Automatic Speaker Recognition System in Adverse Conditions Implication of Noise and Reverberation on System Performance

Transcription

1 Autoatic Speaer Recognition Syste in Adverse Conditions Iplication of Noise and Reverberation on Syste Perforance Khais A. Al-Karawi, Ahed H. Al-Noori, Francis F. Li, and Ti Ritchings Abstract Speaer recognition has been developed and evolved over the past few decades into a supposedly ature technique. Existing ethods typically utilize robust features extracted fro clean speech. In real-world applications, especially security and forensics related ones, reliability of recognition becoes crucial, eanwhile liited speech saples and adverse acoustic conditions, ost notably noise and reverberation, ipose further coplications. This paper is presented fro a study into the behavior of typical speaer recognition systes in adverse retrieval phases. Following a brief review, a speaer recognition syste was ipleented using the MSR Identity Toolbox by Microsoft. Validation tests were carried out with clean speech and the speech containated by noise and/or reverberation of varying degrees. The iage source ethod was adopted to tae into account real acoustic conditions in the spaces. Statistical relationships between recognition accuracy and signal to noise ratios or reverberation ties have therefore been established. Results show noise and reverberation can, to different extents, degrade the perforance of recognition. Both reverberation tie and direct to reverberation ratio can affect recognition accuracy. The findings ay be used to estiate the accuracy of speaer recognition and further deterine the lielihood a particular speaer. Index Ters Clean speech, GMM-UBM, ISM, reverberation, robust speaer recognition, MFCC, MSR toolbox, noise. I. INTRODUCTION The perforance of speaer recognition can be affected by noise and reverberation and hence degraded. The variation in speech signals is caused by the environents, the speaer theselves and signal acquisition equipent. Robust speaer recognition in real world applications reains a technical challenge. While uch of the attention has been paid to channel variability, liited wor has been done to address the issue of roo acoustics and bacground noise in far-field of Autoatic Speaer Verification (ASV), particularly regarding roo reverberation and noise [1]. In far-field applications, it is nown that the signal captured by the icrophone involves the direct path signal, and a huge nuber of reflections off the walls, floor, and ceiling. Reverberation causes coloration of the speech signals and teporal spreading, which severely degrades the perforance of ost autoated speech technologies. To address this issue, Manuscript received Deceber 25, 2014; revised February 28, The authors are with School of Coputing Science and Engineering, University of Salford, UK (e-ail:.a.yousif@edu.salford.ac.u). various techniques have been proposed. Microphone arrays [1], score noralization [2], feature noralization [3] and alternative feature representations [4] have been suggested. The ipacts of reverberation on neural networ based speaer recognition has been studied in [5]. The effect of reverberation on speaer recognition systes using Gaussian ixture odels (GMM), hidden Marov odels (HMM) and quantization odels (VQ) has been tacled in [1]. Speaer verification using GMM with reverberation has been addressed in [6]. Several ethods for Speech enhanceent, such as spectral subtraction, have been investigated for noise-robust speaer recognition [7]. In the feature doain, for instance, techniques such as Cepstral Mean Subtraction (CMS) [1], Relative Spectral (RASTA) processing [2], and feature apping [3] have been utilized to reduce additive and convolutional channel distortions. Previous efforts were typically reistracted to siplistic and specific case reports with too liited inforation to forulate the relations between the syste perforance and acoustic conditions, e.g. [1] only reported 3eperical results. Rrecently, the coputational auditory scene analysis (CASA) has been engaged to reove noise [8]. In general the counity of speaer recognition has concentrated on channel variations in speaer verification. The National Institute of Standards and Technology (NIST) has conducted a series of speaer recognition evaluations (SRE) since State-of-the-art systes include joint factor analysis [9] and i-vector based techniques [10]. May [11] and Gonzalez-Rodriguez [1] studied the cobined ipacts utilizing binaural cues and icrophone arrays. Krishnaoorthy and Prasanna [12] described the results in noisy and reverberant conditions separately. In this paper, we further investigate the relations between noise and reverberation and the perforance of a typical speaer recognition syste through a large nuber of accurately siulated cases. This paper proceeds as follows: Section II, SV bacground; Section III, the role of noise and reverberation; Section IV, the speaer recognition syste; Section V, experients and results; Section VI, experient Result Discussion and Section VII, the concluding rears. II. SV BACKGROUND Speaer recognition is defined as the process of recognising the identity of a person by analysing their speech signals. Speaer recognition can be classified into identification and verification. Speaer identification is the process of deterining which registered speaer provides a given utterance. Speaer verification, on the other hand, is the DOI: /IJIEE.2015.V

2 process of accepting or rejecting the identity clai of a speaer. Speaer recognition ethods can also be divided into text-independent and text dependent ethods. Text-independence eans that the syste doesn t use any prior nowledge about the speech contents used in the training. Statistical approach is often used to odel speaers as well as to achieve verification [13]. Suppose X denotes the collection of feature vectors acquired fro the test data, and refers to the claied speaer identity which has a corresponding reference odel C. So the verification decision is given by: accept speaer, if d( C, x) T or reject speaer, if d( C, x) T. where T is the verification threshold, and d(c, x) represents soe distance easure between the test data and the reference odel C. The distance easure is coputed by finding the lielihood of the test data being created by the reference odel, which is given by: d c, x) log( p( x c )) (1) ( feature. The Cepstral features are ost coonly used fro this toolbox. The bac-end includes the training and testing odules. The training (enrollent) stage is responsible for generating odels for each register. In the test stage, the speech segent under testing is scored against all enrolled speaer odels to deterine if the speaer is the target speaer or just an iposter. The MSR identity toolbox provides two popular tools for speaer odelling the GMM-UBM and i-vector paradigs. In the GMM-UBM fraewor the universal bacground odel (UBM) represents Gaussian ixture odels (GMM) trained in a pool of data fro a large nuber of speaers. The speaer dependent odels are then adapted fro UBM using axiu a posteriori (MAP) estiation. Fig. 1. Illustrate the structure of the GMM-UBM odel. The Gaussian ixture odel (GMM)-based speaer identification algorith is popular due to its good perforance [17]. In this paper GMM -UBM is used. State-of-the-art autoatic speaer verification (ASV) systes, today, are based on the extensions to the joint factor analysis fraewor and constitute the so-called i-vectors, obtained after a total variability feature projection [10]. III. THE ROLE OF NOISE AND REVERBERATION Received speech signals by a icrophone can be odeled as the convolution between the speech signal and the roo ipulse response [14], the latter includes direct sound, early reflections and reverberation. When contributions fro reflections and reverberation are significant copared to the direct sound, the speech is said to be reverberated. Reverberation tie(t60 or RT) is the ain paraeter used to quantify reverberation [14], [15], which is the tie that it taes for the sound pressure in the roo to decay by 60dB after the source is switched off. The ipact of reverberation can be odeled as the processing of a signal by a linear tie invariant syste. Taing into account the bacground noise the received speech signals fro a icrophone can be written as: y( ) x( ) h( ) n( ) (2) where y() represent the received speech signals, x() is the original speech signal, h() denotes the roo ipulse response, and n() is the abient noise. IV. THE SPEAKER RECOGNITION SYSTEM A. MSR Toolbox Microsoft Speaer recognition (MSR) identity toolbox [16] is a Matlab toolbox developed by Microsoft Research which includes a collection of Matlab tools and functions to facilitate the developent of speaer recognition. It provides flexibility for researchers in developing