File size: 8,578 Bytes
2072d0c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 |
import os
import torch
import librosa
import argparse
import numpy as np
import soundfile as sf
import pyworld as pw
import parselmouth
from ast import literal_eval
from slicer import Slicer
from ddsp.vocoder import load_model, F0_Extractor, Volume_Extractor, Units_Encoder
from ddsp.core import upsample
from enhancer import Enhancer
from tqdm import tqdm
def parse_args(args=None, namespace=None):
"""Parse command-line arguments."""
parser = argparse.ArgumentParser()
parser.add_argument(
"-m",
"--model_path",
type=str,
required=True,
help="path to the model file",
)
parser.add_argument(
"-i",
"--input",
type=str,
required=True,
help="path to the input audio file",
)
parser.add_argument(
"-o",
"--output",
type=str,
required=True,
help="path to the output audio file",
)
parser.add_argument(
"-id",
"--spk_id",
type=str,
required=False,
default=1,
help="speaker id (for multi-speaker model) | default: 1",
)
parser.add_argument(
"-mix",
"--spk_mix_dict",
type=str,
required=False,
default="None",
help="mix-speaker dictionary (for multi-speaker model) | default: None",
)
parser.add_argument(
"-k",
"--key",
type=str,
required=False,
default=0,
help="key changed (number of semitones) | default: 0",
)
parser.add_argument(
"-e",
"--enhance",
type=str,
required=False,
default='true',
help="true or false | default: true",
)
parser.add_argument(
"-pe",
"--pitch_extractor",
type=str,
required=False,
default='crepe',
help="pitch extrator type: parselmouth, dio, harvest, crepe (default)",
)
parser.add_argument(
"-fmin",
"--f0_min",
type=str,
required=False,
default=50,
help="min f0 (Hz) | default: 50",
)
parser.add_argument(
"-fmax",
"--f0_max",
type=str,
required=False,
default=1100,
help="max f0 (Hz) | default: 1100",
)
parser.add_argument(
"-th",
"--threhold",
type=str,
required=False,
default=-60,
help="response threhold (dB) | default: -60",
)
parser.add_argument(
"-eak",
"--enhancer_adaptive_key",
type=str,
required=False,
default=0,
help="adapt the enhancer to a higher vocal range (number of semitones) | default: 0",
)
return parser.parse_args(args=args, namespace=namespace)
def split(audio, sample_rate, hop_size, db_thresh = -40, min_len = 5000):
slicer = Slicer(
sr=sample_rate,
threshold=db_thresh,
min_length=min_len)
chunks = dict(slicer.slice(audio))
result = []
for k, v in chunks.items():
tag = v["split_time"].split(",")
if tag[0] != tag[1]:
start_frame = int(int(tag[0]) // hop_size)
end_frame = int(int(tag[1]) // hop_size)
if end_frame > start_frame:
result.append((
start_frame,
audio[int(start_frame * hop_size) : int(end_frame * hop_size)]))
return result
def cross_fade(a: np.ndarray, b: np.ndarray, idx: int):
result = np.zeros(idx + b.shape[0])
fade_len = a.shape[0] - idx
np.copyto(dst=result[:idx], src=a[:idx])
k = np.linspace(0, 1.0, num=fade_len, endpoint=True)
result[idx: a.shape[0]] = (1 - k) * a[idx:] + k * b[: fade_len]
np.copyto(dst=result[a.shape[0]:], src=b[fade_len:])
return result
if __name__ == '__main__':
#device = 'cpu'
device = 'cuda' if torch.cuda.is_available() else 'cpu'
# parse commands
cmd = parse_args()
# load ddsp model
model, args = load_model(cmd.model_path, device=device)
# load input
audio, sample_rate = librosa.load(cmd.input, sr=None)
if len(audio.shape) > 1:
audio = librosa.to_mono(audio)
hop_size = args.data.block_size * sample_rate / args.data.sampling_rate
# extract f0
print('Pitch extractor type: ' + cmd.pitch_extractor)
pitch_extractor = F0_Extractor(
cmd.pitch_extractor,
sample_rate,
hop_size,
float(cmd.f0_min),
float(cmd.f0_max))
print('Extracting the pitch curve of the input audio...')
f0 = pitch_extractor.extract(audio, uv_interp = True, device = device)
f0 = torch.from_numpy(f0).float().to(device).unsqueeze(-1).unsqueeze(0)
# key change
f0 = f0 * 2 ** (float(cmd.key) / 12)
# extract volume
print('Extracting the volume envelope of the input audio...')
volume_extractor = Volume_Extractor(hop_size)
volume = volume_extractor.extract(audio)
mask = (volume > 10 ** (float(cmd.threhold) / 20)).astype('float')
mask = np.pad(mask, (4, 4), constant_values=(mask[0], mask[-1]))
mask = np.array([np.max(mask[n : n + 9]) for n in range(len(mask) - 8)])
mask = torch.from_numpy(mask).float().to(device).unsqueeze(-1).unsqueeze(0)
mask = upsample(mask, args.data.block_size).squeeze(-1)
volume = torch.from_numpy(volume).float().to(device).unsqueeze(-1).unsqueeze(0)
# load units encoder
units_encoder = Units_Encoder(
args.data.encoder,
args.data.encoder_ckpt,
args.data.encoder_sample_rate,
args.data.encoder_hop_size,
device = device)
# load enhancer
if cmd.enhance == 'true':
print('Enhancer type: ' + args.enhancer.type)
enhancer = Enhancer(args.enhancer.type, args.enhancer.ckpt, device=device)
else:
print('Enhancer type: none (using raw output of ddsp)')
# speaker id or mix-speaker dictionary
spk_mix_dict = literal_eval(cmd.spk_mix_dict)
if spk_mix_dict is not None:
print('Mix-speaker mode')
else:
print('Speaker ID: '+ str(int(cmd.spk_id)))
spk_id = torch.LongTensor(np.array([[int(cmd.spk_id)]])).to(device)
# forward and save the output
result = np.zeros(0)
current_length = 0
segments = split(audio, sample_rate, hop_size)
print('Cut the input audio into ' + str(len(segments)) + ' slices')
with torch.no_grad():
for segment in tqdm(segments):
start_frame = segment[0]
seg_input = torch.from_numpy(segment[1]).float().unsqueeze(0).to(device)
seg_units = units_encoder.encode(seg_input, sample_rate, hop_size)
seg_f0 = f0[:, start_frame : start_frame + seg_units.size(1), :]
seg_volume = volume[:, start_frame : start_frame + seg_units.size(1), :]
seg_output, _, (s_h, s_n) = model(seg_units, seg_f0, seg_volume, spk_id = spk_id, spk_mix_dict = spk_mix_dict)
seg_output *= mask[:, start_frame * args.data.block_size : (start_frame + seg_units.size(1)) * args.data.block_size]
if cmd.enhance == 'true':
seg_output, output_sample_rate = enhancer.enhance(
seg_output,
args.data.sampling_rate,
seg_f0,
args.data.block_size,
adaptive_key = float(cmd.enhancer_adaptive_key))
else:
output_sample_rate = args.data.sampling_rate
seg_output = seg_output.squeeze().cpu().numpy()
silent_length = round(start_frame * args.data.block_size * output_sample_rate / args.data.sampling_rate) - current_length
if silent_length >= 0:
result = np.append(result, np.zeros(silent_length))
result = np.append(result, seg_output)
else:
result = cross_fade(result, seg_output, current_length + silent_length)
current_length = current_length + silent_length + len(seg_output)
sf.write(cmd.output, result, output_sample_rate)
|