SPEAKER IDENTIFICATION FROM SHOUTED SPEECH: ANALYSIS AND COMPENSATION

Transcription

1 SPEAKER IDENTIFICATION FROM SHOUTED SPEECH: ANALYSIS AND COMPENSATION Ceal Hanilçi 1,2, Toi Kinnunen 2, Rahi Saeidi 3, Jouni Pohjalainen 4, Paavo Alku 4, Figen Ertaş 1 1 Departent of Electronic Engineering Uludağ University, Bursa, Turkey 2 School of Coputing, University of Eastern Finland, Joensuu, Finland 3 Centre for Language and Speech Technology in Radboud University Nijegen, Netherlands 4 Departent of Signal Processing and Acoustics, Aalto University, Espoo, Finland chanilci@uludag.edu.tr, tkinnu@cs.joensuu.fi, rahi.saeidi@let.ru.nl ABSTRACT Text-independent speaker identification is studied using neutral and shouted speech in Finnish to analyze the effect of vocal ode isatch between training and test utterances. Standard el-frequency cepstral coefficient (MFCC) features with Gaussian ixture odel (GMM) recognizer are used for speaker identification. The results indicate that speaker identification accuracy reduces fro perfect ( %) to 8.71 % under vocal ode isatch. Because of this draatic degradation in recognition accuracy, we propose to use a joint density GMM apping technique for copensating the MFCC features. This apping is trained on a disjoint eotional speech corpus to create a copletely speaker- and speech ode independent eotion-neutralizing apping. As a result of the copensation, the 8.71 % identification accuracy increases to 32. % without degrading the non-isatched train-test conditions uch. Index Ters speaker identification, shouted speech 1. INTRODUCTION Research in both speech and speaker recognition has largely focused on noralizing out undesirable variations caused by transission channel and acoustic environent. Cobating for these technical nuisance factors has lead to any successful noralization techniques in feature [1], odel [2] and atch score doains [3]. A uch less studied proble, however, is that of intra-person variations caused by changes in the vocal production process itself. Of particular interest is variation in speaker s vocal effort. Vocal effort has a counicative purpose, such as an attept to conceal the speech content (whispering), increasing intelligibility in noisy environents (loud speech) or indicating eergency or other type of urgency (shouting). Speech in forensic speaker recognition and accident investigation is likely to have been produced under stress and is therefore cobined with high vocal effort. Even though differences between neutral and shouted/loud speech in traditional acoustic paraeters - forants, fundaental frequency and intensity - are well studied (e.g. [4, 5, 6]), the effect of vocal effort on autoatic speech and speaker recognition [4, 6, 7, 8] has received uch less attention. The question of how shouting affects between-speaker and within-speaker differences is not only relevant for forensic speaker recognition, but of fundaental nature that has iplications to other recognition applications as well. In the NIST speaker recognition evaluation (SRE) capaign, the effect of vocal effort on speaker recognition was analyzed [7]. In that study, speakers produced soft and loud utterances in a The work was supported by Acadey of Finland. controlled set-up. It was reported that isatched vocal effort between training and test (training with noral vocal effort and testing with high vocal effort) cause degradation of recognition accuracy. The authors of [9] have found that features extracted fro nasal syllables are relatively robust to high vocal effort. They have reported that, in context of a GMM-JFA recognizer on the NIST SRE corpus, the nasal constrained cepstral coefficients tend to bring advantage over using all cepstral coefficients. In [4], whispered speech was found to give the lowest identification rate and it was reported that 98.8 % identification accuracy obtained in neutral training-neutral test condition whereas in neutral training-shouted test case identification accuracy reduced to 56.3 %. An HMM based text-dependent speaker identification ethod for shouted speech was proposed in [8] and it was reported that identification accuracy decreases fro 96 % to 73 % when shouted speech is used for testing. In that study, the speech saples used in the experients were collected in different sessions and speaker odels were trained using neutral speech. In this study, we consider