new front-end and bac end techniques, allowing fast prototyping and rapid evaluation of new advanceent. The front-end of this toolbox is responsible for transforing the speech signals to acoustic Fig. 1. Bloc diagra of the GMM-UBM fraewor [16]. B. Iage Source Method The iage-source odel (ISM) [18] is a technique used to generate synthetic roo ipulse responses (RIRs). Once the RIR is available, reverberated speech saples can be siulated by converlution according to (2). The ISM used for typical roos were utilized to siulate the effects of reverberation, and rectangular roos were considered. The siulation progra taes as its input into four sets of values or diensions. The first set is the diensions of the roo, length, width and height, the second set is the source location, the third is the receiver location, and the last is 6 reflection coefficients for each surface of the roo. The ISM technique has a nuber of significant benefits copared to other approaches for roo acoustics siulation [19], it generates a large nuber of virtual roos for the study of the relations between reverberance and speaer recognition in this paper. As the absorption of sound depends upon frequency it is clear that the reverberation tie of an enclosure will also be frequency dependent, it is usual to estiate or esure reverberation tie in octive or third octave bands fro 125-4KHz [20]. Fig. 2 shows the process of ISM, in which a box represents the roo, the blac circle represents the position of the source, and the triangle represents the position of the receiver. 424

3 Fig. 2. Iage source ethod process. C. Siulated Reverberant Speech Iage source ethod developed in [21] and an ipleentation in Matlab by Eric A. Lehann [22] were used in this study to produce reverberated signals fro clean signals. The signals in Fig. 3 deonstrate reverberant speech signals produced by the ISM for T60=0.1, 0.5, and 1s. In the clean wavefor the sharp points relate to the consonants. When the reverberation increases. B. Baseline ASV: GMM-UBM The bloc diagra of the GMM-UBM baseline is illustrated in Fig. 1. MFCC features, calculated through a set of triangles (Mel) bandpass filters, were utilized. Cepstral coefficients, along with log-energy, delta coefficients were utilized to generate a 39-diensional feature vector fro pre-ephasized speech signal, and then the ean and variance noralized. With the GMM-UBM fraewor, Gaussian Mixture Model paraeters were acquired through the expectation-axiization (EM)algorith [13]. During enrollent, speaer odels were acquired through Maxiu a Posteriori (MAP) adaptation. Scoring and the decision was then perfored on log-lielihood thresholding. C. Reverberation Experient To quantify the relations between recognition perforance and reverberance, two ethods for were followed. First, the saples recorded in the Salford university reverberation roo. The MSR accuracy as using these saples in the testing phase against clean saples in the learning phase only achieved 10%. Fig. 4 Shows MSR toolbox process with siulation software. Fig. 4. MSR toolbox process with siulation software. Fig. 3. Wavefors, top to botto: clean and reverberant speech to T 60 =0.1,0.5,1s. A. Speech Database V. EXPERIMENTS AND RESULTS The speech saples were recorded using a Zoo H4 recorder. Speech saples were collected fro 19 speaers 11 ales and 8 feales for both noise and reverberation experients. Speaers were between year of age. Each speaer provided 5 clean speech saples and an additional one was recorded in reverberation condition (in a sall reverberation chaber at Salford University), Lengths of speech saples were between 30s and 40s. The utterances for all speaers in reverberation case included the sae text (text dependent), while each speaer spoe different text and language in the noise case. The speech is sapled at 16 KHz. The other ethod, as discussed in Section 4.2 was utilized to generate reverberation saples with iage source coputer siulation. A different roo diensions ( , and ), five different ties (0.1, 0.5, 1, 1.5 and 2s) were set with each roo. Furtherore, the distance between icrophone and source were also variable and three distances were used as shown in Table II. D. Noise Experient Three types of noises were used in the experients, white, pin and tonal noises. The noisy utterances were obtained by ixing noises with the clean speech with variable ratios according to (3). Fig. 5 illustrates the ixing process of speech and noise Mixingaudi o ( S NR) ( NS 100 NR) (3) where, S the is the speech signal, NR, the noise ratio and NS, is the noise signal. Nine speech saples are ixed with 3 types of noises to produce (297)speech over noise saples 425

