Workshop program

Morning sessions

• 9:00 Opening [Slides: pdf]

Reinhold Haeb-Umbach (University of Paderborn), Tomohiro Nakatani (NTT Corporation)

• 9:10 Challenge summary [Slides: pdf]

Keisuke Kinoshita (NTT Corporation)

• 9:40 Keynote presentation

Chair: Armin Sehr (Beuth Hochschule für Technik Berlin)

• Robustness is dead! Long live Robustness!
  Dr. Mike Seltzer from Microsoft Research (USA)
  Abstract:
  With the widespread use of mobile phones and gaming consoles, and the surging interest in robotics and wearable computing, speech recognition systems are being deployed in more diverse environments than ever before. Thus, improving the robustness of speech recognizers to noise and reverberation would seem to be more important than ever. However, many state of the art systems, including those that use deep neural network-based acoustic models perform remarkable well without any explicit environmental compensation. So, is further effort in robustness critically important or totally unnecessary? I will show that while the DNN has remarkable performance compared to GMM-based acoustic models, it is it not a cure-all and in some cases, the DNN can be as fragile as previous models. To make further progress in these cases, a holistic view of the environmental robustness problem must be considered. I will present some examples of work in this direction, including systems for noise robustness and monaural speech separation.
  
  

• 11:00 Oral session 1

Chair: Steve Renals (University of Edinburgh)

• o1.1: The MERL/MELCO/TUM system for the REVERB Challenge using Deep Recurrent Neural Network Feature Enhancement [Slides: pdf]
  Felix J Weninger (Technische Universität München, Germany); Shinji Watanabe (Mitsubishi Electric Research Laboratories, USA); Jonathan Le Roux (Mitsubishi Electric Research Laboratories, USA); John Hershey (MERL, USA); Yuuki Tachioka (Mitsubishi Electric Corporation, Japan); Jürgen T. Geiger (Technische Universität München, Germany); Björn W Schuller (Imperial College London, Germany); Gerhard Rigoll (Technische Universität München, Germany)
  Abstract:
  This paper describes our joint submission to the REVERB Challenge, which calls for automatic speech recognition systems which are robust against varying room acoustics. Our approach uses deep recurrent neural network (DRNN) based feature enhancement in the log spectral domain as a single-channel front-end. The system is generalized to multi-channel audio by performing single-channel feature enhancement on the output of a sum-and-delay beamformer with direction of arrival estimation. On the back-end side, we employ a state-of-the-art speech recognizer using feature transformations, utterance based adaptation, and discriminative training. Results on the REVERB data indicate that the proposed front-end provides acceptable results already with a simple clean trained recognizer while being complementary to the improved back-end. The proposed ASR system with eight-channel feature enhancement achieves average word error rates of 7.75% and 20.09% on the simulated and real evaluation set, which is a drastic improvement over the Challenge baseline (25.26 and 49.16%). Further improvements can be obtained by system combination with a DRNN tandem recognizer, reaching 7.02% and 19.61% WER.
  
  
• o1.2: Joint Dereverberation and Noise Reduction Using Beamforming and a Single-Channel Speech Enhancement Scheme [Slides: pdf]
  Benjamin Cauchi (Fraunhofer IDMT-HSA, Germany); Ina Kodrasi (University of Oldenburg, Germany); Robert Rehr (University of Oldenburg, Germany); Stephan Gerlach (Fraunhofer IDMT-HSA, Germany); Ante Jukić (University of Oldenburg, Germany); Timo Gerkmann (University of Oldenburg, Germany); Simon Doclo (University of Oldenburg, Germany); Stefan Goetze (Fraunhofer IDMT-HSA, Germany)
  Abstract:
  The REVERB challenge provides a common framework for the evaluation of speech enhancement algorithms in the presence of both reverberation and noise. This contribution proposes a system consisting of a commonly used combination of a beamformer with a single-channel speech enhancement scheme aiming at joint dereverberation and noise reduction. First, a minimum variance distortionless response beamformer with an on-line estimated noise coherence matrix is used to suppress the noise and possibly some reflections. The beamformer output is then processed by a single-channel speech enhancement scheme, incorporating temporal cepstrum smoothing which suppresses both reverberation and residual noise. Experimental results show that improvements are particularly significant in conditions with high reverberation times.
  
