© DeNA Co., Ltd. 1 F0推定の手法を色々 試してみる nagiss

© DeNA Co., Ltd. 2 1 ● F0 推定の手法を同じインタフェースで使えるようにした ● 処理速度や WORLD で再合成した時の品質などを大雑把に比べてみた 概要

© DeNA Co., Ltd. 3 2 ● SWIPE, RAPT, YAAPT, REAPER, DIO, Harvest, pYIN, CREPE, FCPE ○ ここでは各手法の中身は解説しません ● Python のライブラリとして利用可能かつライセンスがゆるかったものを選んだ 試した手法

© DeNA Co., Ltd. 4 3 同じインタフェースで使えるようにした class DIO(PitchEstimator): def __init__( self, sample_rate: int, hop_length: int, fmin: float = 71.0, fmax: float = 800.0, with_stonemask: bool = True, ): self.sample_rate = sample_rate self.hop_length = hop_length self.fmin = fmin self.fmax = fmax self.with_stonemask = with_stonemask def estimate( self, wav: npt.NDArray[np.float32] ) -> tuple[npt.NDArray[np.float32], dict[str, Any]]: wav = wav.astype(np.float64) pitch, t = pyworld.dio( wav, fs=self.sample_rate, f0_ceil=self.fmax, f0_floor=self.fmin, frame_period=1000.0 * self.hop_length / self.sample_rate, ) if self.with_stonemask: pitch = pyworld.stonemask(wav, pitch, t, self.sample_rate) return pitch.astype(np.float32), {"t": t} ● DIO の例 ● 手法ごとに色々な部分が色々と異なる ○ 入力が int16 か float32 か float64 か ○ フレームシフトがサンプル数単位か ミリ秒単位か ○ etc. ● この辺りを吸収して雑に使えるように ● (結果の系列長みたいなところまでは 揃えられていない)

© DeNA Co., Ltd. 5 4 ● 数秒の音声 ● 全て CPU で動作 速度を大雑把に比較 遅 速 EPE RTF: 1 0.1 0.01 pYIN Harvest DIO RAPT SWIPE REAPER FCPE (1スレッド) FCPE (8スレッド) YAAPT

© DeNA Co., Ltd. 6 5 ● いくつかの喋り声で試して体感で確認 ○ 音声にはノイズも多少乗ってる ● 各種パラメータはほぼ各手法のデフォルトのまま ○ F0 の上限と下限は統一した WORLDで再合成したときの自然さを大雑把に比較

© DeNA Co., Ltd. 7 6 ● CREPE ≒ Harvest ≧ FCPE ≒ YAAPT ≧ RAPT ≧ DIO ≧ その他 ● 注意: F0 推定の正確さとはかなり異なる尺度になっていると思われる ○ 有声部を無声部として合成してしまうと耳障りなノイズが発生するが (逆は軽 微)、CREPE はデフォルトで全てのフレームの音高を出すのでこれが無い ■ デフォルト設定での有声判定しやすさと強く相関した尺度になっていそう ○ CREPE は所々不正確なピッチで合成されるが、短い部分なので違和感が少ない WORLDで再合成したときの自然さを大雑把に比較 有声部分が無声判定された時の雑音 (pYIN)

© DeNA Co., Ltd. 8 7 ● CREPE ○ 計算重すぎ ○ Viterbi のせいか短い区間で嘘をつきがちだけど生成結果はそれっぽい感じになる 所感とか 地上の真実 再合成 「です」が下がっていない

© DeNA Co., Ltd. 9 8 ● Harvest ○ 信頼と実績の WORLD ボコーダという感じ ○ デフォルトの下限が結構高い (71Hz) ので注意が必要 ○ 上限と下限の間を狭めると有声判定 が増える (右図) 所感とか

© DeNA Co., Ltd. 10 9 ● RAPT ○ 爆速な割に再合成の品質が良い ○ 若干他より早まった軌跡になっている (原因不明) ● YAAPT ○ 聞いたことない手法だったけど案外良い ● pYIN ○ 自己相関の計算がフレーム内の左側に寄っていて遅れた軌跡になるので、 正確に取りたいなら修正した方が良さそう 所感とか

