def reduce_noise(source_file,
output_file=None,
initial_noise=30,
window_size=0,
noise_threshold=0.15):
logmmse_from_file(source_file,
output_file=(output_file or source_file),
initial_noise=initial_noise,
window_size=window_size,
noise_threshold=noise_threshold)
def prediction(weights_dir, model_name, input_dir, output_dir, sample_rate, frame_length, hop_length_frame, n_fft,
hop_length_fft):
json_file = open(weights_dir + model_name + '.json', 'r')
loaded_model_json = json_file.read()
json_file.close()
loaded_model = model_from_json(loaded_model_json)
loaded_model.load_weights(weights_dir + model_name + '.h5')
print("Loaded model from disk!")
list_files = os.listdir(input_dir)
for file in list_files:
logmmse.logmmse_from_file(input_file=input_dir + file, output_file=output_dir + 'Remastered-' + file)
audio_temp = dt.audio_files_to_file(input_dir, file, sample_rate)
# logmmse.logmmse(data=audio_temp, sampling_rate=32000, output_file=output_dir + 'Remastered-' + file)
# audio_out = nr.reduce_noise(audio_clip=audio_temp, noise_clip=audio_temp, n_fft=n_fft+1, win_length=n_fft+1,
# hop_length=hop_length_fft)
# librosa.output.write_wav(output_dir + 'Remastered-' + file, audio_out, sample_rate)
audio_file = dt.audio_files_to_file(output_dir, 'Remastered-' + file, sample_rate)
# fig, ax = plt.subplots(figsize=(12, 6))
# plt.title('Audio')
# plt.ylabel('Amplitude')
# plt.xlabel('Time(s)')
# ax.plot(audio_temp)
# ax.plot(audio_file, alpha=0.5)
# plt.show()
audio_list = [audio_file]
audio = dt.audio_list_to_numpy(audio_list, frame_length, hop_length_frame)
dim_square_spec = int(n_fft / 2) + 1
m_amp_db_audio, m_pha_audio = dt.audio_numpy_to_matrix_spectrogram(audio, dim_square_spec, n_fft,
hop_length_fft)
x_in = dt.scaled_in(m_amp_db_audio)
x_in = x_in.reshape(x_in.shape[0], x_in.shape[1], x_in.shape[2], 1)
x_pred = loaded_model.predict(x_in)
inv_sca_x_pred = dt.inv_scaled_out(x_pred)
x_denoise = m_amp_db_audio - inv_sca_x_pred[:, :, :, 0]
audio_denoise_recons = dt.matrix_spectrogram_to_numpy_audio(x_denoise, m_pha_audio, frame_length,
hop_length_fft)
nb_samples = audio_denoise_recons.shape[0]
denoise_long = audio_denoise_recons.reshape(1, nb_samples * frame_length) * 10
librosa.output.write_wav(output_dir + 'Final-' + file, denoise_long[0, :], sample_rate)
noise_recons = dt.matrix_spectrogram_to_numpy_audio(inv_sca_x_pred[:, :, :, 0], m_pha_audio, frame_length,
hop_length_fft)
nb_samples = noise_recons.shape[0]
noise_long = noise_recons.reshape(1, nb_samples * frame_length)
librosa.output.write_wav(output_dir + 'Noise-' + file, noise_long[0, :], sample_rate)
def get_command_recording():
_, file_name = tempfile.mkstemp(".wav", prefix="input_command")
wave_file = wave.open(file_name, "wb")
wave_file.setnchannels(1)
wave_file.setsampwidth(pa.get_sample_size(pyaudio.paInt16))
wave_file.setframerate(sample_rate)
nframes = int(sample_rate / frames_per_buffer * DURATION)
for _ in range(nframes):
audio = audio_stream.read(frames_per_buffer)
wave_file.writeframes(audio)
wave_file.close()
logmmse_from_file(file_name, file_name)
return file_name
def reconstruct_audio(signal_filename, output_filename, low_cutoff,
high_cutoff):
data = np.load(signal_filename)
arr = data['arr_0'][:, ::2, 1:].astype(np.float64)
arr = arr.reshape(-1, arr.shape[2])
filter = lambda x: signal.sosfilt(
signal.butter(11, (low_cutoff - 50, high_cutoff + 50),
'bandpass',
fs=2200,
output='sos'), x)
combined_signal = combine(arr, 4, filter)
sos = signal.butter(11, (low_cutoff, high_cutoff),
'bandpass',
fs=2200,
output='sos')
filtered = signal.sosfilt(sos, combined_signal)
wav_res = np.clip((32767 * filtered / np.max(filtered)).astype('i2'),
-32767, 32767)
wavfile.write(output_filename, 2200, wav_res)
denoised_filename = output_filename[:output_filename.
rindex('.')] + '-denoised.wav'
output = logmmse_from_file(output_filename)
wavfile.write(denoised_filename, 2200, output)
def reducenoise():
logmmse_from_file(voicewave,
output_file=voicerd,
initial_noise=60,
window_size=0,
noise_threshold=0.15)
for j in multipliers:
def get_audio():
return randomizer.get_noisy_speech(_set='dev')
set_snr(j[0])
for i in range(0, EPOCHS):
hm, bg, mfcc_feature = generate_sample(_n_filt=NFILT,
_winlen=WINLEN,
_winstep=WINSTEP,
_winfunc=np.hamming,
_generator=get_audio,
to_write=True)
logmmse_from_file(
os.path.join(os.path.dirname(__file__),
'../sample_test/mixed.wav'),
os.path.join(os.path.dirname(__file__),
'../sample_test/logmmse.wav'))
sequence_feature = np.array(
to_sequence_with_stride(mfcc_feature,
left_pad=PAD_L,
right_pad=PAD_R))
# sys model
prediction = predict(sequence_feature, 'sys_model')
estimate = get_estimate(bg,
prediction,
_winlen=WINLEN,
_winstep=WINSTEP,
_winfunc=np.hamming)
write_wav('sys_model.wav', estimate.astype(np.float32))
evaluate(estimated_sources=['sys_model.wav', 'logmmse.wav'])