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 32from torch.utils.data import Dataset 33import numpy as np 34 35from utils.silk_features import silk_feature_factory 36from utils.pitch import hangover, calculate_acorr_window 37 38 39class SilkEnhancementSet(Dataset): 40 def __init__(self, 41 path, 42 frames_per_sample=100, 43 no_pitch_value=256, 44 preemph=0.85, 45 skip=91, 46 acorr_radius=2, 47 pitch_hangover=8, 48 num_bands_clean_spec=64, 49 num_bands_noisy_spec=18, 50 noisy_spec_scale='opus', 51 noisy_apply_dct=True, 52 add_double_lag_acorr=False, 53 ): 54 55 assert frames_per_sample % 4 == 0 56 57 self.frame_size = 80 58 self.frames_per_sample = frames_per_sample 59 self.no_pitch_value = no_pitch_value 60 self.preemph = preemph 61 self.skip = skip 62 self.acorr_radius = acorr_radius 63 self.pitch_hangover = pitch_hangover 64 self.num_bands_clean_spec = num_bands_clean_spec 65 self.num_bands_noisy_spec = num_bands_noisy_spec 66 self.noisy_spec_scale = noisy_spec_scale 67 self.add_double_lag_acorr = add_double_lag_acorr 68 69 self.lpcs = np.fromfile(os.path.join(path, 'features_lpc.f32'), dtype=np.float32).reshape(-1, 16) 70 self.ltps = np.fromfile(os.path.join(path, 'features_ltp.f32'), dtype=np.float32).reshape(-1, 5) 71 self.periods = np.fromfile(os.path.join(path, 'features_period.s16'), dtype=np.int16) 72 self.gains = np.fromfile(os.path.join(path, 'features_gain.f32'), dtype=np.float32) 73 self.num_bits = np.fromfile(os.path.join(path, 'features_num_bits.s32'), dtype=np.int32) 74 self.num_bits_smooth = np.fromfile(os.path.join(path, 'features_num_bits_smooth.f32'), dtype=np.float32) 75 76 self.clean_signal_hp = np.fromfile(os.path.join(path, 'clean_hp.s16'), dtype=np.int16) 77 self.clean_signal = np.fromfile(os.path.join(path, 'clean.s16'), dtype=np.int16) 78 self.coded_signal = np.fromfile(os.path.join(path, 'coded.s16'), dtype=np.int16) 79 80 self.create_features = silk_feature_factory(no_pitch_value, 81 acorr_radius, 82 pitch_hangover, 83 num_bands_clean_spec, 84 num_bands_noisy_spec, 85 noisy_spec_scale, 86 noisy_apply_dct, 87 add_double_lag_acorr) 88 89 self.history_len = 700 if add_double_lag_acorr else 350 90 # discard some frames to have enough signal history 91 self.skip_frames = 4 * ((skip + self.history_len + 319) // 320 + 2) 92 93 num_frames = self.clean_signal_hp.shape[0] // 80 - self.skip_frames 94 95 self.len = num_frames // frames_per_sample 96 97 def __len__(self): 98 return self.len 99 100 def __getitem__(self, index): 101 102 frame_start = self.frames_per_sample * index + self.skip_frames 103 frame_stop = frame_start + self.frames_per_sample 104 105 signal_start = frame_start * self.frame_size - self.skip 106 signal_stop = frame_stop * self.frame_size - self.skip 107 108 clean_signal_hp = self.clean_signal_hp[signal_start : signal_stop].astype(np.float32) / 2**15 109 clean_signal = self.clean_signal[signal_start : signal_stop].astype(np.float32) / 2**15 110 coded_signal = self.coded_signal[signal_start : signal_stop].astype(np.float32) / 2**15 111 112 coded_signal_history = self.coded_signal[signal_start - self.history_len : signal_start].astype(np.float32) / 2**15 113 114 features, periods = self.create_features( 115 coded_signal, 116 coded_signal_history, 117 self.lpcs[frame_start : frame_stop], 118 self.gains[frame_start : frame_stop], 119 self.ltps[frame_start : frame_stop], 120 self.periods[frame_start : frame_stop] 121 ) 122 123 if self.preemph > 0: 124 clean_signal[1:] -= self.preemph * clean_signal[: -1] 125 clean_signal_hp[1:] -= self.preemph * clean_signal_hp[: -1] 126 coded_signal[1:] -= self.preemph * coded_signal[: -1] 127 128 num_bits = np.repeat(self.num_bits[frame_start // 4 : frame_stop // 4], 4).astype(np.float32).reshape(-1, 1) 129 num_bits_smooth = np.repeat(self.num_bits_smooth[frame_start // 4 : frame_stop // 4], 4).astype(np.float32).reshape(-1, 1) 130 131 numbits = np.concatenate((num_bits, num_bits_smooth), axis=-1) 132 133 return { 134 'features' : features, 135 'periods' : periods.astype(np.int64), 136 'target_orig' : clean_signal.astype(np.float32), 137 'target' : clean_signal_hp.astype(np.float32), 138 'signals' : coded_signal.reshape(-1, 1).astype(np.float32), 139 'numbits' : numbits.astype(np.float32) 140 } 141