Contaminated speech training methods for robust DNN-HMM distant speech recognition

Contaminated speech training methods f or r obust DNN-HMM distant speech r ecognition Mir co Ravanelli, Maurizio Omologo Fondazione Bruno K essler , T rento, Italy mravanelli@fbk.eu, omologo@fbk.eu Abstract Despite the significant progress made in the last years, state-of- the-art speech recognition technologies provide a satisfactory performance only in the close-talking condition. Robustness of distant speech recognition in adverse acoustic conditions, on the other hand, remains a crucial open issue for future applications of human-machine interaction. T o this end, several adv ances in speech enhancement, acoustic scene analysis as well as acous- tic modeling, hav e recently contributed to improv e the state-of- the-art in the field. One of the most effecti ve approaches to de- riv e a robust acoustic modeling is based on using contaminated speech, which proved helpful in reducing the acoustic mismatch between training and testing conditions. In this paper , we re vise this classical approach in the context of modern DNN-HMM systems, and propose the adoption of three methods, namely , asymmetric context windowing, close- talk based supervision, and close-talk based pre-training. The experimental results, obtained using both real and simulated data, sho w a significant advantage in using these three meth- ods, ov erall providing a 15% error rate reduction compared to the baseline systems. The same trend in performance is con- firmed either using a high-quality training set of small size, and a large one. Index T erms : distant speech recognition, multi-condition train- ing, DNN. 1. Introduction During the last decade, much research has been dev oted to im- proving Automatic Speech Recognition (ASR) performance. Nev ertheless, most state-of-the-art systems are still based on close-talking solutions, forcing the user to speak v ery close to a microphone-equipped de vice. There are, ho wev er , various real- life situations where Distant Speech Recognition (DSR) is more natural, con venient and attractiv e [1]. An emerging applica- tion is, for instance, speech-based domestic control, where users might prefer to freely interact with their home appliances with- out wearing or e ven handling any microphone-equipped de vice. This scenario was addressed under the EU DIRHA project 1 , which had the ultimate goal of de veloping real-time systems for motor-impaired users. Unfortunately , current DSR technologies still exhibit a significant lack of robustness and flexibility , due to the adverse acoustic conditions originated by non-stationary noises and acoustic rev erberation [2]. Sev eral efforts have been devoted during the last years to improv e DSR, as also witnessed by some recent international challenges such as REVERB [3], CHIME [4, 5] and ASpIRE. 1 This work was partially funded by EU’ s 7th Framework Pro- gramme under grant agreement n. 288121 DIRHA. More details can be found in http://dirha.fbk.eu/ . Considerable progresses were fostered by substantial advances in spatial filtering [6, 7], microphone selection [8], source sepa- ration [9], speech dere verberation [10] as well as speaker local- ization [11], acoustic ev ent detection [12] and acoustic model- ing. Concerning the latter field, the research community is cur- rently e xperiencing a revolution due to the introduction of Deep Neural Netw orks (DNNs) [13], which hav e consistently outper- formed previous Gaussian Mixture Models (GMMs) for both close [14] and distant-talking scenarios [15, 16, 17, 18, 19]. One of the most effecti ve and straightforward approach so far ap- plied to DSR for acoustic modeling is based on multi-condition training with contaminated speech corpora [20, 21, 22]. Con- taminated speech is generated by con volving close-talking sig- nals with an Impulse Response (IR) usually measured in the targeted en vironment. Background noise is then typically added to the con voluted signals, in order to make the resulting corpus more realistic and better matching real-world conditions. This approach has also been adopted recently in [3, 4, 23, 24]. This paper deals with the use of contaminated speech