DeepLearningExamples/PyTorch/SpeechRecognition/Jasper/parts/perturb.py
Przemek Strzelczyk fa400a7367 Adding Jasper/PyT
2019-07-26 20:08:16 +02:00

112 lines
3.9 KiB
Python

# Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import random
import librosa
from .manifest import Manifest
from .segment import AudioSegment
class Perturbation(object):
def max_augmentation_length(self, length):
return length
def perturb(self, data):
raise NotImplementedError
class SpeedPerturbation(Perturbation):
def __init__(self, min_speed_rate=0.85, max_speed_rate=1.15, rng=None):
self._min_rate = min_speed_rate
self._max_rate = max_speed_rate
self._rng = random.Random() if rng is None else rng
def max_augmentation_length(self, length):
return length * self._max_rate
def perturb(self, data):
speed_rate = self._rng.uniform(self._min_rate, self._max_rate)
if speed_rate <= 0:
raise ValueError("speed_rate should be greater than zero.")
data._samples = librosa.effects.time_stretch(data._samples, speed_rate)
class GainPerturbation(Perturbation):
def __init__(self, min_gain_dbfs=-10, max_gain_dbfs=10, rng=None):
self._min_gain_dbfs = min_gain_dbfs
self._max_gain_dbfs = max_gain_dbfs
self._rng = random.Random() if rng is None else rng
def perturb(self, data):
gain = self._rng.uniform(self._min_gain_dbfs, self._max_gain_dbfs)
data._samples = data._samples * (10. ** (gain / 20.))
class ShiftPerturbation(Perturbation):
def __init__(self, min_shift_ms=-5.0, max_shift_ms=5.0, rng=None):
self._min_shift_ms = min_shift_ms
self._max_shift_ms = max_shift_ms
self._rng = random.Random() if rng is None else rng
def perturb(self, data):
shift_ms = self._rng.uniform(self._min_shift_ms, self._max_shift_ms)
if abs(shift_ms) / 1000 > data.duration:
# TODO: do something smarter than just ignore this condition
return
shift_samples = int(shift_ms * data.sample_rate // 1000)
# print("DEBUG: shift:", shift_samples)
if shift_samples < 0:
data._samples[-shift_samples:] = data._samples[:shift_samples]
data._samples[:-shift_samples] = 0
elif shift_samples > 0:
data._samples[:-shift_samples] = data._samples[shift_samples:]
data._samples[-shift_samples:] = 0
perturbation_types = {
"speed": SpeedPerturbation,
"gain": GainPerturbation,
"shift": ShiftPerturbation,
}
class AudioAugmentor(object):
def __init__(self, perturbations=None, rng=None):
self._rng = random.Random() if rng is None else rng
self._pipeline = perturbations if perturbations is not None else []
def perturb(self, segment):
for (prob, p) in self._pipeline:
if self._rng.random() < prob:
p.perturb(segment)
return
def max_augmentation_length(self, length):
newlen = length
for (prob, p) in self._pipeline:
newlen = p.max_augmentation_length(newlen)
return newlen
@classmethod
def from_config(cls, config):
ptbs = []
for p in config:
if p['aug_type'] not in perturbation_types:
print(p['aug_type'], "perturbation not known. Skipping.")
continue
perturbation = perturbation_types[p['aug_type']]
ptbs.append((p['prob'], perturbation(**p['cfg'])))
return cls(perturbations=ptbs)