idealized speaker identification conditions where the typically included effects of channel isatch, environental noise and reverberation are copletely excluded. To this end, we consider closed-set speaker identification using Finnish utterances recorded in an anechoic chaber. This approach, iportantly, enables studying speaker identifiability solely under varying vocal odes; if one cannot correctly identify speakers even under such idealized setting, one should not expect accurate recognition under additional nuisance factors due to channel or environent. In order to study speaker identification in isatch conditions between neutral and shouted speech, a ethod based on joint density GMM apping is proposed to copensate the effect of shouting. To this end, we adopt ethods fro voice conversion [] typically used for speaker identity conversion to train a speaker-independent joint density Gaussian ixture odel apping on the MFCC feature space. This apping, intended to reove any expressive factors fro a given strea of MFCCs, is independently trained on a disjoint eotional Geran speech corpus including parallel recordings of neutral and eotionally colored speech saples. 2. NEUTRAL VS. SHOUTED SPEECH The authors of [4] categorize speech as having five different odes: whispered, soft, neutral, loud and shouted. Vocal intensity is lowest in whispered speech which is acoustically generated by an aperiodic weak excitation wavefor in the absence of the vocal fold vibration. Due to the lack of vocal fold vibration, whispered speech is the lowest vocal ode. Shouted speech, in turn, is the highest vocal ode /13/$ IEEE 827 ICASSP 13

2 Frequency (khz) Neutral Power Spectru (db) Neutral FFT LP Power Spectru (db) Shouted Frequency (khz) Shouted Tie (s) Fig. 1. Spectrogras and first three forants of neutral and shouted versions of the sae utterance spoken by a feale speaker. It calls for increased lung effort generating rapid period fluctuation of the vocal folds and a proinent voice excitation, which result in axial vocal intensity. Current speaker recognition studies ostly focus on neutral, norally spoken speech. A nuber of authors have analyzed acoustic differences of neutral and shouted speech. In [5], acoustic differences between noral and shouted speech were analyzed in forensic settings. In that study, it was found that the fundaental frequency (F ) and the first forant frequency (F 1) increase in shouting whereas the second and the third forants (F 2 andf 3) were less affected by shouting. In [4], different speech odes were analyzed in ters of the sound intensity level, duration and frae energies. It was found that the average sound intensity level of shouted speech is higher than that of neutral speech and sentence duration of shouted sentence is longer than neutral. Nuber of low energy fraes, on the other hand, is saller in shouting than in neutral speech. This is in line with [7] where statistically significant differences between the average energy levels of noral and high vocal effort utterances in NIST SRE were reported. In [6], eergency situation detection was studied for an indoor acoustic-based security syste and it was found that bothf 1 and F 2 and their standard deviations increase in shouting. Recognition of shouted speech was also considered and it was reported that word recognition accuracy decreases for shouted speech. Acoustic differences between noral and shouted speech can easily be seen fro spectrogras. Fig. 1 displays the wideband spectrogras and the first three forants (F 1-F 3) calculated using Praat 1 for neutral and shouted version of the sae utterance. As seen fro the figure, the forants (especially F 1) are shifted to higher frequencies in shouted speech. Differences in neutral and shouted speech are further described in Fig. 2, which show spectra of these two vocal odes coputed by fast Fourier transfor (FFT) and linear prediction (LP). Clearly, the shouted speech is characterized by sharper peaks in the spectral envelope. The spectral dissiilarities between neutral and shouted speech will affect MFCCs utilized in the feature extraction of speaker recognition resulting in a speech ode isatch between training and test. In the following, we propose a feature copensation to itigate for such isatches. 1 db 4 Frequency (Hz) Frequency (Hz) 4 Fig. 2. Power spectra of a voiced speech frae in neutral (F = 297 Hz) and shouted ode (F = 375 Hz). 