4 distributed in 11 ixing groups for each type of noise depends on ixing percentage. Table I shows this process. TABLE I: SPEECH AND NOISE MIXING PERCENTAGE Speech Noise Fig. 5. Mixing of speech and noise with different ratio. VI. EXPERIMENT RESULT DISCUSSION A. Noise Experient Result Discussion The ai of this experient is to characterize accuracy of the MSR in different speech to noise ratio (SNR). Fig. 6 illustrates the degradation in recognition accuracy when noise increases. The X axis represents the ratio of speech over noise, while the Y axis represents the recognition accuracy of the syste. Regarding additive white noise, the syste accuracy drops rapidly fro 100% in clean speech to 30% when adding 10% noise. Then the accuracy of the syste continues to decrease to 10% when increasing the noise to a 70/30 speech over noise ratio. The pin noise degrades syste accuracy to approxiately 56% when ixing with 10% noise, and the syste accuracy continues to decrease to 0 with a 30/70 speech to noise ratio. On the other hand, tonal noise shows little effect copared with two other noises, but it still has a tangible effect on the accuracy of the syste, when ixing with a 40/60 speech to noise ratio, recognition accuracy becoes 55%. B. Reverberation Experient Result Discussion Reverberation was added to the speech utilizing the Iage source ethod [21]. Three reverberant ipulse responses were utilizing. The specifications of the roo used to generate the reverberant ipulses are shown in Table II. We first establish a set of baseline results using clean signals, i.e. training and testing with independent but clean speech saples. For the baseline the accuracy of MSR in verification case was 100%. However, the MSR results of reverberation saples which recorded in the described roo was only 10%. To further investigate the relations between reverberation and syste perforance, siulated roo ipulse responses are used. On the other hand, The direct-to-reverberant energy ratio (DRR) is also a significant paraeter. Moreover, there are soe applications that depend on DRR such as speech enhanceent at a specific distance, source distance estiation, derevereberation and spatial audio coding. Several ways are available for estiating DRR. The Direct to Reverberation Ratio (DRR) is defined as [23]: 2 h ( ) DDR 10 log 10 db 1 2 h ( ) 0( ) where, h(n) is the speaer-to-receiver ipulse response M, it is length in saples and δ the tie-index of the direct path in saples. MSR accuracy results for the siulation of the SV in reverb 1, reverb 2 and reverb 3 using clean training signals and reverberant test signals are shown in Fig. 7. TABLE II: REVERBERATION SPECIFICATIONS Specification Reverberation Model Reverb 1 Reverb 2 Reverb 3 Roo diensions Ro volue Mic. Position Source position Fixed RT , 0.5, 1, 1.5, 2 Second Walls reflection 0.5, 0.6, 0.1, 0.8 Ceiling reflection Carpeted floor reflection (4) Fig. 6. Perforance of speaer recognition in different speech to noise ratio. Fig. 7. MSR accuracy in three different reverberation roo. 426

5 According to the syste accuracy data fro clean and reverberated saples, a regression odel can be obtained RT DR 100 [ (1 RT )] (5) Rvolue where, DR denotes the degradation rate, RT denote reverberation tie and Rvoule denote roo volue. VII. CONCLUSION This paper investigates the relations between the accuracy of speaer recognition and adverse acoustic conditions. In particular a regression forula has been established to predict the recognition accuracy of a typical speaer recognition syste. Validation is left for future wor with ore roo types and speaer recognition engines. REFERENCES [1] J. González-Rodríguez, J. Ortega-García, C. Martín, and L. Hernández, Increasing robustness in GMM speaer recognition systes for noisy and reverberant speech with low coplexity icrophone arrays, in Proc. Fourth International Conference on Spoen Language, 1996, pp [2] I. Peer, B. Rafaely, and Y. Zigel, Reverberation atching for speaer recognition, in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing, 2008, pp [3] S. Ganapathy, J. Pelecanos, and M. K. Oar, Feature noralization for speaer verification in roo reverberation, in Proc IEEE International Conference on Acoustics, Speech and Signal Processing, 2011, pp [4] T. H. Fal and W.