  
• o1.3: A Multichannel Feature Compensation Approach for Robust ASR in Noisy and Reverberant Environments [Slides: pdf]
  Ramon Fernandez Astudillo (INESC-ID-Lisboa, Portugal); Sebastian Braun (International Audio Laboratories Erlangen, Germany); Emanuël Habets (International Audio Laboratories Erlangen, Germany)
  Abstract:
  In this paper we propose a multichannel feature compensation approach for automatic speech recognition in reverberant and noisy environments. The proposed technique propagates the posterior of the clean signal estimated by a multichannel Wiener filter in short-time Fourier transform (STFT) domain into Mel-frequency cepstrum coefficients (MFCC) domain. The multichannel Wiener filter reduces both reverberation and additive noise. Furthermore, we approximate the propagation of the prior distributions of speech and interference through the inverse STFT and the STFT with different time-frequency resolutions. This allows us to derive a multichannel minimum mean square error MFCC estimator with an STFT resolution that is different from the resolution in the speech enhancement stage. The proposed approach is able to outperform a multichannel short-time spectral amplitude estimation approach on both the clean training and multi-condition training ASR tasks of the REVERB challenge.
  
  

Afternoon sessions

• 13:30 Poster session 1

Chair: Markus Buck (Nuance Communications)

• p1.1: The DREAMS Initial Training Network: Dereverberation and Reverberation of Audio, Music, and Speech [Poster: pdf]
  Toon van Waterschoot (KU Leuven, Department of Electrical Engineering (ESAT-STADIUS),Belgium)
  Abstract:
  The DREAMS Initial Training Network investigates the problem of modeling, controlling, removing, and synthesizing acoustic reverberation with the aim of enhancing the quality and intelligibility of audio, music, and speech signals. This is achieved through a research and training program that builds upon four disciplines that are equally important in understanding and tackling the (de-)reverberation problem: room acoustics, signal processing, psychoacoustics, and speech and audio processing. The DREAMS ITN hosts 12 early stage researchers and 4 experienced researchers, each performing an individual research project around one of four themes that reflect the most challenging open problems in the area of (de-)reverberation.
  
  
• p1.2: A Speech Dereverberation Method Using Adaptive Sparse Dictionary Learning [Poster: pdf]
  Mohammadreza Moshirynia (Islamic Azad University, Science & Research Branch, Iran); Farbod Razzazi (Islamic Azad University, Science and Research Branch, Iran); Afrooz Haghbin (Tarbiat Modarres University, Iran)
  Abstract:
  We present a monaural blind de-reverberation method based on sparse coding of de-convolved version of reverberated speech signal in a dictionary which is learned by joint dictionary learning method, consisting of the concatenation of a clean speech and a non-negative matrix factor de-convolution result of the reverberated copy. The environment specific dictionary is originally learned off-line on a training corpus for different locations, while adaptive dictionary learning continues on-line for any other surroundings. Our approach uses both non-negative blind de-convolution and sparse coding, and achieves some improvements on objective voice quality testing's like perceptual evaluation of speech quality.
  
  
• p1.3: Dual system combination approach for various reverberant environments with dereverberation techniques [Poster: pdf]
  Yuuki Tachioka (Mitsubishi Electric Corporation, Japan); Tomohiro Narita (Mitsubishi Electric, Japan); Felix J Weninger (Technische Universität München, Germany); Shinji Watanabe (Mitsubishi Electric Research Laboratories, USA)
  Abstract:
  The recently introduced REVERB challenge includes a reverberant speech recognition task. We focus on state-of-the-art ASR techniques such as discriminative training and various feature transformations including Gaussian mixture model, sub-space Gaussian mixture model, and deep neural networks, in addition to the proposed single channel dereverberation method with reverberation time estimation and multi-channel beamforming that enhances direct sound compared with the reflected sound. In addition, because the best performing system is different from environment to environment, we perform a system combination approach using different feature and different types of systems to handle these various environments in the challenge. Moreover, we use our discriminative training technique for system combination that improves system combination by making systems complementary. Experiments show the effectiveness of these approaches, reaching 6.76% and 18.60% word error rate on the REVERB simulated and real test sets, which are 68.8% and 61.5% relative improvements over the baseline.
  
  
• p1.4: Single-channel reverberant speech recognition using C50 estimation [Poster: pdf]
  Pablo Peso Parada (Nuance Communications, United Kingdom); Dushyant Sharma (Nuance Communications, United Kingdom); Patrick A Naylor (Imperial College London, United Kingdom); Toon van Waterschoot (KU Leuven, Belgium)
  Abstract:
  We present several approaches to robust speech recognition in reverberant environments based on single-channel estimation of C50 . Our best method includes this estimation in the feature vector as an additional parameter and also uses C50 to select the most suitable acoustic model according to the reverberation level. We evaluate our method on the REVERB challenge database and show that our method outperforms the best baseline of the challenge, reducing the word error rate by 5.7% (corresponding to 16.8% relative word error rate reduction).
  