3. SHOUT COMPENSATION To copensate the effect of different speech odes, voice conversion can be utilized to convert an utterance fro one ode to another. Voice conversion refers to ethodologies for converting one speaker s (source) utterances to given an ipression that they are spoken by another speaker (target) []. A voice conversion syste consists of two ain coponents, signal paraeterization and feature apping function. Signal paraetrization odel such as STRAIGHT [11] is used for analyzing (and synthesizing) utterances, whereas apping is used for learning a regression function between the vocal spaces of the source and the target speakers. As we do recognition rather than synthesis or conversion, we only consider the feature apping part. We directly plug-in our feature apping function to our recognizer MFCC front-end as will be detailed below. A generic feature apping function is denoted here by f Θ (x) : R d R d, where Θ denotes the odel paraeters and d is the diensionality of the acoustic vectors. In the training phase, the paraeters Θ are learnt fro a training set consisting of frae-aligned feature vector pairs{(x t,y t) t = 1,2,...,T}. To ensure that training utterances are phonetically aligned, they are usually taken to be parallel so that both the source and the target speakers read the sae sentences. Alignent of the feature vectors is achieved using dynaic tie warping (DTW). In the conversion phase which is copletely text-independent one applies ŷ t = f Θ (x t) for each source vector x t to find predicted target speaker vector y t for that observation. In this study, we adopt feature apping techniques fro voice conversion to copensate for shouted speech. To this end, nowx andy represent non-neutral and neutral vocal spaces of the sae speaker rather than two different speakers. We copensate non-neutral speech using Gaussian ixture odel (GMM) conversion [12]. In particular, we adopt the joint density GMM originally proposed in [13]. In this odel, the joint distribution of the source (non-neutral) and the target (neutral) features is odeled by GMMs trained using the stacked feature vectors z t = [x t,yt ] of diensionality2d. The joint probability density function is given by, where µ (z) = p(z t Θ (z) ) = [ µ (x) µ (y) ] M =1 P (z) N(z t µ (z),σ (z) ), and Σ (z) = [ Σ (xx) Σ (yx) Σ (xy) Σ (yy) the ean vector and covariance atrix of the ultivariate Gaussian density N(z t µ (z),σ (z) ), respectively, and P (z) are the prior probabilities constrained by P (z) and P(z) = 1. The joint odel paraeters are estiated to axiize likelihood for the training data set using the conventional expectation-axiization (EM) algorith [14]. In our ipleentation, we use full covariance atrices and 4 EM iterations starting fro randoized initial solution. Even though speaker recognition systes typically use di- ] are 828

3 Original neutral speech Frae nuber Neutral speech after copensation (4 Gaussians) Frae nuber Original shouted speech 4 Frae nuber Shouted speech after copensation (4 Gaussians) 4 Frae nuber Fig. 3. MFCCs of neutral, shouted and their copensated counterparts. agonal covariances, full covariances are coon in voice conversion. They capture cross-correlations across the source and the target spaces, while diagonal covariance (for all the four subatrices Σ (xx), Σ (xy), Σ (yx) and Σ (yy) ) iplies independent conversion of each cepstral coefficient. In preliinary tests, we ipleented both variants and, despite sall aount of training data, full covariance with less Gaussians outperfored systeatically all trialed diagonal conversions (up to 256 Gaussians). To reduce sensitivity to paraeter initialization, we repeat training ties, each starting fro a different rando guess, and pick the GMM which yields largest likelihood. Given the trained joint density odel, the predictor for future data points is, ŷ = f(x) = M =1 p (x)(µ (y) +Σ (yx) (Σ (xx) ) 1 (x µ (x) )), where p (x) = P N(x µ x,σ xx )/ k P kn(x µ x k,σ xx k ) denotes the posterior probability of x originating fro the th Gaussian. Fig. 3 shows the MFCCs of the sae utterance spoken with neutral and shouted speech odes and their copensated versions using 4 Gaussians, as an exaple. It can be seen that the variations between neutral and shouted speech odes of the sae utterance are highly reduced after copensation. 