Add_noise_and_rir_to_speech

The purpose of this code base is to add a specified signal-to-noise ratio noise from MUSAN dataset to a pure speech signal and to generate far-field speech data using room impulse response data from BUT Speech@FIT Reverb Database.

Noise and RIR dataset description:

• BUT Speech@FIT Reverb Database:

  The database is being built with respect to collect a large number of various Room Impulse Responses, Room environmental noises (or "silences"), Retransmitted speech (for ASR and SID testing), and meta-data (positions of microphones, speakers etc.).

  The goal is to provide speech community with a dataset for data enhancement and distant microphone or microphone array experiments in ASR and SID.

  In this codebase, we only use the RIR data, which is used to synthesize far-field speech, the composition of the RIR dataset and citation details are as follows.

  Room Name	Room Type	Size (length, depth, height) (m)	(microphone_num x loudspeaker_num)
  Q301	Office	10.7x6.9x2.6	31 x 3
  L207	Office	4.6x6.9x3.1	31 x 6
  L212	Office	7.5x4.6x3.1	31 x 5
  L227	Stairs	6.2x2.6x14.2	31 x 5
  R112	Hotel room	4.4x2.8x2.6	31 x 5
  CR2	Conference room	28.2x11.1x3.3	31 x 4
  E112	Lecture room	11.5x20.1x4.8	31 x 2
  D105	Lecture room	17.2x22.8x6.9	31 x 6
  C236	Meeting room	7.0x4.1x3.6	31 x 10

  @ARTICLE{8717722,
           author={Szöke, Igor and Skácel, Miroslav and Mošner, Ladislav and Paliesek, Jakub and Černocký, Jan},
           journal={IEEE Journal of Selected Topics in Signal Processing}, 
           title={Building and evaluation of a real room impulse response dataset}, 
           year={2019},
           volume={13},
           number={4},
           pages={863-876},
           doi={10.1109/JSTSP.2019.2917582}
   }
  

• MUSAN database:

  The database consists of music from several genres, speech from twelve languages, and a wide assortment of technical and non-technical noises and we only use the noise data in this database. Citation details are as follows.

  @misc{snyder2015musan,
        title={MUSAN: A Music, Speech, and Noise Corpus}, 
        author={David Snyder and Guoguo Chen and Daniel Povey},
        year={2015},
        eprint={1510.08484},
        archivePrefix={arXiv},
        primaryClass={cs.SD}
  }
  

Before using the data-processing code:

• If you do not want the original dataset to be overwritten, please download the dataset again for use

• You need to create three files: 'training_list.txt', 'validation_list.txt', 'testing_list.txt', based on your training, validation and test data file paths respectively, and ensure the audio in the file paths can be read and written.

• The content of the aforementioned '*_list.txt' files are in the following form:

  *_list.txt
  	/../...../*.wav
  	/../...../*.wav
  	/../...../*.wav
  

Instruction for using the following data-processing code:

1. mix_cleanaudio_with_rir_offline.py: Generate far-field speech offline

   • two parameters are needed:

     • --data_root: the data path which you want to download and store the RIR dataset in.
     • --clean_data_list_path: the path of the folder in which 'training_list.txt', 'validation_list.txt', 'testing_list.txt' are stored in

   • 2 folders will be created in data_root: 'ReverDB_data (Removable if needed)', 'ReverDB_mix'

2. download_and_extract_noise_file.py: Generate musan noise file

   • one parameters are needed:
     • --data_root: the data path which you want to download and store the noise dataset in.
   • 2 folder will be created in data_root: 'musan (Removable if needed)', 'noise'

3. vad_torch.py: Voice activity detection when adding noise to the speech

   The noise data is usually added online according to the SNR requirements, several pieces of code are provided below, please add them in the appropriate places according to your needs!

   import torchaudio
   import numpy as np
   import torch
   import random
   from vad_torch import VoiceActivityDetector
   
   
   def _add_noise(speech_sig, vad_duration, noise_sig, snr):
       """add noise to the audio.
       :param speech_sig: The input audio signal (Tensor).
       :param vad_duration: The length of the human voice (int).
       :param noise_sig: The input noise signal (Tensor).
       :param snr: the SNR you want to add (int).
       :returns: noisy speech sig with specific snr.
       """
       if vad_duration != 0:
           snr = 10**(snr/10.0)
           speech_power = torch.sum(speech_sig**2)/vad_duration
           noise_power = torch.sum(noise_sig**2)/noise_sig.shape[1]
           noise_update = noise_sig / torch.sqrt(snr * noise_power/speech_power)
   
           if speech_sig.shape[1] > noise_update.shape[1]:
               # padding
               temp_wav = torch.zeros(1, speech_sig.shape[1])
               temp_wav[0, 0:noise_update.shape[1]] = noise_update
               noise_update = temp_wav
           else:
               # cutting
               noise_update = noise_update[0, 0:speech_sig.shape[1]]
   
           return noise_update + speech_sig
       
       else:
           return speech_sig
       
   def main():
       # loading speech file
       speech_file = './speech.wav'
   	waveform, sr = torchaudio.load(speech_file)
   	waveform = waveform - waveform.mean()
   	
       # loading noise file and set snr
   	snr = 0       
   	noise_file = random.randint(1, 930)
   	
       # Voice activity detection
   	v = VoiceActivityDetector(waveform, sr)
   	raw_detection = v.detect_speech()
   	speech_labels = v.convert_windows_to_readible_labels(raw_detection)
   	vad_duration = 0
       if not len(speech_labels) == 0:
           for i in range(len(speech_labels)):
               start = speech_labels[i]['speech_begin']
               end = speech_labels[i]['speech_end']
               vad_duration = vad_duration + end-start
               
   	# adding noise
       noise, _ = torchaudio.load('/notebooks/noise/' + str(noise_file) + '.wav')
       waveform = _add_noise(waveform, vad_duration, noise, snr)
   
   if __name__ == '__main__':
       main()