  
• p1.5: Adaptive dereverberation method based on complementary Wiener filter and modulation transfer function [Poster: pdf]
  Kento Ohtani (Nagoya University, Japan); Tatsuya Komatsu (Nagoya University, Japan); Takanori Nishino (Mie University, Japan); Kazuya Takeda (Nagoya University, Japan)
  Abstract:
  Many methods for dereverberation of speech have been proposed, however most of these methods require large computational resources. In this paper, we evaluate a method based on a complementary Wiener filter which requires very little computation. Because the filter coefficients used with this method have a large impact on dereverberation performance, we propose method to estimate these filter coefficients. Our filter estimation method is based on the observation that modulation transfer functions have inverse characteristics of dereverberation filter. We also use spectral subtraction for noise reduction and an ML estimator to estimate reverberation time. The results of our experimental evaluation show that our proposed method can effectively suppress reverberation with only a slight time delay.
  
  
• p1.6: The TUM system for the REVERB Challenge: Recognition of Reverberated Speech using Multi-Channel Correlation Shaping Dereverberation and BLSTM Recurrent Neural Networks
  Jürgen T. Geiger (Technische Universität München, Germany); Erik Marchi (Technische Universität München, Germany); Björn W Schuller (Imperial College London, Germany); Gerhard Rigoll (Technische Universität München, Germany)
  Abstract:
  This paper presents the TUM contribution to the 2014 REVERB Challenge: we describe a system for robust recognition of reverberated speech. In addition to an HMM-GMM recogniser, we use bidirectional long short-term memory (LSTM) recurrent neural networks. These networks can exploit long-range temporal context by using memory cells in the hidden units, which increases the robustness against reverberation. The LSTM is trained with phonemes as targets, and the predictions are converted into observation likelihoods and used as an acoustic model. Furthermore, we apply a dereverberation method called correlation shaping on the 8-channel recordings. This method applies a reduction of the long-term correlation energy in the received reverberant speech. The linear prediction residual, which generally contains information about reverberation, is processed to suppress the long-term correlation that is mostly due to the speaker-to-receiver impulse response. Using dereverberation as a front-end of the GMM in combination with the LSTM predictions leads to substantial improvements of the word error rate, achieving 11.19 % (relative improvement of about 35 %) and 28.13 % (improvement of about 30 %) with simulated and real data test sets, respectively. In the single-channel case, in which the dereverberation technique can not be applied, improvements of about 20 % (for simulated data) and 7 % (for real data) are obtained with the LSTM technique.
  
  
• p1.7: Speech Quality Enhancement based on Spectral Subtraction [Poster: pdf]
  Jéssica Veras (Federal University of Rio de Janeiro, Brazil); Thiago Prego (CEFET-RJ, Brazil); Amaro Lima (CEFET-RJ, Brazil); Tadeu Ferreira (Fluminense Federal University, Brazil); Sergio Netto (UFRJ, Brazil)
  Abstract:
  This paper presents an algorithm for speech enhancement of reverberant speech signals based on single channel blind spectral subtraction. This algorithm deals with the late components of the reverberation effect and it was optimized using 18 speech signals from the NBP database. Experimental results show that the proposed algorithm is well suited for speech enhancement in teleconference and telepresence environments and it can increase the perceptual quality by up to 31% and 62% of reverberant and noisy speech signals from databases with simulated and real reverberation and noise effects, respectively.
  
  
• p1.8: Reverberant speech recognition combining deep neural networks and deep autoencoders [Poster: pdf]
  M. Mimura (Kyoto University, Japan); Shinsuke Sakai (Kyoto University, Japan); Tatsuya Kawahara (Kyoto University, Japan)
  Abstract:
  We propose an approach to reverberant speech recognition adopting deep learning in front end as well as back end of the system. At the front end, we adopt a deep autoencoder for enhancing the speech feature parameters, and the decoding is performed using a DNN-HMM acoustic models trained on multi-condtion data. The system was evaluated through the ASR task in Chime Challenge 2014. The DNN-HMM system trained on the multi-condition training set achieved a conspicuously higher word accuracy compared to the MLLR-adapted GMM-HMM system trained on the same data. Furthermore, feature enhancement with the deep autoencoder contributed to the improvement of recognition accuracy especially in the more adverse conditions. When the DNN-HMM was used without the deep autoencoder front end, it resulted in a better performance than the non-adoped GMM-HMM system, but was not as good as the adapted GMM-HMM system. However, it outperformed the adapted GMM-HMM system when combined with the deep autoencoder.
  
