engine.synthesis(query=query, speaker_id=target_speaker).astype(
"float"
)
return create_morphing_parameter(
base_wave=base_wave,
target_wave=target_wave,
fs=query.outputSamplingRate,
)
def synthesis_morphing(
morph_param: MorphingParameter,
morph_rate: float,
output_fs: int,
output_stereo: bool = False,
) -> np.ndarray:
"""
指定した割合で、パラメータをもとにモーフィングした音声を生成します。
Parameters
----------
morph_param : MorphingParameter
`synthesis_morphing_parameter`または`create_morphing_parameter`で作成したパラメータ
morph_rate : float
モーフィングの割合
0.0でベースの話者、1.0でターゲットの話者に近づきます。
Returns
-------
generated : np.ndarray
モーフィングした音声
Raises
-------
ValueError
morph_rate ∈ [0, 1]
"""
if morph_rate < 0.0 or morph_rate > 1.0:
raise ValueError("morph_rateは0.0から1.0の範囲で指定してください")
morph_spectrogram = (
morph_param.base_spectrogram * (1.0 - morph_rate)
+ morph_param.target_spectrogram * morph_rate
)
y_h = pw.synthesize(
morph_param.base_f0,
morph_spectrogram,
morph_param.base_aperiodicity,
morph_param.fs,
morph_param.frame_period,
)
# TODO: synthesis_engine.py でのリサンプル処理と共通化する
if output_fs != morph_param.fs:
y_h = resample(y_h, output_fs * len(y_h) // morph_param.fs)
if output_stereo:
y_h = np.array([y_h, y_h]).T
return y_h