to train a DNN-HMM speech recognition system. T raining such very complex systems is typically dif ficult and tricky . The basic strategy to run it, and the way one performs optimization steps, often play a crucial role in order to get best performance. Our work, which relies on the Kaldi toolkit [25], has concerned all those optimization ef forts, e.g., including an optimal choice of number of hidden layers and of units. Howe ver , in this paper we focus on two specific aspects, about which we believe the exper- imental results demonstrate nov el and interesting evidences. First, we propose some methods able to exploit informa- tion from both distant and close-talking datasets to deriv e ro- bust acoustic models. In particular , one proposed approach con- sists in inheriting the high-quality labels that might be generated from close-talk datasets to train distant-talking DNNs. Another nov el method consists in replacing a standard Restricted Boltz- mann Machines (RBMs) based pre-training [26] with a super - vised close-talk pre-training, leading to a smarter initialization of the distant-talking DNN. As a second contribution, we propose to replace the tradi- tional symmetric conte xt windows, which are typically adopted to gather sev eral consecutiv e features in the DNN framework, with an Asymmetric Conte xt W indow (ACW). Interestingly , this approach has prov ed effecti ve in counteracting the harm- ful effects of the acoustic rev erberation, due to a better use of the contextual information. The experimental validation is carried out in a living-room of a real apartment using both simulated and real test data. Sim- ilarly to [23], the e valuation is performed on a phone-loop task, in order to av oid any non-linear influence and artifacts possibly originated by a language model. The rest of the paper is organized as follows. In Sec. 2 the 453 cm 485 cm A B T1 D1 R1 R2 R3 R4 T r aining D ev elopmen t T est -Sim T est -R eal C orp or a P ositions: Figure 1: The reference DIRHA living-room en vironment. Squares and arrows represent the speaker positions and orien- tations adopted for the real and simulated corpora. contamination approach is outlined, while Sec. 3 and 4 describe the adopted corpora and the ASR system, respectively . Finally , the proposed methods are experimentally v alidated in Sec. 5, while Sec. 6 dra ws the conclusions. 2. Speech contamination The speech contamination process is represented by the follo w- ing basic relationship: y ( t ) = x ( t ) ∗ h ( t ) + αn ( t ) (1) where x ( t ) and y ( t ) are the close and distant-talking signals, re- spectiv ely , while n ( t ) consists in recorded environmental noise controlled by the gain α in order to obtain dif ferent SNRs. h ( t ) is an IR that describes the acoustic rev erberation effects of a generic source-microphone path. In particular, if one assumes to deal with a linear time-inv ariant acoustical transmission sys- tem, the IR provides a complete description of the changes a sound signal under goes when it tra vels from a particular posi- tion in space to a giv en microphone [27]. The method adopted here for estimating IR consists in diffusing in the target envi- ronment a known Exponential Sine Sweep (ESS) signal, and recording it by a microphone properly placed in space [28]. This method of measuring the IR has a remarkable impact on the speech recognition performance, as discussed in [29]. 3. Simulated and Real Corpora The contamination process described in Sec. 2, can be exploited to generate simulated data starting from dif ferent close-talking datasets. In this work, simulated data are used for training and dev elopment purposes only , while test is performed also on real data. The sampling rate is 16 kHz for all the datasets. In the following sections, the reference environment, the close-talking corpora, and the simulated and real datasets are described. 