  
• p1.9: The CMU-MIT REVERB Challenge 2014 System: Description and Results [Poster: pdf]
  Xue Feng (Massachusetts Institute of Technology, USA); Kenichi Kumatani (Spansion Inc., USA); John McDonough (Carnegie Mellon University, USA)
  Abstract:
  To evaluate state-of-the-art algorithms and draw new insights regarding potential future research directions in distant speech recognition, Kinoshita et al. launched the REverberant Voice Enhancement and Recognition Benchmark Challenge, commonly known as the REVERB Challenge, intended to provide a test bed for researchers to evaluate their methods based on common corpora and evaluation metrics. In this work, we describe our system and present our results on the 2014 REVERB Challenge. Our system is comprised of four primary components: an acoustic speaker tracking system to determine the speaker's position; this position is used for beamforming to focus on the desired speech while suppressing noise and reverberation; speaker clustering to determine sets of utterances spoken by the same speaker; and a speech recognition engine with speaker adaptation to extract word hypotheses from the enhanced waveforms produced by the beamformer. On the REAL RC evaluation data, our system obtained a word error rate of 39.9% with a single channel of the array, and 16.9% with the best beamformed signal.
  
  
• p1.10: Reverberation Robust Speech Recognition by Matching Distributions of Spectrally and Temporally Decorrelated Features [Poster: pdf]
  Kalle Palomäki (Aalto University, Finland); Heikki Kallasjoki (Aalto University, Finland)
  Abstract:
  This paper addresses dereverberation of speech using an unsupervised approach utilizing speech prior and taking only weak assumptions on reverberation. Our approach uses a long time context representation of reverberated speech in spectral-temporal supervectors which are decorrelated by the PCA. In the decorrelated domain supervectors are mapped from reverberant speech distribution to clean speech distribution and then to mel-spectral vectors. Mel-domain Wiener filter is applied as post processing. Our results demonstrate performance gains over the provided baseline recognizer, and show that the method can be coupled to CMLLR adaptation with cumulative benefits for clean trained models. Furthermore, we show that using dimensionality reduction coupled with the Wiener filter is better than using full dimensional PCA in representing small variance components in speech.
  
  
• p1.11: Enhancement of reverberant and noisy speech by extending its coherence [Poster: pdf]
  Scott Wisdom (University of Washington, USA); Thomas Powers (University of Washington, USA); Les Atlas (University of Washington, USA); James Pitton (University of Washington, USA)
  Abstract:
  We introduce a novel speech enhancement algorithm for removing reverberation and noise from recorded speech data. Our approach centers around using a single-channel minimum mean-square error log-spectral amplitude (MMSE-LSA) estimator, which applies gain coefficients in a time-frequency domain to suppress noise and reverberation. The main contribution of this paper is that the enhancement is done in a time-frequency domain that is coherent with speech signals over longer analysis durations than the short-time Fourier transform (STFT) domain. This extended coherence is gained by using a linear model of fundamental frequency variation over the analysis frame. In the multichannel case, we preprocess the data with either a minimum variance distortionless response (MVDR) beamformer, or a delay-and-sum beamformer (DSB). We evaluate our algorithm on the REVERB challenge dataset. Compared to the same processing done in the STFT domain, our approach achieves significant improvement on the REVERB challenge objective metrics, and according to informal listening tests, results in fewer artifacts in the enhanced speech.
  
  

• 14:50 Oral session 2

Chair: Patrick Naylor (Imperial College London)

• o2.1: Recognition of Reverberant Speech by Missing Data Imputation and NMF Feature Enhancement [Slides: pdf]
  Heikki Kallasjoki (Aalto University, Finland); Jort Gemmeke (KU Leuven, Belgium); Kalle Palomäki (Aalto University, Finland); Amy Beeston (University of Sheffield, United Kingdom); Guy Brown (University of Sheffield, United Kingdom)
  Abstract:
  The problem of reverberation in speech recognition is addressed in this study by extending a noise-robust feature enhancement method based on non-negative matrix factorization. The signal model of the observation as a linear combination of sample spectrograms is augmented by a mel-spectral feature domain convolution to account for the effects of room reverberation. The proposed method is contrasted with missing data techniques for reverberant speech, and evaluated for speech recognition performance using the REVERB challenge corpus. Our results indicate consistent gains in recognition performance compared to the baseline system, with a relative improvement in word error rate of 42.6% for the optimal case.
  
