def prediction(weights_path, name_model, audio_dir_prediction,
dir_save_prediction, audio_input_prediction,
audio_output_prediction, sample_rate, min_duration,
frame_length, hop_length_frame, n_fft, hop_length_fft):
""" This function takes as input pretrained weights, noisy voice sound to denoise, predict
the denoise sound and save it to disk.
"""
# load json and create model
json_file = open(weights_path + '/' + name_model + '.json', 'r')
loaded_model_json = json_file.read()
json_file.close()
loaded_model = model_from_json(loaded_model_json)
# load weights into new model
loaded_model.load_weights(weights_path + '/' + name_model + '.h5')
print("Loaded model from disk")
# Extracting noise and voice from folder and convert to numpy
audio = audio_files_to_numpy(audio_dir_prediction, audio_input_prediction,
sample_rate, frame_length, hop_length_frame,
min_duration)
#Dimensions of squared spectrogram
dim_square_spec = int(n_fft / 2) + 1
print(dim_square_spec)
# Create Amplitude and phase of the sounds
m_amp_db_audio, m_pha_audio = numpy_audio_to_matrix_spectrogram(
audio, dim_square_spec, n_fft, hop_length_fft)
#global scaling to have distribution -1/1
X_in = scaled_in(m_amp_db_audio)
#Reshape for prediction
X_in = X_in.reshape(X_in.shape[0], X_in.shape[1], X_in.shape[2], 1)
#Prediction using loaded network
X_pred = loaded_model.predict(X_in)
#Rescale back the noise model
inv_sca_X_pred = inv_scaled_ou(X_pred)
#Remove noise model from noisy speech
X_denoise = m_amp_db_audio - inv_sca_X_pred[:, :, :, 0]
#Reconstruct audio from denoised spectrogram and phase
print(X_denoise.shape)
print(m_pha_audio.shape)
print(frame_length)
print(hop_length_fft)
audio_denoise_recons = matrix_spectrogram_to_numpy_audio(
X_denoise, m_pha_audio, frame_length, hop_length_fft)
#Number of frames
nb_samples = audio_denoise_recons.shape[0]
#Save all frames in one file
denoise_long = audio_denoise_recons.reshape(1,
nb_samples * frame_length) * 10
# Update of Librosa no longer uses .output use sf.write instead
#librosa.output.write_wav(dir_save_prediction + audio_output_prediction, denoise_long[0, :], sample_rate)
import soundfile as sf
sf.write(dir_save_prediction + audio_output_prediction, denoise_long[0, :],
sample_rate, 'PCM_16')
def prediction(weights_path, model, audio_input_prediction, sample_rate,
min_duration, frame_length, hop_length_frame, n_fft,
hop_length_fft):
""" This function takes as input pretrained weights, noisy voice sound to denoise, predict
the denoise sound and save it to disk.
"""
loaded_model = model
# load weights into new model
loaded_model.load_weights(weights_path + '/' + 'model_best.h5')
print("Loaded model from disk")
audio_dir_prediction = ""
# Extracting noise and voice from folder and convert to numpy
audio = audio_files_to_numpy(audio_dir_prediction, audio_input_prediction,
sample_rate, frame_length, hop_length_frame,
min_duration)
#Dimensions of squared spectrogram
dim_square_spec = int(n_fft / 2) + 1
print(dim_square_spec)
# Create Amplitude and phase of the sounds
m_amp_db_audio, m_pha_audio = numpy_audio_to_matrix_spectrogram(
audio, dim_square_spec, n_fft, hop_length_fft)
#global scaling to have distribution -1/1
X_in = scaled_in(m_amp_db_audio)
#Reshape for prediction
X_in = X_in.reshape(X_in.shape[0], X_in.shape[1], X_in.shape[2], 1)
#Prediction using loaded network
X_pred = loaded_model.predict(X_in)
#Rescale back the noise model
inv_sca_X_pred = inv_scaled_ou(X_pred)
#Remove noise model from noisy speech