© DeNA Co., Ltd. 11 10 ● FCPE ○ 速度と品質のバランスが良い感じ ○ 最近傍補間周りの処理が雑めな気がするのは気になる ● 全体 ○ 再合成では細かいピッチの違いは意外と気にならないなあと思った 所感とか

© DeNA Co., Ltd. 12

© DeNA Co., Ltd. 13 実験に使ったコード ex from abc import ABC, abstractmethod from typing import Any, Callable, Literal, Optional, Union import amfm_decompy.basic_tools import amfm_decompy.pYAAPT import librosa import numpy as np import numpy.typing as npt import pyreaper import pysptk import pyworld import torch import torchcrepe import torchfcpe class PitchEstimator(ABC): @abstractmethod def estimate( self, wav: npt.NDArray[np.float32] ) -> tuple[npt.NDArray[np.float32], dict[str, Any]]: pass [tool.poetry.dependencies] python = "3.10" pyworld = "0.3.4" librosa = "0.10.1" torchcrepe = "0.0.22" pyreaper = "0.0.10" pysptk = "0.2.2" torchfcpe = "0.0.4" amfm-decompy = "1.0.11" matplotlib = "3.8.3"

© DeNA Co., Ltd. 14 class CREPE(PitchEstimator): def __init__( self, sample_rate: int, hop_length: int, fmin: float = 50.0, fmax: float = torchcrepe.MAX_FMAX, model: Literal["full", "tiny"] = "full", decoder: Callable = torchcrepe.decode.viterbi, return_periodicity: bool = False, batch_size: Optional[int] = None, device="cpu", pad: bool = True, ): self.sample_rate = sample_rate self.hop_length = hop_length self.fmin = fmin self.fmax = fmax self.model = model self.decoder = decoder self.return_periodicity = return_periodicity self.batch_size = batch_size self.device = device self.pad = pad def estimate( self, wav: npt.NDArray[np.float32] ) -> tuple[npt.NDArray[np.float32], dict[str, Any]]: if isinstance(wav, np.ndarray): wav = torch.from_numpy(wav).float() results = torchcrepe.predict( wav[None], sample_rate=self.sample_rate, hop_length=self.hop_length, fmin=self.fmin, fmax=self.fmax, model=self.model, decoder=self.decoder, return_periodicity=self.return_periodicity, batch_size=self.batch_size, device=self.device, pad=self.pad, ) if self.return_periodicity: pitch, periodicity = results pitch = pitch.squeeze_(0).numpy() periodicity = periodicity.numpy() return pitch, {"periodicity": periodicity} else: pitch = results.squeeze_(0).numpy() return pitch, {} class PYIN(PitchEstimator): def __init__( self, sample_rate: int, hop_length: int, fmin: float = librosa.note_to_hz("C2"), fmax: float = librosa.note_to_hz("C7"), frame_length: int = None, win_length: Optional[int] = None, n_thresholds: int = 100, beta_parameters: tuple[float, float] = (2, 18), boltzmann_parameter: float = 2, resolution: float = 0.1, max_transition_rate: float = 35.92, switch_prob: float = 0.01, no_trough_prob: float = 0.01, fill_na: Optional[float] = np.nan, center: bool = True, pad_mode: Union[ Literal[ "constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty", ], Callable, ] = "constant", ): self.sample_rate = sample_rate self.hop_length = hop_length self.fmin = fmin self.fmax = fmax self.frame_length = sample_rate // 20 if frame_length is None else frame_length self.win_length = win_length self.n_thresholds = n_thresholds self.beta_parameters = beta_parameters self.boltzmann_parameter = boltzmann_parameter self.resolution = resolution self.max_transition_rate = max_transition_rate self.switch_prob = switch_prob self.no_trough_prob = no_trough_prob self.fill_na = fill_na self.center = center self.pad_mode = pad_mode def estimate( self, wav: npt.NDArray[np.float32] ) -> tuple[npt.NDArray[np.float32], dict[str, Any]]: pitch, voiced_flag, voiced_prob = librosa.pyin( wav, fmin=self.fmin, fmax=self.fmax, sr=self.sample_rate, frame_length=self.frame_length, win_length=self.win_length, hop_length=self.hop_length, n_thresholds=self.n_thresholds, beta_parameters=self.beta_parameters, boltzmann_parameter=self.boltzmann_parameter, resolution=self.resolution, max_transition_rate=self.max_transition_rate, switch_prob=self.switch_prob, no_trough_prob=self.no_trough_prob, fill_na=self.fill_na, center=self.center, pad_mode=self.pad_mode, ) return np.nan_to_num(pitch.astype(np.float32)), { "voiced_flag": voiced_flag, "voiced_prob": voiced_prob, }