3.1. Acoustic En vironment and Microphone Set-up The reference en vironment considered for this study is the living-room of a real apartment, which was av ailable for ex- periments under the DIRHA project. The original microphone network includes several microphones installed on the walls and on the ceiling. Howe ver , due to the purpose of this work, in the following we will consider only one microphone, i.e., the central one of the ceiling array . As sho wn in Fig. 1, an IR measurement session exploring a lar ge number of positions and orientations of the sound source has been conducted. The mean rev erberation time T 60 measured in the room is 0 . 7 s. 3.2. Close-talking Corpora In this work, tw o close-talk datasets are used for training, namely AP ASCI and Euronews. AP ASCI [30] consists of more than 5200 phonetically-rich sentences uttered by 163 speakers and recorded in a professional studio, ov erall corresponding to about 6 hours. For comparison purposes, the Italian part of the Euronews database [31], consisting of about 100 hours of TV news, is also considered for training. The Eurone ws signals are frequently corrupted by cross-talk, background music, en viron- mental noise as well as spontaneous speech phenomena. A small dev elopment set is also adopted to tune the free- parameters of the DSR systems, as discussed in Sec. 5.1. T o this purpose, a portion of the DIRHA corpus [32], consisting of 220 high-quality phonetically-rich sentences uttered by 10 speakers is employed. For testing purposes, a different portion of the DIRHA cor - pus, composed of 430 phonetically-rich sentences uttered by 20 speakers is adopted. 3.3. Simulated Corpora For each of the close-talking corpora outlined in Sec. 3.2, two contaminated versions were generated. The first version is based on a con volution of the close-talking sentences with im- pulse responses measured in the reference li ving-room en viron- ment. Such simulated datasets are intended to study a scenario where only rev erberation acts as a source of disturbance. Fur- thermore, a second version of the contaminated databases was directly deriv ed from the reverberated datasets by adding noisy background sequences recorded in the target environment, en- suring a mean SNR of about 10 dB . This version is intended to study more challenging scenarios where both re verberation and noises act as a source of disturbance. For the training corpora a single impulse response (from T1 in Fig.1) was considered, while for dev elopment a different IR (D1) was adopted. As shown in [23], note that a single impulse response is sufficient to model room rev erberation effects for DSR purposes. 3.4. DIRHA Real Corpus Besides the simulated test-sets, a corpus of real phonetically- rich sentences was recorded in the reference living-room. These real recordings in volved 18 nati ve Italian speakers (9 males and 9 females), who read a list of 40 phonetically-rich sentences in four different positions (R1-4 in Fig.1). In total, 2880 sentences, corresponding to more than two hours of speech material were acquired. These recordings were conducted in quiet acoustic conditions, where re verberation was the main source of distur - bance. The estimated SNR is about 22 dB . 4. DNN based ASR In this work, we use a Context-Dependent DNN-HMM speech recognizer , where every unit is modeled by a three state left- to-right HMM, and the tied-state observation probabilities are estimated through a DNN. Feature extraction is based on blocking the signal into frames of 25 ms with 10 ms overlapping. For each frame, 13 MFCCs plus pitch and Probability of V oicing (PoV) are ex- tracted. The pitch and PoV are estimated through the normal- ized autocorrelation method discussed in [33]. The resulting CW Frames 20 -20 -10 0 10 PER (%) 14 15 16 17 Clean CW Frames -20 -10 0 10 20 PER (%) 38 40 42 44 46 48 Rev CW Frames -20 -10 0 10 20 PER (%) 54 56 58 60 62 Rev&Noise Figure 2: PER(%) obtained by DNN conte xt windows which progressi vely inte grate past or future frames. features, together with their first and second order deriv ati ves, are then arranged into a single observation vector of 45 compo- nents. Finally , a context window gathering several consecutiv e frames followed by a mean and variance normalization of the feature space are applied before feeding the DNN. The DNN, trained with the Kaldi toolkit (Karel’ s recipe) [25], is composed of sigmoid-based hidden neurons, while the output layer is based on softmax activ ation functions. The pre- training phase is carried