X_denoise = m_amp_db_audio - inv_sca_X_pred[:, :, :, 0]
#Reconstruct audio from denoised spectrogram and phase
print(X_denoise.shape)
print(m_pha_audio.shape)
print(frame_length)
print(hop_length_fft)
audio_denoise_recons = matrix_spectrogram_to_numpy_audio(
X_denoise, m_pha_audio, frame_length, hop_length_fft)
#Number of frames
nb_samples = audio_denoise_recons.shape[0]
#Save all frames in one file
denoise_long = audio_denoise_recons.reshape(1,
nb_samples * frame_length) * 10
return [denoise_long[0, :], sample_rate]
def decode(weights_path=args.weights_folder,
name_model=args.name_model,
audio_dir_prediction=args.audio_dir_prediction,
dir_save_prediction=args.dir_save_prediction,
audio_input_prediction=args.audio_input_prediction,
audio_output_prediction=args.audio_output_prediction,
sample_rate=args.sample_rate,
min_duration=args.min_duration,
frame_length=args.frame_length,
hop_length_frame=args.hop_length_frame,
n_fft=args.n_fft,
hop_length_fft=args.hop_length_fft):
loaded_model = ConvAutoEncoder(weights_path=weights_path)
loaded_model.load_weights()
loaded_model.info()
print("Loaded model from disk")
audio = audio_files_to_numpy(audio_dir_prediction, audio_input_prediction,
sample_rate, frame_length, hop_length_frame,
min_duration)
#Dimensions of squared spectrogram
dim_square_spec = int(n_fft / 2) + 1
# Create Amplitude and phase of the sounds
m_amp_db_audio, m_pha_audio = numpy_audio_to_matrix_spectrogram(
audio, dim_square_spec, n_fft, hop_length_fft)
data_compress = np.load('aaa.npy')
print(data_compress.shape)
decoded = loaded_model.decode(data_compress)
#Rescale back the noise model
inv_sca_X_pred = inv_scaled_ou(decoded)
#Remove noise model from noisy speech
X_denoise = m_amp_db_audio - inv_sca_X_pred[:, :, :, 0]
#Reconstruct audio from denoised spectrogram and phase
print(X_denoise.shape)
print(m_pha_audio.shape)
print(frame_length)
print(hop_length_fft)
audio_denoise_recons = matrix_spectrogram_to_numpy_audio(
X_denoise, m_pha_audio, frame_length, hop_length_fft)
#Number of frames
nb_samples = audio_denoise_recons.shape[0]
#Save all frames in one file
denoise_long = audio_denoise_recons.reshape(1,
nb_samples * frame_length) * 10
librosa.output.write_wav(dir_save_prediction + audio_output_prediction,
denoise_long[0, :], sample_rate)
print('saved audio decoded file in:',
dir_save_prediction + audio_output_prediction)
hop_length_fft = 63
dim_square_spec = int(n_fft / 2) + 1
# Create Amplitude and phase of the sounds
m_amp_db_audio, m_pha_audio = numpy_audio_to_matrix_spectrogram(
audio, dim_square_spec, n_fft, hop_length_fft)
#global scaling to have distribution -1/1
X_in = scaled_in(m_amp_db_audio)
#Reshape for prediction
X_in = X_in.reshape(X_in.shape[0], X_in.shape[1], X_in.shape[2], 1)
#Prediction using loaded network
X_pred = loaded_model.predict(X_in)
#Rescale back the noise model
inv_sca_X_pred = inv_scaled_ou(X_pred)
#Remove noise model from noisy speech
X_denoise = m_amp_db_audio - inv_sca_X_pred[:, :, :, 0]
#Reconstruct audio from denoised spectrogram and phase
audio_denoise_recons = matrix_spectrogram_to_numpy_audio(
X_denoise, m_pha_audio, frame_length, hop_length_fft)
#Number of frames
nb_samples = audio_denoise_recons.shape[0]
def test_dimensions_spectrogram():
""" test that dimensions are correct"""
assert dim_square_spec == 128
assert dim_square_spec == m_amp_db_audio.shape[1]
assert dim_square_spec == m_amp_db_audio.shape[2]
assert dim_square_spec == X_denoise.shape[1]