1""" 2/* Copyright (c) 2023 Amazon 3 Written by Jan Buethe */ 4/* 5 Redistribution and use in source and binary forms, with or without 6 modification, are permitted provided that the following conditions 7 are met: 8 9 - Redistributions of source code must retain the above copyright 10 notice, this list of conditions and the following disclaimer. 11 12 - Redistributions in binary form must reproduce the above copyright 13 notice, this list of conditions and the following disclaimer in the 14 documentation and/or other materials provided with the distribution. 15 16 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 17 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 18 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 19 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER 20 OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 21 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 22 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 23 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 24 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 25 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 26 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27*/ 28""" 29 30import os 31 32import numpy as np 33import torch 34 35import scipy 36import scipy.signal 37 38from utils.pitch import hangover, calculate_acorr_window 39from utils.spec import create_filter_bank, cepstrum, log_spectrum, log_spectrum_from_lpc 40 41def spec_from_lpc(a, n_fft=128, eps=1e-9): 42 order = a.shape[-1] 43 assert order + 1 < n_fft 44 45 x = np.zeros((*a.shape[:-1], n_fft )) 46 x[..., 0] = 1 47 x[..., 1:1 + order] = -a 48 49 X = np.fft.fft(x, axis=-1) 50 X = np.abs(X[..., :n_fft//2 + 1]) ** 2 51 52 S = 1 / (X + eps) 53 54 return S 55 56def silk_feature_factory(no_pitch_value=256, 57 acorr_radius=2, 58 pitch_hangover=8, 59 num_bands_clean_spec=64, 60 num_bands_noisy_spec=18, 61 noisy_spec_scale='opus', 62 noisy_apply_dct=True, 63 add_double_lag_acorr=False 64 ): 65 66 w = scipy.signal.windows.cosine(320) 67 fb_clean_spec = create_filter_bank(num_bands_clean_spec, 320, scale='erb', round_center_bins=True, normalize=True) 68 fb_noisy_spec = create_filter_bank(num_bands_noisy_spec, 320, scale=noisy_spec_scale, round_center_bins=True, normalize=True) 69 70 def create_features(noisy, noisy_history, lpcs, gains, ltps, periods): 71 72 periods = periods.copy() 73 74 if pitch_hangover > 0: 75 periods = hangover(periods, num_frames=pitch_hangover) 76 77 periods[periods == 0] = no_pitch_value 78 79 clean_spectrum = 0.3 * log_spectrum_from_lpc(lpcs, fb=fb_clean_spec, n_fft=320) 80 81 if noisy_apply_dct: 82 noisy_cepstrum = np.repeat( 83 cepstrum(np.concatenate((noisy_history[-160:], noisy), dtype=np.float32), 320, fb_noisy_spec, w), 2, 0) 84 else: 85 noisy_cepstrum = np.repeat( 86 log_spectrum(np.concatenate((noisy_history[-160:], noisy), dtype=np.float32), 320, fb_noisy_spec, w), 2, 0) 87 88 log_gains = np.log(gains + 1e-9).reshape(-1, 1) 89 90 acorr, _ = calculate_acorr_window(noisy, 80, periods, noisy_history, radius=acorr_radius, add_double_lag_acorr=add_double_lag_acorr) 91 92 features = np.concatenate((clean_spectrum, noisy_cepstrum, acorr, ltps, log_gains), axis=-1, dtype=np.float32) 93 94 return features, periods.astype(np.int64) 95 96 return create_features 97 98 99 100def load_inference_data(path, 101 no_pitch_value=256, 102 skip=92, 103 preemph=0.85, 104 acorr_radius=2, 105 pitch_hangover=8, 106 num_bands_clean_spec=64, 107 num_bands_noisy_spec=18, 108 noisy_spec_scale='opus', 109 noisy_apply_dct=True, 110 add_double_lag_acorr=False, 111 **kwargs): 112 113 print(f"[load_inference_data]: ignoring keyword arguments {kwargs.keys()}...") 114 115 lpcs = np.fromfile(os.path.join(path, 'features_lpc.f32'), dtype=np.float32).reshape(-1, 16) 116 ltps = np.fromfile(os.path.join(path, 'features_ltp.f32'), dtype=np.float32).reshape(-1, 5) 117 gains = np.fromfile(os.path.join(path, 'features_gain.f32'), dtype=np.float32) 118 periods = np.fromfile(os.path.join(path, 'features_period.s16'), dtype=np.int16) 119 num_bits = np.fromfile(os.path.join(path, 'features_num_bits.s32'), dtype=np.int32).astype(np.float32).reshape(-1, 1) 120 num_bits_smooth = np.fromfile(os.path.join(path, 'features_num_bits_smooth.f32'), dtype=np.float32).reshape(-1, 1) 121 122 # load signal, add back delay and pre-emphasize 123 signal = np.fromfile(os.path.join(path, 'noisy.s16'), dtype=np.int16).astype(np.float32) / (2 ** 15) 124 signal = np.concatenate((np.zeros(skip, dtype=np.float32), signal), dtype=np.float32) 125 126 create_features = silk_feature_factory(no_pitch_value, acorr_radius, pitch_hangover, num_bands_clean_spec, num_bands_noisy_spec, noisy_spec_scale, noisy_apply_dct, add_double_lag_acorr) 127 128 num_frames = min((len(signal) // 320) * 4, len(lpcs)) 129 signal = signal[: num_frames * 80] 130 lpcs = lpcs[: num_frames] 131 ltps = ltps[: num_frames] 132 gains = gains[: num_frames] 133 periods = periods[: num_frames] 134 num_bits = num_bits[: num_frames // 4] 135 num_bits_smooth = num_bits[: num_frames // 4] 136 137 numbits = np.repeat(np.concatenate((num_bits, num_bits_smooth), axis=-1, dtype=np.float32), 4, axis=0) 138 139 features, periods = create_features(signal, np.zeros(350, dtype=signal.dtype), lpcs, gains, ltps, periods) 140 141 if preemph > 0: 142 signal[1:] -= preemph * signal[:-1] 143 144 return torch.from_numpy(signal), torch.from_numpy(features), torch.from_numpy(periods), torch.from_numpy(numbits) 145