-Y. Chan, Modulation spectral features for robust far-field speaer identification, IEEE Transactions on Audio, Speech, and Language Processing, vol. 18, pp , [5] P. J. Castellano, S. Sradharan, and D. Cole, Speaer recognition in reverberant enclosures, in Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing, 1996, pp [6] I. A. McCowan, J. Pelecanos, and S. Sridharan, Robust speaer recognition using icrophone arrays, in Proc. a Speaer Odyssey-The Speaer Recognition Worshop, [7] W. Ning, P. C. Ching, Z. Nengheng, and L. Tan, Robust speaer recognition using denoised vocal source and vocal tract features, IEEE Transactions on Audio, Speech, and Language Processing, vol. 19, pp , [8] X. Zhao, Y. Shao, and D. Wang, CASA-based robust speaer identification, IEEE Transactions on Audio, Speech, and Language Processing, vol. 20, pp , [9] P. Kenny, Joint factor analysis of speaer and session variability: Theory and algoriths, CRIM, Montreal,(Report) CRIM-06/08-13, [10] N. Deha, P. Kenny, R. Deha, P. Duouchel, and P. Ouellet, Front-end factor analysis for speaer verification, IEEE Transactions on Audio, Speech, and Language Processing, vol. 19, pp , [11] T. May, S. van de Par, and A. Kohlrausch, A binaural scene analyzer for joint localization and recognition of speaers in the presence of interfering noise sources and reverberation, IEEE Transactions on Audio, Speech, and Language Processing, vol. 20, pp , [12] P. Krishnaoorthy and S. M. Prasanna, Application of cobined teporal and spectral processing ethods for speaer recognition under noisy, reverberant or ulti-speaer environents, Sadhana, vol. 34, pp , [13] D. A. Reynolds, T. F. Quatieri, and R. B. Dunn, Speaer verification using adapted Gaussian ixture odels, Digital Signal Processing, vol. 10, pp , [14] H. Kuttruff, Roo Acoustics, CRC Press, [15] M. R. Schroeder, New ethod of easuring reverberation tie, The Journal of the Acoustical Society of Aerica, vol. 37, pp , [16] S. O. Sadjadi, M. Slaney, and L. Hec, MSR identity toolbox - A atlab toolbox for speaer recognition research, Microsoft Research, Conversational Systes Research Center (CSRC), [17] W. M. Capbell, J. P. Capbell, D. A. Reynolds, E. Singer, and P. A. Torres-Carrasquillo, Support vector achines for speaer and language recognition, Coputer Speech & Language, vol. 20, pp , [18] S. Van Vuuren, Coparison of text-independent speaer recognition ethods on telephone speech with acoustic isatch, in Proc. of Fourth International Conference on Spoen Language, 1996, pp [19] E. A. Lehann and A. M. Johansson, Diffuse reverberation odel for efficient iage-source siulation of roo ipulse responses, IEEE Transactions on Audio, Speech, and Language Processing, vol. 18, pp , [20] M. Kahrs and K. Brandenburg, Applications of Digital Signal Processing to Audio and Acoustics, vol. 1, Springer, [21] J. B. Allen and D. A. Berley, Iage ethod for efficiently siulating sall roo acoustics, The Journal of the Acoustical Society of Aerica, vol. 65, pp , [22] E. A. Lehann. (October). [Online]. Available: [23] M. Tonelli, Blind reverberation cancellation techniques, Khais A. Al-Karawi received a BSc degree in coputer science fro Baghdad University, Iraq in He received an Mc.s in coputer science fro Pune University, India in He is currently studying towards a PhD at Salford University, United Kindo. Ahed H. Al-Noori received a BSc degree in coputer science fro Al-Nahrain University, Iraq in He also received an Mc.s in coputer science/genetic algorith-prograing fro Al-Nahrain University, Iraq in He is currently doing his PhD at Salford University, United Kindo. Francis F. Li received a B.Eng. fro the East China University of Science and Technology, an MPhil fro University of Brighton, and a PhD fro the University of Salford, UK. Francis is currently with the School of Coputing Science and Engineering, Salford University, where he teaches on BSc and MSc levels, supervises PhDs, and carries out research. His research interests include speech, usic and ultiedia signal processing; artificial intelligence; soft-coputing; architectural acoustics; data counications; bio-edical engineering; and instruentation. Francis published over 100 research papers and a boo. He is an associate editor in chief for SPIJ and an associate technical editor for J. AES. 427