© DeNA Co., Ltd. 15 class DIO(PitchEstimator): def __init__( self, sample_rate: int, hop_length: int, fmin: float = 71.0, fmax: float = 800.0, with_stonemask: bool = True, ): self.sample_rate = sample_rate self.hop_length = hop_length self.fmin = fmin self.fmax = fmax self.with_stonemask = with_stonemask def estimate( self, wav: npt.NDArray[np.float32] ) -> tuple[npt.NDArray[np.float32], dict[str, Any]]: wav = wav.astype(np.float64) pitch, t = pyworld.dio( wav, fs=self.sample_rate, f0_ceil=self.fmax, f0_floor=self.fmin, frame_period=1000.0 * self.hop_length / self.sample_rate, ) if self.with_stonemask: pitch = pyworld.stonemask(wav, pitch, t, self.sample_rate) return pitch.astype(np.float32), {"t": t} class Harvest(PitchEstimator): def __init__( self, sample_rate: int, hop_length: int, fmin: float = 71.0, fmax: float = 800.0, ): self.sample_rate = sample_rate self.hop_length = hop_length self.fmin = fmin self.fmax = fmax def estimate( self, wav: npt.NDArray[np.float32] ) -> tuple[npt.NDArray[np.float32], dict[str, Any]]: pitch, t = pyworld.harvest( wav.astype(np.double), fs=self.sample_rate, f0_ceil=self.fmax, f0_floor=self.fmin, frame_period=1000.0 * self.hop_length / self.sample_rate, ) return pitch.astype(np.float32), {"t": t} class RAPT(PitchEstimator): def __init__( self, sample_rate: int, hop_length: int, fmin: float = 60.0, fmax: float = 240.0, voice_bias: float = 0.0, ): self.sample_rate = sample_rate self.hop_length = hop_length self.fmin = fmin self.fmax = fmax self.voice_bias = voice_bias def estimate( self, wav: npt.NDArray[np.float32] ) -> tuple[npt.NDArray[np.float32], dict[str, Any]]: pitch = pysptk.rapt( (wav * 32767.0).astype(np.float32), fs=self.sample_rate, hopsize=self.hop_length, min=self.fmin, max=self.fmax, voice_bias=self.voice_bias, ) return pitch.astype(np.float32), {} class SWIPE(PitchEstimator): def __init__( self, sample_rate: int, hop_length: int, fmin: float = 60.0, fmax: float = 240.0, threshold: float = 0.3, ): self.sample_rate = sample_rate self.hop_length = hop_length self.fmin = fmin self.fmax = fmax self.threshold = threshold def estimate( self, wav: npt.NDArray[np.float32] ) -> tuple[npt.NDArray[np.float32], dict[str, Any]]: pitch = pysptk.swipe( (wav * 32767.0).astype(np.float64), fs=self.sample_rate, hopsize=self.hop_length, min=self.fmin, max=self.fmax, threshold=self.threshold, ) return pitch.astype(np.float32), {}