4. EXPERIMENTAL SETUP The speech corpus used in experients consists of 11 ale and 11 feale speakers. Each speaker produced 24 Finnish utterances using neutral speech ode. The sae 24 utterances were also produced with shouting. The sentences were recorded using a high-quality icrophone in an anechoic chaber so that the device, environental and channel effects are copletely excluded. The average duration of utterances is approxiately 3 seconds. Half of the sentences are in iperative and half in indicative ood. For ore details about the database, refer to [15]. In training the joint density GMM feature apping, we utilize the Berlin database of eotional speech [16] 2. This corpus consists of Geran speech saples fro ten speakers (5 ales and 5 feales) recorded also in an anechoic chaber. Each speaker produces 5 short and 5 longer sentences in seven different eotional odes: 2 neutral, anger, happiness, fear, boredo, disgust and sadness. Using this corpus, we train a speaker-independent feature apping that attepts to noralize out any eotional effects of a given speech utterance. To this end, we consider all the non-neutral utterances of a given speaker as our source utterances and the corresponding neutral utterance of that speaker as the target utterance. The DTW alignent is first coputed to the MFCC vectors within each speaker by using cosine siilarity as a vector siilarity easure so that all the non-neutral utterances utilize the corresponding neutral utterance as a target. Additional care is taken to exclude any-to-one and oneto-any assignents of the training vectors [17]. The aligned vector pairs fro all the speakers are then pooled and used for training a speaker-independent joint density GMM as detailed above. This feature apping is then applied to all training and test utterances in our evaluation set. Copensating for eotions rather than shouting is naturally a ore general proble setup. In fact, in preliinary experients, we trained only angry-to-neutral apping on the sae corpus as angriness is aong the seven eotions of [16] the one which corresponds best with shouting. However, since we have a rather liited training set with full covariance GMM odeling, including the other source eotions helped preventing nuerical probles in GMM training. For the sae reason (sall training set relative to the diensionality of the joint feature space), we also experient with two alternative feature appings. In the first approach, we train apping on base MFCC coefficients only and add the delta and double delta coefficients after feature apping. In the second approach, we train the apping function directly on the higher diensional MFCC features (see below). In the speaker identification experients, we use standard MFCCs extracted fro s Haing windowed speech fraes every s. We use two standard spectru estiation ethods, FFT and LP with prediction order of p =, to copute spectra of windowed fraes. The power spectra are processed through a 27-channel triangular filterbank. The logarithic filterbank outputs are converted into MFCCs by discrete cosine transfor. The first and second tie derivatives ( and 2 ) are appended to the first 16 MFCCs which leads to 48 diensional feature vectors. Finally, cepstral ean and variance noralization (CMVN) are applied to the features. Gaussian ixture odel (GMM) is used as the classifier. We use GMMs with 32 Gaussians trained by axiu likelihood (ML) criterion [14] using 5 EM iterations. We consider text-independent speaker identification in the experients. Due to relatively sall aount of data, the speaker identification experients are carried out using leave-one-out cross validation to axiize the nuber of test trials. That is, each speaker odel is trained using his/her 23 sentences and the held-out utterance is used for testing. Rotating over all 24 utterances and 22 speakers, this yields = 528 identification trials. In the experients we consider four different training and test conditions: Neutral - Neutral (N-N): Training and test utterances are both in neutral speech ode. Shouted - Shouted (S-S): Shouted speech is used in both training and test. Neutral - Shouted (N-S): Each speaker odel is trained using neutral speech and tested with shouted speech. Shouted - Neutral (S-N): Each speaker odel is trained using shouted speech and tested with neutral speech As the perforance criterion, we use identification accuracy, which is the ratio of the correctly identified trials to the total nuber of trials. 829

