def addFx(sound, effects, pad=3000, fade_in=100, fade_out=100):
# Add padding
if pad > 0:
sound += AudioSegment.silent(duration=pad, frame_rate=sound.frame_rate)
# convert pydub sound to np array
samples = np.array(sound.get_array_of_samples())
samples = samples.astype(np.int16)
chain = AudioEffectsChain()
for effect, value in effects:
if effect == "reverb" and value > 0:
chain.reverb(reverberance=value)
elif effect == "distortion" and value > 0:
chain.overdrive(gain=value)
elif effect == "highpass" and value > 0:
chain.highpass(value)
elif effect == "lowpass" and value > 0:
chain.lowpass(value)
elif effect == "bass":
frequency = 100
gain = value
if isinstance(value, tuple):
gain, frequency = value
print("%s, %s" % (gain, frequency))
chain.highshelf(gain=gain, frequency=frequency)
elif effect == "echo":
echoStr = "echo 0.8 0.9"
amount = value
count = 1
# check if we have echo count indicated
if isinstance(value, tuple):
amount, count = value
for i in range(count):
# amount between 10 (robot) and 1000 (mountains)
echoStr += " %s 0.3" % amount
chain.custom(echoStr)
elif effect == "tempo" and value != 1.0 and value != 1:
chain.tempo(factor=value)
# apply reverb effect
fx = (chain)
y = fx(samples)
# convert it back to an array and create a new sound clip
newData = array.array(sound.array_type, y)
newSound = sound._spawn(newData)
dur = len(newSound)
newSound = newSound.fade_in(min(fade_in, dur)).fade_out(min(fade_out, dur))
return newSound
def worker(_):
sig = np.random.standard_normal(44100 * 8)
effect = AudioEffectsChain()
effect = effect.pitch(np.random.uniform(-300, 300))
effect = effect.tempo(np.random.uniform(0.8, 1.2))
effect = effect.reverb(np.random.uniform(0, 100))
return effect(sig)
def __call__(self, input):
effect = AudioEffectsChain()
if np.random.uniform() > 0.5:
effect = effect.pitch(np.random.uniform(-300, 300))
if np.random.uniform() > 0.5:
effect = effect.tempo(np.random.uniform(0.8, 1.2))
if np.random.uniform() > 0.5:
effect = effect.reverb(np.random.uniform(0, 100))
# if np.random.uniform() > 0.5:
# effect = effect.overdrive(np.random.uniform(0, 10))
# if np.random.uniform() > 0.5:
# effect = effect.limiter(np.random.uniform(-10, 10))
# if np.random.uniform() > 0.5:
# effect = effect.lowshelf()
# if np.random.uniform() > 0.5:
# effect = effect.highshelf()
return effect(input)
def __call__(self, wav_file):
if not Path(wav_file).exists():
print(wav_file)
raise IOError
sr, wav = scipy.io.wavfile.read(wav_file)
if wav.ndim > 1 and wav.shape[1] > 1:
logger.error("wav file has two or more channels")
sys.exit(1)
if type(wav[0]) is np.int32:
wav = wav.astype('float32', copy=False) / 2147483648.0
elif type(wav[0]) is np.int16:
wav = wav.astype('float32', copy=False) / 32768.0
elif type(wav[0]) is np.uint8:
wav = wav.astype('float32', copy=False) / 256.0 - 128.0
fx = AudioEffectsChain()
if self.resample:
if self.sample_rate > sr:
ratio = int(self.sample_rate / sr)
fx.upsample(ratio)
elif self.sample_rate < sr:
ratio = int(sr / self.sample_rate)
fx.custom(f"downsample {ratio}")
if self.tempo:
tempo_change = np.random.uniform(*self.tempo_range)
fx.tempo(tempo_change, opt_flag="s")
if self.pitch:
pitch_change = np.random.uniform(*self.pitch_range)
fx.pitch(pitch_change)
# dithering
fx.custom(f"dither -s")
wav = fx(wav, sample_in=sr, sample_out=self.sample_rate)
#wav = wav / max(abs(wav))
# normalize audio power
gain = 0.1
wav_energy = np.sqrt(np.sum(np.power(wav, 2)) / wav.size)
wav = gain * wav / wav_energy
# sample-domain padding
if self.padding:
wav = np.pad(wav, self.num_padding, mode='constant')
# sample-domain offset
if self.offset:
offset = np.random.randint(*self.offset_range)
wav = np.roll(wav, offset, axis=0)
if self.noise:
snr = 10.0**(np.random.uniform(*self.noise_range) / 10.0)
noise = np.random.normal(0, 1, wav.shape)
noise_energy = np.sqrt(np.sum(np.power(noise, 2)) / noise.size)
wav = wav + snr * gain * noise / noise_energy
#filename = wav_file.replace(".wav", "_augmented.wav")
#scipy.io.wavfile.write(filename, self.sample_rate, wav)
return torch.FloatTensor(wav)