  
• o2.2: The NTU-ADSC Systems for Reverberation Challenge 2014 [Slides: pdf]
  Xiong Xiao (Nanyang Technological University, Singapore); Zhao Shengkui (Advanced Digital Sciences Center, Singapore); Duc Hoang Ha Nguyen (Nanyang Technological University, Singapore); Zhong Xionghu (Nanyang Technological University, Singapore); Douglas Jones (University of Illinois at Urbana-Champaign, USA); Eng-Siong Chng (Nanyang Technological University, Singapore); Haizhou Li (Institute for Infocomm Research, Singapore)
  Abstract:
  This paper describes our speech enhancement and recognition systems developed for the Reverberation Challenge 2014. To enhance the noisy and reverberant speech for human listening, besides using conventional methods such as delay and sum beamformer and late reverberation reduction by spectral subtraction, we also studied a novel learning-based speech enhancement. Specifically, we train deep neural networks (DNN) to map reverberant spectrogram to the corresponding clean spectrogram by using parallel data of clean and reverberant speech. Results show that the trained DNN is able to reduce reverberation significantly for unseen test data. For the speech recognition task, when parallel data is available, we train a DNN to map reverberant features to clean features, following the same spirit as the DNN-based speech enhancement. Results show that the DNN-based feature compensation improves speech recognition performance even when a DNN acoustic model is already used, showing the benefit of explicitly cleansing the features. When parallel data is not available in the clean condition training scheme, we focus on reducing the training-test mismatch by using our proposed cross transform feature adaptation that uses both temporal and spectral information. The cross transform works complementarily with traditional model adaptation.
  
  
• o2.3: Linear prediction-based dereverberation with advanced speech enhancement and recognition technologies for the REVERB challenge [Slides: pdf]
  Marc Delcroix (NTT Corporation, Japan); Takuya Yoshioka (NTT Corporation, Japan); Atsunori Ogawa (NTT Corporation, Japan); Yotaro Kubo (NTT Corporation, Japan); Masakiyo Fujimoto (NTT Corporation, Japan); Ito Nobutaka (NTT Corporation, Japan); Keisuke Kinoshita (NTT Corporation, Japan); Miquel Espi (NTT Corporation, Japan); Takaaki Hori (NTT Corporation, Japan); Tomohiro Nakatani (NTT Corporation, Japan); Atsushi Nakamura (NTT Corporation, Japan)
  Abstract:
  This paper describes systems for the enhancement and recognition of distant speech recorded in reverberant rooms. Our speech enhancement (SE) system handles reverberation with blind deconvolution using linear filtering estimated by exploiting the temporal correlation of observed reverberant speech signals. Additional noise reduction is then performed using an MVDR beamformer and advanced model-based SE. We employ this SE system as a front-end for our advanced automatic speech recognition (ASR) back-end, which uses deep neural network (DNN) based acoustic models and recurrent neural network based language models. Moreover, we ensure good interconnection between the SE front-end and ASR back-end using unsupervised model adaptation to reduce the mismatch caused by, for example, front-end processing artifacts. Our SE front-end greatly improves speech quality and achieves up to a 60% relative word error rate reduction for the real recordings of the REVERB challenge data, compared with a strong DNN-based ASR baseline.
  
  

• 16:10 Poster session 2

Chair: Dorothea Kolossa (Ruhr-Universität Bochum)

• p2.1: Acoustic Scene Aware Dereverberation using 2-channel spectral enhancement for REVERB Challenge [Poster: pdf]
  Xiaofei Wang (Institute of Acoustics, Chinese Academy of Sciences, P.R. China); Yanmeng Guo (Key Laboratory of Speech Acoustics and Content Understanding, P.R. China); Xi Yang (Institute of Acoustics, Chinese Academy of Sciences, P.R. China); Qiang Fu (Institute of Acoustics, Chinese Academy of Sciences, P.R. China); Yonghong Yan (Chinese Academy of Sciences, Key Laboratory of Speech Acoustics and Content Understanding, P.R. China)
  Abstract:
  A 2-channel dereverberation method contributed to the REVERB challenge is proposed in this paper. It aims at achieving robust dereverberation under different reverberant conditions. 2-channel spectral enhancement method is used where the gain of each time-frequency point is controlled by acoustic scene, which is detected based on the analysis of full-band coherent property. A preprocessing module is supplied to satisfy the requirement of enhancement. In addition, we describe the back-end re-trained acoustic model to match the front-end signal processing. Results, including enhancement indexes and improvement on recognition rate, are evaluated on the simulated data and really recorded data provided by REVERB organizers.
  