4 Table 1. Identification accuracy (%) for different speech odes using feature apping Training- Baseline: no Copensation applied to Test copensation MFCCs MFCCs+ + 2 condition FFT LP FFT LP FFT LP N-N S-S N-S S-N EXPERIMENTAL RESULTS We first analyze the perforance of the baseline speaker identification syste without any feature copensations. The identification accuracy for different scenarios and with different features are provided as the first two coluns of Table 1. In the atched vocal ode cases (N-N and S-S), both the FFT and LP spectru estiators yield high identification accuracies. In the case of the isatched vocal ode cases (N-S and S-N), both ethods degrade to unusable levels which confirs the general observation on previous studies on the topic. In the isatched cases, LP outperfors FFT. We next evaluated the shout copensation technique described in Section 3 using different nuber of Gaussian coponents. Fig. 4 shows the identification rates for the N-S and the S-N conditions using FFT spectru estiator. Feature apping iproves the identification rates considerably in coparison to the uncopensated baseline syste. Coparing the two types of feature appings, apping the full front-end (MFCC ) works generally slightly better. This ight be because the full front-end presents richer feature space and directly copensates also for the cepstral dynaics. Regarding the nuber of Gaussians, single Gaussian is not enough as expected. Using 32 Gaussians yields the highest identification accuracy for both the N-S and the S-N conditions. Identification rates using feature apping are given in Table 1. Feature apping iproves recognition accuracies for isatched odes (N-S and S-N) by a wide argin whereas identification rates decreases in coparison to the uncopensated baseline on the atched conditions (N-N and S-S). However, these relative degradations on N-N (5.5 %) and S-S conditions (2.48 %) are acceptable, given that the isatched vocal odes experience ipressive iproveents (for instance, around 4-fold increase for FFT in the N-S condition). In contrast to baseline perforances, now FFT outperfors LP in ost cases. Finally, the nuber of isidentified trials are given in Fig. 5 for FFT features before and after copensation. In the case of no copensation (baseline MFCC) the errors are uniforly distributed and for ost speakers all the 24 trials are isidentified. However, the copensation reduces the nuber of errors alost for every speaker. Fig. 5 reveals that, while shout copensation is successful for soe speakers (e.g., 11 and 12), it akes no difference for soe speakers (e.g. 8,, 14 and 19). The rest of the speakers fall in between these two extrees. There are two possible reasons for such behavior. Firstly, the shout copensation apping training set is both sall and language-isatched with our evaluation data, as no additional parallel Finnish shouted speech corpus was available. Secondly, being trained fro a pool of any speakers, the apping function does statistical averaging that ay reove speaker cues in addition to copensating shouting. 6. DISCUSSION We evaluated text-independent speaker identification using shouted speech and proposed a first step towards explicit shout copensa- Identification Rate (%) No Copensation Mapping (base features only) Mapping (base features and deltas) Neutral training Shouted test Nuber of Gaussian coponents Identification Rate (%) Shouted training Neutral test Nuber of Gaussian coponents Fig. 4. Identification rates for different nuber of Gaussians used for copensation. Nuber of isidentified trials Speaker Nuber of isidentified trials Speaker Fig. 5. Nuber of isidentified trials per each speaker for no copensation (left) and after copensation (right). tion using joint density GMM apping. Identification accuracy is reasonable when the training and test conditions are atched but large degradation on the recognition accuracy occurs in the case of isatched vocal odes. It was shown that this degradation on recognition accuracy can partly be copensated by training feature apping on the MFCCs. It is iportant to note that the proposed copensation apping is speaker-independent and was trained on a different set of speakers actually even different spoken language, due to the lack of Finnish data to train the apping function. While the authors of [4] uses a sall