out by stacking Restricted Boltzmann Machines (RBM) [26] to form a deep belief network, while the fine-tuning is performed by a stochastic gradient descent opti- mizing cross-entropy loss function. In the latter phase, the ini- tial learning rate is kept fixed as long as the increment of the frame accuracy on the dev-set is higher than 0.5 % . For the fol- lowing epochs, the learning rate is halved until the increment of frame accuracy is less than the stopping threshold of 0.1 % . The decoding is performed by adopting a phone-loop based grammar . As previously outlined, even though the use of more complex grammars or language models is certainly helpful in increasing the recognition performance, the adoption of a sim- ple phone-loop is due to the need of an experimental evidence not biased by a LM. In this study , as in [23], a set of 26 phone units of the Italian language was chosen for e valuation purposes. 5. Experiments In the following section, the proposed techniques are better de- scribed and an experimental validation is pro vided 2 . The ex- periments are conducted in three dif ferent acoustic scenarios of increasing complexity: close-talking, distant-talking with re ver- beration (Rev), and distant-talking with both noise and reverber - ation (Rev&Noise). The reported performance is obtained when a matching acoustic condition between test and training is met (e.g., for the reverberated speech test, the reverberated training is used). The ev aluation is performed on both simulated and real data, using either AP ASCI or Eurone ws training-sets. 5.1. Baselines This section reports the results obtained with a standard DNN- HMM system, after optimizing the main ASR free-parameters on the dev-set. The best performance is reached with a DNN composed of 6 hidden layers with 1500 neurons per layer, adopting a symmetric context window of 17 frames (8 before and 8 after the current frame) and using a learning rate of 0.008. Note that such free-parameters are used for both the training corpora and for all the addressed acoustic scenarios, due to a 2 Part of the experiments are conducted with a T esla K40 donated by the NVIDIA Corporation. similar outcome of the optimization step. Only a dif ference con- cerning the optimal number of tied-states is observed, leading to 2.5k and 5k tied-states for AP ASCI and Eurone ws, respecti vely . T able 1 sho ws the performance achiev ed on the test-sets. P P P P P P P T est T rain AP ASCI (6 h) Euronews (100 h) Close-talk 14.0 14.5 Sim-Rev 37.6 37.0 Real-Rev 41.0 40.1 Sim-Rev&Noise 52.3 51.8 T able 1: Phone Error Rate (PER%) of the baseline system. T able 1 clearly highlights that the performance in distant- talking scenarios is dramatically reduced, if compared to a close-talking case. This loss of performance further confirms how challenging DSR is under adverse acoustic conditions and without any help from grammars or LMs. It is also highly rel- ev ant that with only 6 hours of high-quality speech material a performance similar to that of a 100 hours corpus is obtained. 5.2. Asymmetric Context Window (A CW) In distant-talking scenarios, the acoustic rev erberation intro- duces a long-term redundancy in the signal. This disturbance can be modeled as a long causal FIR filter, whose taps introduce a forward memory in the speech sequence. As a consequence, since the length of the window adopted for features extraction (i.e., 25 ms) is much smaller than the reverberation time (i.e, 700 ms), a progressiv e concatenation of future frames tends to embed information correlated with the central frame contents. On the other hand, a concatenation of past frames seems to be more helpful since, on average, dif ferent and complementary information can potentially be analyzed to perform a frame- lev el prediction. Based on this concept, we explore the use of an Asymmetric Context W indow (A CW), which analyzes more past frames than future frames. Fig. 2 reports a first e xperiment in volving context windows which progressively integrate only past (negati ve x-axis) or fu- ture frames (positive x-axis). In this section, only results ob- tained using AP ASCI as training-set are reported. Howe ver , a similar trend is observed with Eurone