  
• p2.2: Single channel speech enhancement based on zero phase transformation in reverberated environments
  Dayana Ribas Gonzalez (Advanced Technologies Application Center, Cuba); Serguey Crespo Arias (Advanced Technologies Application Center (CENATAV), Cuba); Jose Ramon Calvo de Lara (Advanced Technologies Application Center, Cuba)
  Abstract:
  In this paper we present a single channel speech enhancement proposal in the autocorrelation domain. The main goal of this work is to compensate for speech in reverberated environments. However the proposed method is designed to handle any non-periodic corruption and it can carry out dereverberation and noise reduction at the same time. The work is framed in the REVERB (REverberant Voice Enhancement and Recognition Benchmark) challenge as part of the Speech Enhancement task. Results are given in both, WSJCAM0 Corpus for the simulated data and the MC-WSJ-AV corpus for the real recordings. Four quality measures have been computed for the experiments: cepstral distance, speech-to-reverberation modulation energy ratio, frequency weighted segmental SNR and log-likelihood ratio. Also the computational time of the speech enhancement proposal has been computed by measuring the Wall Clock Time. Promising results have been obtained both for the quality measures and for the computational time. In quality measures, the proposal achieves improvements until 57% using the proposed speech enhancement procedure as well as performing 16.8069 seconds faster than the reference.
  
  
• p2.3: Reverberation suppression based on sparse linear prediction in noisy environments
  Nicolás López (Institut Mines-Télécom, Télécom ParisTech, CNRS-LTCI, France); Gaël Richard (Institut Mines-Télécom, Télécom ParisTech, CNRS-LTCI, France); Yves Grenier (Institut Mines-Télécom, Télécom ParisTech, CNRS LTCI, France); Ivan Bourmeyster (Arkamys, France)
  Abstract:
  We present a single channel method for late reverberation suppression. The proposed approach estimates late reverberation as a linear combination of previous time-frequency frames. We impose a sparsity constraint on the predictor in order to select the most relevant signal frames for the estimation. The dataset used for the evaluation is corrupted by background noise, thus we propose to jointly suppress background noise and late reverberation. This leads to an important improvement in the quality of the processed signals as well as an improvement of the automatic speech recognition scores. The method appears to be efficient mainly in far field conditions and in highly reverberant environments. In addition, it is suitable for real time processing.
  
  
• p2.4: A computationally restrained and single-channel blind dereverberation method utilizing iterative spectral modifications
  Kazunobu Kondo (Yamaha Corporation, Japan)
  Abstract:
  A computationally restrained, single-channel, blind dereverberation method is proposed. The proposed method consists of two iterative spectral modifications, which employs spectral subtraction for noise reduction, and a complementary Wiener filter for dereverberation. Modulation transfer function is employed to calculate the dereverberation parameters. Late reverberation is estimated without any delaying operations, in contrast to other commonly used dereverberation methods. The proposed method achieves very balanced dereverberation and distortion reduction performance, in spite of the proposed rough T60 estimation technique. Some signal delay occurs as the result of Short Time Fourier Transform, but this delay is equivalent to the delay caused by conventional noise reduction methods such as spectral subtraction. Computational cost is sufficiently restrained, despite of the use of iterative spectral processing.
  
  
• p2.5: Robust features and system fusion for reverberation-robust speech recognition [Poster: pdf]
  Vikramjit Mitra (SRI International, USA); Wen Wang (SRI International, USA); Yun Lei (SRI International, USA); Andreas Kathol (Research Linguist, USA); Ganesh Sivaraman (University of Maryland, College Park, MD, USA); Carol Espy-Wilson (University of Maryland, USA)
  Abstract:
  Reverberation in speech degrades the performance of speech recognition systems, leading to higher word error rates. Human listeners are robust to reverberation effects, indicating that the auditory system somehow compensates for reverberation degradations. In this work, we present robust acoustic features motivated by the knowledge gained from human speech perception and production, and we demonstrate that these features provide reasonable robustness to reverberation effects compared to traditional mel-filterbank-based features. Using a single-feature system trained with the data distributed through the REVERB 2014 challenge on automatic speech recognition, we show a modest 12% and 0.2% relative reduction in word error rate (WER) compared to the mel-scale-feature-based baseline system for simulated and real reverberation conditions. The reduction is more pronounced when three systems are combined, resulting in a relative 20% reduction in WER for the simulated reverberation condition and a relative 11.7% reduction in WER for the real reverberated condition compared to the mel-scale-feature-based baseline system. The WER was found to reduce even further with addition of more systems trained with robust acoustic features. HLDA transform of features and MLLR adaptation of speaker clusters were also explored in this study and both of them were found to improve the recognition performance under reverberant conditions.
  