© DeNA Co., Ltd. 16 self.fft_length = fft_length self.bp_forder = bp_forder self.bp_low = bp_low self.bp_high = bp_high self.nlfer_thresh1 = nlfer_thresh1 self.nlfer_thresh2 = nlfer_thresh2 self.shc_numharms = shc_numharms self.shc_window = shc_window self.shc_maxpeaks = shc_maxpeaks self.shc_pwidth = shc_pwidth self.shc_thresh1 = shc_thresh1 self.shc_thresh2 = shc_thresh2 self.f0_double = f0_double self.f0_half = f0_half self.dp5_k1 = dp5_k1 self.dec_factor = dec_factor self.nccf_thresh1 = nccf_thresh1 self.nccf_thresh2 = nccf_thresh2 self.nccf_maxcands = nccf_maxcands self.nccf_pwidth = nccf_pwidth self.merit_boost = merit_boost self.merit_pivot = merit_pivot self.merit_extra = merit_extra self.median_value = median_value self.dp_w1 = dp_w1 self.dp_w2 = dp_w2 self.dp_w3 = dp_w3 self.dp_w4 = dp_w4 self.spec_pitch_min_std = spec_pitch_min_std self.pitch_half = pitch_half self.pitch_half_sens = pitch_half_sens self.pitch_double = pitch_double self.pitch_double_sens = pitch_double_sens self.smooth_factor = smooth_factor self.smooth = smooth self.ptch_typ = ptch_typ def estimate( self, wav: npt.NDArray[np.float32] ) -> tuple[npt.NDArray[np.float32], dict[str, Any]]: wav = amfm_decompy.basic_tools.SignalObj(wav, self.sample_rate) amfm_decompy.pYAAPT.PitchObj.PITCH_HALF = self.pitch_half amfm_decompy.pYAAPT.PitchObj.PITCH_HALF_SENS = self.pitch_half_sens amfm_decompy.pYAAPT.PitchObj.PITCH_DOUBLE = self.pitch_double amfm_decompy.pYAAPT.PitchObj.PITCH_DOUBLE_SENS = self.pitch_double_sens amfm_decompy.pYAAPT.PitchObj.SMOOTH_FACTOR = self.smooth_factor amfm_decompy.pYAAPT.PitchObj.SMOOTH = self.smooth amfm_decompy.pYAAPT.PitchObj.PTCH_TYP = self.ptch_typ pitch = amfm_decompy.pYAAPT.yaapt( wav, frame_length=self.frame_length, tda_frame_length=self.tda_frame_length, frame_space=self.frame_space, class YAAPT(PitchEstimator): def __init__( self, sample_rate: int, hop_length: int, fmin: float = 60.0, fmax: float = 400.0, frame_length: float = 35.0, tda_frame_length: float = 35.0, frame_space: float = 10.0, fft_length: int = 8192, bp_forder: int = 150, bp_low: float = 50.0, bp_high: float = 1500.0, nlfer_thresh1: float = 0.75, nlfer_thresh2: float = 0.1, shc_numharms: int = 3, shc_window: float = 40.0, shc_maxpeaks: int = 4, shc_pwidth: float = 50.0, shc_thresh1: float = 5.0, shc_thresh2: float = 1.25, f0_double: float = 150.0, f0_half: float = 150.0, dp5_k1: float = 11.0, dec_factor: int = 1, nccf_thresh1: float = 0.3, nccf_thresh2: float = 0.9, nccf_maxcands: int = 3, nccf_pwidth: int = 5, merit_boost: float = 0.20, merit_pivot: float = 0.99, merit_extra: float = 0.4, median_value: int = 7, dp_w1: float = 0.15, dp_w2: float = 0.5, dp_w3: float = 0.1, dp_w4: float = 0.9, spec_pitch_min_std: float = 0.05, pitch_half: int = 0, pitch_half_sens: float = 2.9, pitch_double: int = 0, pitch_double_sens: float = 2.9, smooth_factor: int = 5, smooth: int = 5, ptch_typ: float = 100.0, ): self.sample_rate = sample_rate self.hop_length = hop_length self.fmin = fmin self.fmax = fmax self.frame_length = frame_length self.tda_frame_length = tda_frame_length self.frame_space = frame_space f0_min=self.fmin, f0_max=self.fmax, fft_length=self.fft_length, bp_forder=self.bp_forder, bp_low=self.bp_low, bp_high=self.bp_high, nlfer_thresh1=self.nlfer_thresh1, nlfer_thresh2=self.nlfer_thresh2, shc_numharms=self.shc_numharms, shc_window=self.shc_window, shc_maxpeaks=self.shc_maxpeaks, shc_pwidth=self.shc_pwidth, shc_thresh1=self.shc_thresh1, shc_thresh2=self.shc_thresh2, f0_double=self.f0_double, f0_half=self.f0_half, dp5_k1=self.dp5_k1, dec_factor=self.dec_factor, nccf_thresh1=self.nccf_thresh1, nccf_thresh2=self.nccf_thresh2, nccf_maxcands=self.nccf_maxcands, nccf_pwidth=self.nccf_pwidth, merit_boost=self.merit_boost, merit_pivot=self.merit_pivot, merit_extra=self.merit_extra, median_value=self.median_value, dp_w1=self.dp_w1, dp_w2=self.dp_w2, dp_w3=self.dp_w3, dp_w4=self.dp_w4, spec_pitch_min_std=self.spec_pitch_min_std, ) stats = {} for name in [ "nfft", "frame_size", "frame_jump", "noverlap", "mean_energy", "energy", "vuv", "frame_pos", "nframes", "samp_values", "values", "edges", "samp_interp", "values_interp", ]: if hasattr(pitch, name): stats[name] = getattr(pitch, name) return pitch.values[:: self.hop_length].copy(), stats