ws. From Fig. 2, it is clear that in a close-talking scenario a symmetric behaviour is attained, meaning that past and fu- ture information provides a similar contribution in impro ving the system performance. Conv ersely , the role of the past in- formation seems significantly more important in distant-talking scenarios, since a faster decrease of the PER(%) is progres- siv ely obtained when past frames are concatenated. Along this line, another experiment, reported in T able 2, shows the results achiev ed when the symmetric context window of 17 frames in- troduced in Sec. 5.1 is replaced by different asymmetric context windows. P P P P P P P T est T rain Context W indow P16-F0 P10-F6 P8-F8 P6-F10 P0-F16 Close-talk 15.5 14.2 14.0 14.4 16.3 Sim-Rev 38.0 37.0 37.6 38.3 43.6 Real-Rev 41.6 40.2 41.0 41.6 46.8 Sim-Rev&Noise 53.8 51.8 52.3 53.0 57.0 T able 2: PER(%) obtained with Symmetric and Asymmetric context windows composed of 17 frames. “P” refers to the num- ber of past frames, while “F” refers to future frames. Consistently with the previous experiment, in a close- talking scenario the typical symmetric window performs better than any asymmetric window . Differently , an asymmetric con- text window which embeds more past information (10 frames) than future information (6 frames) performs slightly better in distant-talking conditions. Note that this is consistent for both real and simulated test-sets. Moreover , results obtained with different context windo w lengths (e.g., 9, 11, 13, 21, 27 frames) further confirms that ACW is a better choice in distant-talking scenarios, thanks to an improved robustness against reverber - ation. Note also that the use of an ACW does not imply any additional computational effort and is particularly suitable for on-line systems [34, 35], since the latency originated by wait- ing for the future frames is minimized. 5.3. Close-talking labels for Distant-talking DNN In standard ASR, the labels for DNN training are derived by a forced-alignment of the training corpus ov er the tied-states. Al- though some GMM-free solutions have been proposed [36], this alignment is typically performed using a standard CD-GMM- HMM system. Howe ver , this phase can be v ery critical because a precise alignment could be difficult to reach, especially in challenging acoustic scenarios characterized by noise and rev er- beration. As a consequence, the DNN learning process might be more critical due to a poor supervision. In the contaminated speech training frame work, howe ver , a more precise supervision can be obtained from the close-talking dataset. F or this reason, we propose to inherit the high-quality labels that can be generated from the close-talking data to train the distant-talking DNNs. This requires to train, with the origi- nal clean datasets, a standard CD-GMM-HMM system, and e x- ploit it to generate a precise tied-state forced alignment ov er the close-talking training corpus, later inherited as supervision for the distant-talking DNN. T able 3 reports the results obtained with the standard approach, which is based on labels deriv ed from distant-talking signals, and the proposed solution, which directly inherits labels derived from close-talking signals. The A CW of 17 frames (P10-F6) discussed in Sec. 5.2 is still used for the following e xperiments. Results show that the proposed approach provides a sub- stantial improvement in the performance, over both real and simulated data. Specifically , a relativ e improv ement of 10% and 12% is achie ved for AP ASCI and Eurone ws, respecti vely . Be- sides this significant performance improvement, another inter- esting aspect is the faster con vergence of the iterative learning procedure, due to a better supervision provided to the DNN. In P P P P P P P T est T rain AP ASCI (6 h) Euronews (100 h) Standard CT -lab Standard CT -lab Sim-Rev 37.0 33.0 36.1 32.1 Real-Rev 40.2 35.9 39.3 34.3 Sim-Rev&Noise 51.8 47.3 50.1 46.4 T able 3: PER(%) obtained in distant-talking scenarios with the standard and with the proposed technique based on close- talking labels (CT -lab). these experiments, 15 epochs are needed to con verge with the standard solution, while only 12 are sufficient with the proposed approach, reducing the training time of 20%. 