← Back to research group page
← Back to personal page
This page contains audio samples and source code related to the following work:
Heikki Kallasjoki, Jort F. Gemmeke, Kalle J. Palomäki, Amy V. Beeston, Guy J. Brown. "Recognition of reverberant speech by missing data imputation and NMF feature enhancement". Proceedings of the REVERB Challenge Workshop, Florence, 2014.
Note that the algorithm has been designed and evaluated with speech recognition performance (rather than speech enhancement) as the primary benchmark. The method used for the generation of audio samples based on the enhanced features may also add its own artefacts, which would not be present when the features are used directly for speech recognition.
The following samples have been generated using public domain audiobook readings from the LibriVox project, artificially convoluted with multi-condition training data impulse responses distributed as part of the tools of the REVERB challenge. For detailed information, see below.
Column "Clean" contains the original clean speech recording, while column "Rev." refers to its convolution with the corresponding room impulse response. Columns "MD" and "NMF" contain the end result of filtering the reverberated signal based on the enhanced features from the missing data and NMF-based approaches, respectively.
|Male, 1||Large, far||.wav||.wav||.wav||.wav|
|Male, 2||Large, far||.wav||.wav||.wav||.wav|
|Female, 1||Large, far||.wav||.wav||.wav||.wav|
|Female, 2||Large, far||.wav||.wav||.wav||.wav|
Source code for the missing data and NMF-based feature enhancement systems can be found in: kallasjoki-reverb14.zip. The archive contains the following functions:
The source code is made available under the following BSD license:
Copyright (c) 2014, Heikki Kallasjoki, Kalle Palomäki, Ulpu Remes and Sami Keronen
All rights reserved.
Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
- Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
- Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
The audio samples are based on the LibriVox reading of The Principles of Economics with Applications to Practical Problems, Frank Albert Fetter, 1905. The male voice samples are extracts from Chapter 18, read by user "PDyer". The female voice samples are extracts from Chapter 21, read by user "Marian Cervassi". The imputation GMM model and the NMF dictionary was constructed from the entirety of chapters 12, 23 and 28, read by (different) users "Ron Lockhart", "Andrew Nelson" and "Olenka", respectively.
The room impulse responses used in all cases were the responses of channel 1 of array 1 in the corresponding room size and microphone distance setting.
Feature enhancement for the audio samples was performed with methods essentially equivalent to those presented in the paper, with the exception of using a mel filterbank of higher spectral resolution (60 channels). Other parameters of the algorithms were unmodified.
After feature enhancement, the enhanced and original features were used to construct a per-frame mel-domain Wiener filter, which was then frequency-warped back to the linear FFT domain, and applied to the magnitudes of STFT spectra extracted from the original audio signal, while keeping the phase information unchanged. Overlap-add averaging of the IFFT results of the modified spectra was used to construct the final sample waveforms.
Mean cepstrum distances (as defined by ) between the original clean speech and the reverberated/enhanced samples for the above files are shown in the table below:
|Male, 1||Large, far||6.144||5.980||6.016|
|Male, 2||Large, far||5.456||5.115||5.223|
|Female, 1||Large, far||2.501||2.363||2.146|
|Female, 2||Large, far||2.538||2.404||2.118|
 Kalle J. Palomäki, Guy J. Brown, Jon P. Barker. "Recognition of reverberant speech using full cepstral features and spectral missing data," Proc. ICASSP, 2006.
 Ulpu Remes. "Bounded conditional mean imputation with an approximate posterior," Proc. Interspeech, 2013.
 Kalle J. Palomäki, Guy J. Brown, Jon P. Barker. "Techniques for handling convolutional distortion with 'missing data' automatic speech recognition," Speech Communication, 43(1–2), pp. 123–142, 2004.
 Yi Hu, P.C. Loizou. "Evaluation of Objective Quality Measures for Speech Enhancement," IEEE Transactions on Audio, Speech, and Language Processing, 16(1), pp. 229–238, Jan. 2008, doi:10.1109/TASL.2007.911054.