  
• p2.6: Speech Dereverberation by Constrained and Regularized Multi-Channel Spectral Decomposition: Evaluated on REVERB Challenge
  Meng Yu (Audience Inc., USA); Frank Soong (Microsoft Research Asia, USA)
  Abstract:
  We present our contribution to the REVERB Challenge in this paper. A multi-channel speech dereverberation system combines cross-channel cancellation and spectral decomposition. The reverberation is modeled as a convolution operation in the spectral domain. Using the generalized Kullback-Leibler (KL) divergence, we decompose the reverberant magnitude spectrum into clean magnitude spectrum convolved with a deconvolution filter. The magnitude spectrum is constrained and regularized by non-negativity and sparsity, respectively, while the deconvolution filter is constrained by non-negativity and cross-channel cancellation. Spectral decomposition of individual channels and cross-channel cancellation are jointly optimized by a multiplicative algorithm to achieve multi-channel speech dereverberation. Experimental evaluations on "speech enhancement task" are carried out according to the evaluation guidelines of the REVERB challenge, showing promising results. The objective metrics for measuring reverberation are investigated through the algorithm evaluation.
  
  
• p2.7: Sparse Recovery Method for Dereverberation [Poster: pdf]
  Nicolas Epain (University of Sydney, Australia); Tahereh Noohi (University of Sydney, Australia); Craig Jin (University of Sydney, Australia)
  Abstract:
  In this paper, we present our contribution to the REverberant Voice Enhancement and Recognition Benchmark (REVERB) challenge. We propose a multichannel dereverberation algorithm which enhances a target speech signal in a reverberant environment. The algorithm consists of two main steps. First, the directive component of the sound field is extracted from the microphone signals. Second, sparse recovery is employed to beamformed the extracted directive signals towards the target speaker. Experiments were carried out on the evaluation dataset of the challenge. The results of these experiments indicate that the method effectively de-reverberates the signals.
  
  
• p2.8: Robust ASR in Reverberant Environments using Temporal Cepstrum Smoothing for Speech Enhancement and an Amplitude Modulation Filterbank for Feature Extraction [Poster: pdf]
  Feifei Xiong (Fraunhofer IDMT, Germany); Niko Moritz (Fraunhofer IDMT-HSA, Germany); Robert Rehr (University of Oldenburg, Germany); Joern Anemueller (University of Oldenburg, Germany); Bernd Meyer (University of Oldenburg, Germany); Timo Gerkmann (University of Oldenburg, Germany); Simon Doclo (University of Oldenburg, Germany); Stefan Goetze (Fraunhofer IDMT-HSA, Germany)
  Abstract:
  This paper presents techniques aiming at improving automatic speech recognition (ASR) in single channel scenarios in the context of the REVERB (REverberant Voice Enhancement and Recognition Benchmark) challenge. System improvements range from speech enhancement over robust feature extraction to model adaptation and word-based integration of multiple classifiers. Temporal cepstrum smoothing (TCS) technique is applied to enhance the reverberant speech signal at moderate noise levels, based on a statistical model of room impulse responses (RIRs) and minimum statistics (MS), considering estimates of late reverberations and the noise power spectrum densities (PSDs). Robust feature extraction is performed by amplitude modulation filtering of the cepstrogram to extract its temporal modulation information. As an alternative classifier the acoustic models have been adopted using different RIRs and a RIR selection scheme based on a multi-layer perceptron (MLP) system that uses spectro-temporal features as the input. In the final stage, a system combination approach achieved by recognizer output voting error reduction (ROVER) is employed to obtain a jointly optimal recognized transcription. The proposed system has been evaluated in two different processing modes, i.e. utterance-based batch processing and full batch processing, which results in an overall average absolute improvement of 11% under variant reverberant conditions compared to the baseline system.
  
