File size: 2,334 Bytes
88b5dc0 |
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 |
import logging
import random
from dataclasses import dataclass
from functools import cached_property
from pathlib import Path
import librosa
import numpy as np
from scipy import signal
from ..utils import walk_paths
from .base import Effect
_logger = logging.getLogger(__name__)
@dataclass
class RandomRIR(Effect):
rir_dir: Path | None
rir_rate: int = 44_000
rir_suffix: str = ".npy"
deterministic: bool = False
@cached_property
def rir_paths(self):
if self.rir_dir is None:
return []
return list(walk_paths(self.rir_dir, self.rir_suffix))
def _sample_rir(self):
if len(self.rir_paths) == 0:
return None
if self.deterministic:
rir_path = self.rir_paths[0]
else:
rir_path = random.choice(self.rir_paths)
rir = np.squeeze(np.load(rir_path))
assert isinstance(rir, np.ndarray)
return rir
def apply(self, wav, sr):
# ref: https://github.com/haoheliu/voicefixer_main/blob/b06e07c945ac1d309b8a57ddcd599ca376b98cd9/dataloaders/augmentation/magical_effects.py#L158
if len(self.rir_paths) == 0:
return wav
length = len(wav)
wav = librosa.resample(wav, orig_sr=sr, target_sr=self.rir_rate, res_type="kaiser_fast")
rir = self._sample_rir()
wav = signal.convolve(wav, rir, mode="same")
actlev = np.max(np.abs(wav))
if actlev > 0.99:
wav = (wav / actlev) * 0.98
wav = librosa.resample(wav, orig_sr=self.rir_rate, target_sr=sr, res_type="kaiser_fast")
if abs(length - len(wav)) > 10:
_logger.warning(f"length mismatch: {length} vs {len(wav)}")
if length > len(wav):
wav = np.pad(wav, (0, length - len(wav)))
elif length < len(wav):
wav = wav[:length]
return wav
class RandomGaussianNoise(Effect):
def __init__(self, alpha_range=(0.8, 1)):
super().__init__()
self.alpha_range = alpha_range
def apply(self, wav, sr):
noise = np.random.randn(*wav.shape)
noise_energy = np.sum(noise**2)
wav_energy = np.sum(wav**2)
noise = noise * np.sqrt(wav_energy / noise_energy)
alpha = random.uniform(*self.alpha_range)
return wav * alpha + noise * (1 - alpha)
|