5.4. Close-talking Pre-T raining for Distant-talking DNN In this section, we propose a further way of taking advantage of the rich information embedded in the close-talking dataset. In particular , instead of adopting a standard unsupervised RBM pre-training of the DNN, we propose to use a supervised pre- training method based on the close-talking data. More precisely , the idea is to train a close-talking DNN and inherit its parame- ters for initializing the distant-talking DNN. A subsequent fine tuning phase is then carried out on distant-talking data using a slightly reduced learning rate (LR=0.005 is used here). T able 4 reports the results obtained using such approach on real and simulated datasets. The A CW (P10-F6) discussed in Sec. 5.2 and the close-talking labels proposed in Sec. 5.3 are still used for this experiment. P P P P P P P T est T rain AP ASCI (6 h) Euronews (100 h) Stand-PT CT -PT Stand-PT CT -PT Sim-Rev 33.0 31.4 32.1 31.2 Real-Rev 35.9 34.7 34.3 33.0 Sim-Rev&Noise 47.3 46.2 46.4 45.1 T able 4: PER(%) obtained in distant-talking scenarios with a standard RBM pre-training (Stand-PT) and with the proposed supervised close-talking pre-training (CT -PT). Results clearly sho w an impro vement deriving from this ap- proach, which is consistent over both simulated and real data. This suggests that a close-talk pre-training is a smart way of initializing the DNN, which someho w first learns the speech characteristics and only at a later stage learns ho w to counteract the adverse acoustic conditions. 6. Conclusions and Future W ork In this paper, some nov el methods for taking advantage of con- taminated speech training in modern DNN-HMM systems are proposed. In particular, we explored approaches based on an asymmetric context window , a close-talk supervision and a su- pervised close-talk pre-training. Experimental results are very promising with a relativ e performance improvement of more than 15% over the baseline system on both real and simulated data. Along this line, our next acti vities will concern the explo- ration of different architectures, such as conv olutional and re- current NNs. W e also plan to consider different acoustic en vi- ronments, languages and tasks and, at a later stage, to address a multi-microphone case. 7. References [1] M. W ¨ olfel and J. McDonough, Distant Speech Recognition . W i- ley , 2009. [2] E. H ¨ ansler and G. Schmidt, Speech and Audio Processing in Ad- verse En vir onments . Springer , 2008. [3] K. Kinoshita, M. Delcroix, T . Y oshioka, T . Nakatani, E. Habets, R. Haeb-Umbach, V . Leutnant, A. Sehr , W . Kellermann, R. Maas, S. Gannot, and B. Raj, “The reverb challenge: A common e valua- tion framework for derev erberation and recognition of rev erberant speech, ” in W ASP AA 2013 , 2013, pp. 1–4. [4] J. Barker , E. V incent, N. Ma, H. Christensen, and P . Green, “The P ASCAL CHiME speech separation and recognition challenge.” Computer Speech and Language , vol. 27, no. 3, pp. 621–633, 2013. [5] J. Barker , R. Marxer , E. V incent, and S. W atanabe, “The third CHiME Speech Separation and Recognition Challenge: Dataset, task and baselines, ” in Submitted to ASRU 2015 , 2015. [6] M. Brandstein and D. W ard, Microphone arrays . Springer , Berlin, 2000. [7] W . Kellermann, Beamforming for Speech and Audio Signals . in HandBook of Signal Processing in Acoustics, Springer , 2008. [8] M. W olf and C. Nadeu, “Channel selection measures for multi- microphone speech recognition, ” Speech Communication , vol. 57, pp. 170–180, Feb . 2014. [9] S. Makino, T . Lee, and H. Saw ada, Blind Speec h Separation . Springer , 2010. [10] P . A. Naylor and N. D. Gaubitch, Speech Dere verberation. Springer , 2010. [11] A. Brutti, M. Rav anelli, P . Svaizer , and M. Omologo, “ A speech ev ent detection/localization task for multiroom en vironments, ” in Pr oc. of HSCMA , 2014, pp. 157–161. [12] A. T emko, C. Nadeu, D. Macho, R. Malkin, C. Zieger , and M. Omologo, “ Acoustic event detection and classification, ” in Computers in the Human Interaction Loop . Springer London, 2009, pp. 61–73. [13] D. Y u and L. Deng, Automatic Speech Recognition - A Deep Learning Appr oach . Springer , 2015. [14] G. Dahl, D. Y u, L. Deng, and A. Acero, “Context-dependent pre- trained deep neural networks for large vocabulary speech recog- nition, ” IEEE T ransactions on Audio, Speech, and Language Pr o- cessing , vol. 20, no. 1, pp. 30–42, 2012. [15] P . Swietojanski, A. Ghoshal, and S. Renals, “Hybrid acous- tic models for distant and multichannel large vocabulary speech recognition, ” in Pr oc. of ASRU 2013 , 2013, pp. 285–290. [16] Y . Liu, P . Zhang, and T . Hain, “Using neural network front-ends on far field multiple microphones based speech recognition, ” in Pr oc. of ICASSP 2014 , 2014, pp. 5542–5546. [17] F . W eninger, S. W atanabe, J. Le Roux, J. Hershey , Y . T achioka, J. Geiger , B. Schuller , and G. Rigoll, “The MERL/MELCO/TUM System for the REVERB Challenge Using Deep Recurrent Neu- ral Network Feature Enhancement, ” in IEEE REVERB W orkshop , 2014. [18] S. S. Masato Mimura and T . Kawahara, “Reverberant speech recognition combining deep neural networks and deep autoen- coders, ” in IEEE REVERB W orkshop , 2014. [19] A. Schwarz, C. Huemmer, R. Maas, and W . Kellermann, “Spatial diffuseness features for dnn-based speech recognition in noisy and rev erberant en vironments, ” in Accepted at ICASSP 2015 , 2015. [20] M. Matassoni, M. Omologo, D. Giuliani, and P . Svaizer , “Hid- den Markov model training with contaminated speech material for distant-talking speech recognition.” Computer Speech & Lan- guage , v ol. 16, no. 2, pp. 205–223, 2002. [21] L. Couvreur, C. Couvreur, and C. Ris, “A corpus-based approach for robust ASR in reverberant en vironments.” in Pr oc. of INTER- SPEECH , 2000, pp. 397–400. [22] T . Haderlein, E. Nth, W . Herbordt, W . Kellermann, and H. Nie- mann, “Using Artificially Reverberated Training Data in Distant- T alking ASR.” ser . Lecture Notes in Computer Science, vol. 3658. Springer , 2005, pp. 226–233. [23] M. Ravanelli and M. Omologo, “On the selection of the impulse responses for distant-speech recognition based on contaminated speech training, ” in Pr oc. of INTERSPEECH 2014 , 2014, pp. 1028–1032. [24] A. Brutti and M. Matassoni, “On the use of Early-to-Late Rever - beration Ratio for ASR in reverberant environments, ” in Pr oc. of ICASSP 2014 , 2014, pp. 4670–4675. [25] D. Pov ey , A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P . Motlicek, Y . Qian, P . Schwarz, J. Silovsky , G. Stemmer , and K. V esely , “The Kaldi Speech Recognition T oolkit, ” in ASRU 2011 , 2011. [26] O. S. T . Y . Hinton, G.E., “ A fast learning algorithm for deep belief nets, ” Neural Computation , vol. 18, no. 1, pp. 1527–1554,, 2006. [27] H. Kuttruff, “Room acoustic-fifth edition, ” in Spon Press , 2009. [28] A. Farina, “Simultaneous measurement of impulse response and distortion with a swept-sine technique, ” in Pr oc. of the 108th AES Con vention , 2000, pp. 18–22. [29] M. Rav anelli, A. Sosi, P . Svaizer , and M. Omologo, “Impulse re- sponse estimation for robust speech recognition in a reverberant en vironment, ” in Proc. of EUSIPCO , 2012, pp. 1668–1672. [30] B. Angelini, F . Brugnara, D. Falavigna, D. Giuliani, R. Gretter, and M. Omologo, “Speaker independent continuous speech recog- nition using an acoustic-phonetic Italian corpus, ” in Pr oc. of IC- SLP , 1994, pp. 1391–1394. [31] R. Gretter, “Euronews: a multilingual speech corpus for ASR, ” in Pr oc. of LREC 2014 , 2014. [32] L. Cristoforetti, M. Rav anelli, M. Omologo, A. Sosi, A. Abad, M. Hagmueller , and P . Maragos, “The DIRHA simulated corpus, ” in Pr oc. of LREC 2014 , 2014, pp. 2629–2634. [33] P . Ghahremani, B. BabaAli, D. Pove y , K. Riedhammer, J. Trmal, and S. Khudanpur, “ A pitch extraction algorithm tuned for auto- matic speech recognition, ” in Pr oc. of ICASSP 2014 , 2014, pp. 2494–2498. [34] X. Lei, A. Senior , A. Gruenstein, and J. Sorensen, “ Accurate and compact large v ocabulary speech recognition on mobile de vices. ” in in Pr oc. of INTERSPEECH 2013 , 2013, pp. 662–665. [35] G. Chen, C. Parada, and G. Heigold, “Small-footprint keyword spotting using deep neural networks, ” in ICASSP 2014 , 2014, pp. 4087–4091. [36] A. W . Senior, G. Heigold, M. Bacchiani, and H. Liao, “GMM- free DNN acoustic model training, ” in ICASSP 20014 , 2014, pp. 5602–5606.