  
• p2.9: Extraction of Robust Features by Combining Noise Reduction and FDLP for the Recognition of Noisy Speech Signals in Hands-free Mode [Poster: pdf]
  Hans-Günter Hirsch (Niederrhein University of Applied Sciences, Germany)
  Abstract:
  A processing scheme is presented to create robust acoustic features for the recognition of noisy speech signals at a hands-free speech input in reverberant environments. The robustness is achieved by extending a Mel cepstral analysis scheme by the additional processing steps of a noise reduction and a frequency domain linear prediction (FDLP). FDLP is a signal processing technique to modify the energy contours of subband signals. We analyze the effects of FDLP and show its ability to create robust features by looking at the energy contours of clean and reverberant speech signals. This technique is combined with a noise reduction approach that is based on an adaptive filtering in the spectral domain. We analyze the effectiveness of our approach by running several different recognition experiments and evaluating the achieved recognition results. The experiments contain recognition tasks with different complexity and with a different modeling of speech by applying whole-word or phoneme based reference models. Besides the recognition of reverberant versions of the TIDigits we focus on the experiments as defined by the Reverb Challenge task where the focus is put on the effects of a hands-free speech input without a lot of noise in the background. Furthermore, we created two further versions of the development test data to include the effects of typical noise scenarios at a realistic signal-to-noise ratio (SNR) in room environments.
  
  
• p2.10: Use Of Multiple Front-Ends And I-Vector-Based Speaker Adaptation For Robust Speech Recognition [Poster: pdf]
  Md. Jahangir Alam (INRS Énergie, Matériaux et Télécommunications, Canada); Vishwa Gupta (CRIM, Canada); Patrick Kenny (CRIM, Canada); Pierre Dumouchel (Ecole de technologie superieure, Canada)
  Abstract:
  Although state-of-the-art speech recognition systems perform well in controlled environments they work poorly in realistic acoustical conditions in reverberant environments. Here, we use multiple front-ends (conventional mel-filterbank, multitaper spectrum estimation-based mel filterbank, robust Mel and compressive gammachirp filterbank, iterative deconvolution-based dereverberated mel-filterbank, and maximum likelihood inverse filtering-based dereverberated mel-frequency cepstral coefficients) - based recognition systems with multi-condition training data and combined their results using ROVER (Recognizer Output Voting Error Reduction). For 2- and 8- channel tasks, to get benefited from more than one channel, we also utilize ROVER instead of the multi-microphone signal processing method. Similar to Vishwa Gupta's previous work we also apply i-vector -based speaker adaptation which was found effective. Speech recognition Experiments are conducted on the REVERB challenge 2014 corpora using Kaldi recognizer. For 2-channel task (using full batch processing) we obtained an average WER of 9.0% and 23.4% on the SimData and RealData respectively. Whereas for 8-channel task on the SimData and RealData the average WER found were 8.9% and 21.7%, respectively.
  
  
• p2.11: PBF-GSC Beamforming for ASR and Speech Enhancement in reverberant environments [Poster: pdf]
  Yi Ren Leng (Institute for Infocomm Research, Singapore); Jonathan Dennis (Institute for Infocomm Research, Singapore); Wen Zheng Terence Ng (Institute for Infocomm Research, Singapore); Tran Huy Dat (Institute for Infocomm Research, Singapore)
  Abstract:
  Reverberant noise is present in most practical applications of Automatic Speech Recognition (ASR) that do not rely on close-talk microphones. Unlike additive noise, there is limited progress in reducing the effect of reverberant noise in speech. For our REverberant Voice Enhancement and Recognition Benchmark (REVERB) challenge submission, we propose to use of a beamforming system combining the Generalized Sidelobe Canceller (GSC) with Phase-error based filtering (PBF). Our system is calibrated to make use of the baseline HMM recognizer with CMLLR and multi-condition training as it is the most practical setup for realistic applications.From the results we obtained, it appears that the main contribution to reverberant noise reduction comes from the combination of multiple audio streams while the adaptive filtering of GSC provides a slight improvement on top of this combination. For the speech enhancement task, we simply apply a previous noise reduction method, which is based on generalized gamma distribution modeling of speech spectral magnitude, on the output of developed PBF-GSC beamforming.
  
  

• 17:30 Panel session [Slides: pdf, Summary: pdf]

Moderators: Reinhold Haeb-Umbach (University of Paderborn), Sharon Gannot (Bar-Ilan University), Tomohiro Nakatani (NTT Corporation)

• 18:00 Closing session