database which consists of 12 ale speakers with total of 48 identification trials, our results are in reasonable agreeent with the results of that study. However, the author in [8] reported saller degradation in shouted case and this is probably because text-dependent speaker identification were considered using a database of 5 speakers (25 ale and 25 feale speakers) and each speaker trained using 4 utterances (alost two ties ore than our training data) and identification experients carried out with 16 neutral and 36 shouted identification trials whereas in this study we have 528 identification trials. The results for N-N and S-S in Table 1 after applying the transforation reveals that the proposed transforation is soothing out soe speaker specific inforation fro MFCCs. This is also seen fro Fig. 3 where, by applying the transforation, ost of the MFCC fluctuations are softened for both neutral and shouted speech. On the other hand, the reason for iproved recognition accuracy in N-S and S-N condition after applying the proposed transforation could be also found in reduced isatch between neutral and shouted MFCCs as can be seen fro the second row of Fig CONCLUSION In this paper we evaluated the text-independent speaker identification using shouted speech. Four different training/test conditions have been analyzed and it has been found that recognition perforance of speaker identification is quite reasonable when the training and test conditions are atched but large degradation on the recognition accuracy occurs in the case of vocal effort isatch between training and testing. It was shown that this degradation on recognition accuracy can be partly copensated by applying the feature apping on the MFCCs. Future work should address how such apping could be trained ensuring that speaker features are retained. 83

5 8. REFERENCES [1] J. Pelecanos and S. Sridharan, Feature warping for robust speaker verification, in Proc. Odyssey 1, 1, pp [2] P. Kenny, Joint factor analysis of speaker and session variability : theory and algoriths, Technical Report, 5. [3] R. Auckenthaler, Score noralization for text-independent speaker verification systes, Dig. Sig. Proc., vol., no. 1-3, pp ,. [4] C. Zhang and J. H. L. Hansen, Analysis and classification of speech ode: whispered through shouted, in Proc. Interspeech 7, 7, pp [5] J. Elliot, Coparing the acoustic properties of noral and shouted speech: a study in forensic phonetics, in Proc. Aust. Int. Conf. Speech Sci. and Tech.,, pp [6] H. Nanjo, T. Nishiura, and H. Kawano, Acoustic-based security syste:towards robust understanding of eergency shout, in Proc. Int. Conf. Inf. Assurance and Sec., 9, pp [7] C. S. Greenberg, A. F. Martin, B. N. Barr, and G. R. Doddington, Report on perforance results in the NIST speaker recognition evaluation, in Proc. Interspeech 11, 11, pp [8] I. Shahin, Eploying second-order circular suprasegental hidden Markov odels to enhance speaker identification perforance in shouted talking environents, EURASIP J. on Audio, Speech and Music Proc., p. pages,. [9] N. Scheffer, L. Ferrer, M. Gracinarena, S. Kajarekar, E. Shriberg, and A. Stolcke, The SRI NIST speaker recognition evaluation syste, in Proc. ICASSP 11, 11, pp [] Y. Stylianou, Voice transforation: A survey, in ICASSP- 9, 9, pp [11] H. Kawahara, I. Masuda-Katsuse, and A. de Cheveigné, Restructuring speech representations using a pitch-adaptive tiefrequency soothing and an instantaneous-frequency-based f extraction: Possible role of a repetitive structure in sounds, Speech Counication, vol. 27, no. 3-4, pp , [12] Y. Stylianou, O. Cappé, and E. Moulines, Continuous probabilistic transfor for voice conversion, IEEE T. Speech and Audio Process, vol. 6, no. 2, pp , [13] A. Kain and M.W. Macon, Spectral voice conversion for textto-speech synthesis, in ICASSP 98, 1998, vol. 1, pp [14] D. A. Reynolds and R. C. Rose, Robust text-independent speaker identification using Gaussian ixture speaker odels, IEEE T. Speech and Audio Process., pp , [15] J. Pohjalainen, P. Alku, and T. Kinnunen, Shout detection in noise, in Proc. ICASSP 11, 11, pp [16] F. Burkhard, A. Paeschke, M. Rolfes, W. Sendleier, and B. Weiss, A database of Geran eotional speech, in Proc. Interspeech 5, 5, pp [17] E. Helander, J. Schwarz, J. Nurinen, H. Silen, and M. Gabbouj, On the ipact of alignent on voice conversion perforance, in Proc. Interspeech 8, 8, pp