app.py

import librosa
import tensorflow as tf
from tensorflow.keras.models import model_from_json
import soundfile as sf
import numpy as np
import os
import scipy
from scipy.io import wavfile
import gradio as gr

def audio_to_audio_frame_stack(sound_data, frame_length, hop_length_frame):
    """This function take an audio and split into several frame
       in a numpy matrix of size (nb_frame,frame_length)"""

    sequence_sample_length = sound_data.shape[0]

    sound_data_list = [sound_data[start:start + frame_length] for start in range(
    0, sequence_sample_length - frame_length + 1, hop_length_frame)]  # get sliding windows
    sound_data_array = np.vstack(sound_data_list)

    return sound_data_array


def audio_files_to_numpy(audio_dir, list_audio_files, sample_rate, frame_length, hop_length_frame, min_duration):
    """This function take audio files of a directory and merge them
    in a numpy matrix of size (nb_frame,frame_length) for a sliding window of size hop_length_frame"""

    list_sound_array = []

    for file in list_audio_files:
        # open the audio file
        y, sr = librosa.load(os.path.join(audio_dir, file), sr=sample_rate)
        total_duration = librosa.get_duration(y=y, sr=sr)

        if (total_duration >= min_duration):
            list_sound_array.append(audio_to_audio_frame_stack(
                y, frame_length, hop_length_frame))
        else:
            print(
                f"The following file {os.path.join(audio_dir,file)} is below the min duration")

    return np.vstack(list_sound_array)


def blend_noise_randomly(voice, noise, nb_samples, frame_length):
    """This function takes as input numpy arrays representing frames
    of voice sounds, noise sounds and the number of frames to be created
    and return numpy arrays with voice randomly blend with noise"""

    prod_voice = np.zeros((nb_samples, frame_length))
    prod_noise = np.zeros((nb_samples, frame_length))
    prod_noisy_voice = np.zeros((nb_samples, frame_length))

    for i in range(nb_samples):
        id_voice = np.random.randint(0, voice.shape[0])
        id_noise = np.random.randint(0, noise.shape[0])
        level_noise = np.random.uniform(0.2, 0.8)
        prod_voice[i, :] = voice[id_voice, :]
        prod_noise[i, :] = level_noise * noise[id_noise, :]
        prod_noisy_voice[i, :] = prod_voice[i, :] + prod_noise[i, :]

    return prod_voice, prod_noise, prod_noisy_voice


def audio_to_magnitude_db_and_phase(n_fft, hop_length_fft, audio):
    """This function takes an audio and convert into spectrogram,
       it returns the magnitude in dB and the phase"""

    stftaudio = librosa.stft(audio, n_fft=n_fft, hop_length=hop_length_fft)
    stftaudio_magnitude, stftaudio_phase = librosa.magphase(stftaudio)

    stftaudio_magnitude_db = librosa.amplitude_to_db(
        stftaudio_magnitude, ref=np.max)

    return stftaudio_magnitude_db, stftaudio_phase


def numpy_audio_to_matrix_spectrogram(numpy_audio, dim_square_spec, n_fft, hop_length_fft):
    """This function takes as input a numpy audi of size (nb_frame,frame_length), and return
    a numpy containing the matrix spectrogram for amplitude in dB and phase. It will have the size
    (nb_frame,dim_square_spec,dim_square_spec)"""

    nb_audio = numpy_audio.shape[0]

    m_mag_db = np.zeros((nb_audio, dim_square_spec, dim_square_spec))
    m_phase = np.zeros((nb_audio, dim_square_spec, dim_square_spec), dtype=complex)

    for i in range(nb_audio):
        m_mag_db[i, :, :], m_phase[i, :, :] = audio_to_magnitude_db_and_phase(
            n_fft, hop_length_fft, numpy_audio[i])

    return m_mag_db, m_phase


def magnitude_db_and_phase_to_audio(frame_length, hop_length_fft, stftaudio_magnitude_db, stftaudio_phase):
    """This functions reverts a spectrogram to an audio"""

    stftaudio_magnitude_rev = librosa.db_to_amplitude(stftaudio_magnitude_db, ref=1.0)

    # taking magnitude and phase of audio
    audio_reverse_stft = stftaudio_magnitude_rev * stftaudio_phase
    audio_reconstruct = librosa.core.istft(audio_reverse_stft, hop_length=hop_length_fft, length=frame_length)

    return audio_reconstruct

def matrix_spectrogram_to_numpy_audio(m_mag_db, m_phase, frame_length, hop_length_fft)  :
    """This functions reverts the matrix spectrograms to numpy audio"""

    list_audio = []

    nb_spec = m_mag_db.shape[0]

    for i in range(nb_spec):

        audio_reconstruct = magnitude_db_and_phase_to_audio(frame_length, hop_length_fft, m_mag_db[i], m_phase[i])
        list_audio.append(audio_reconstruct)

    return np.vstack(list_audio)

def scaled_in(matrix_spec):
    "global scaling apply to noisy voice spectrograms (scale between -1 and 1)"
    matrix_spec = (matrix_spec + 46)/50
    return matrix_spec

def scaled_ou(matrix_spec):
    "global scaling apply to noise models spectrograms (scale between -1 and 1)"
    matrix_spec = (matrix_spec -6 )/82
    return matrix_spec

def inv_scaled_in(matrix_spec):
    "inverse global scaling apply to noisy voices spectrograms"
    matrix_spec = matrix_spec * 50 - 46
    return matrix_spec

def inv_scaled_ou(matrix_spec):
    "inverse global scaling apply to noise models spectrograms"
    matrix_spec = matrix_spec * 82 + 6
    return matrix_spec


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)
    # audio = audioData
    #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
    # librosa.output.write_wav(dir_save_prediction + audio_output_prediction, denoise_long[0, :], sample_rate)
    print(dir_save_prediction + audio_output_prediction)
    sf.write(dir_save_prediction + audio_output_prediction , denoise_long[0, :], sample_rate)

def denoise_audio(audioName):
  sr, data = audioName
  sf.write("temp.wav",data, 44100)
  testNo = "temp"
  audio_dir_prediction = os.path.abspath("/")+ str(testNo) +".wav"
  sample_rate, data = audioName[0], audioName[1]
  len_data = len(data)  # holds length of the numpy array

  
  t = len_data / sample_rate # returns duration but in floats
  print("t:",t)
  weights_path = os.path.abspath("./")
  name_model = "model_unet"
  audio_dir_prediction = os.path.abspath("./")
  dir_save_prediction = os.path.abspath("./")
  audio_output_prediction = "test.wav"
  audio_input_prediction = ["temp.wav"]

  min_duration = t
  frame_length = 8064
  hop_length_frame = 8064
  n_fft = 255
  hop_length_fft = 63

  dim_square_spec = int(n_fft / 2) + 1

  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)
  print(audio_output_prediction)
  return audio_output_prediction


examples = [
    [os.path.abspath("3.wav")],
    [os.path.abspath("2.wav")]
]



iface = gr.Interface(fn = denoise_audio,
                     inputs = 'audio',
                     outputs = 'audio', 
                     verbose = True,
                     title = 'audio to denoised Audio Application',
                     description = 'A simple application to denoise audio speech usinf UNet deep learning model. Upload your own audio, or click one of the examples to load them.',
                     article = 
                        '''<div>
                            <p style="text-align: center"> All you need to do is to upload the pdf file and hit submit, then wait for compiling. After that click on Play/Pause for listing to the audio. The audio is saved in a wav format.</p>
                        </div>''',
                     examples=examples
                    )

iface.launch()