© DeNA Co., Ltd. 17 class REAPER(PitchEstimator): def __init__( self, sample_rate: int, hop_length: int, fmin: float = 40.0, fmax: float = 500.0, do_high_pass: bool = True, do_hilbert_transform: bool = False, inter_pulse: float = 0.01, unvoiced_cost: float = 0.1, ): self.sample_rate = sample_rate self.hop_length = hop_length self.fmin = fmin self.fmax = fmax self.do_high_pass = do_high_pass self.do_hilbert_transform = do_hilbert_transform self.inter_pulse = inter_pulse self.unvoiced_cost = unvoiced_cost def estimate( self, wav: npt.NDArray[np.float32] ) -> tuple[npt.NDArray[np.float32], dict[str, Any]]: pm_times, pm, f0_times, f0, corr = pyreaper.reaper( np.round(wav * 32767.0).astype(np.int16), fs=self.sample_rate, minf0=self.fmin, maxf0=self.fmax, do_high_pass=self.do_high_pass, do_hilbert_transform=self.do_hilbert_transform, inter_pulse=self.inter_pulse, frame_period=self.hop_length / self.sample_rate, unvoiced_cost=self.unvoiced_cost, ) f0 = f0.clip(min=0.0) return f0.astype(np.float32), { "pm_times": pm_times, "pm": pm, "f0_times": f0_times, "corr": corr, } class FCPE(PitchEstimator): def __init__( self, sample_rate: int, hop_length: int, fmin: float = 80.0, fmax: float = 880.0, decoder_mode: Literal["argmax", "local_argmax"] = "local_argmax", threshold: float = 0.006, interp_uv: bool = False, ): self.sample_rate = sample_rate self.hop_length = hop_length self.fmin = fmin self.fmax = fmax self.decoder_mode = decoder_mode self.threshold = threshold self.interp_uv = interp_uv self.model = torchfcpe.spawn_bundled_infer_model() def estimate( self, wav: npt.NDArray[np.float32] ) -> tuple[npt.NDArray[np.float32], dict[str, Any]]: output_interp_target_length = (len(wav) // self.hop_length) + 1 wav = torch.from_numpy(wav).float()[None, ..., None] f0, uv = self.model.infer( wav, sr=self.sample_rate, decoder_mode="local_argmax", threshold=self.threshold, f0_min=self.fmin, f0_max=self.fmax, interp_uv=self.interp_uv, output_interp_target_length=output_interp_target_length, retur_uv=True, ) f0 = f0.squeeze(0).squeeze(-1).numpy() uv = uv.squeeze(0).squeeze(-1).numpy() return f0, {"uv": uv}