Speech Communication 42 (2004) 271–287 www.elsevier.com/locate/specom Efficient voice activity detection algorithms using long-term speech information
Javier Ram
ırez *, Jos
e C. Segura 1, Carmen Ben
ıtez, Angel de la Torre, 2 Antonio Rubio Dpto. Electr
onica y Tecnolog
ıa de Computadores, Universidad de Granada, Campus Universitario Fuentenueva, 18071 Granada, Spain Received 5 May 2003; received in revised form 8 October 2003; accepted 8 October 2003 Abstract Currently, there are technology barriers inhibiting speech processing systems working under extreme noisy conditions. The emerging applications of speech technology, especially in the fields of wireless communications, digital hearing aids or speech recognition, are examples of such systems and often require a noise reduction technique operating in combination with a precise voice activity detector (VAD). This paper presents a new VAD algorithm for improving speech detection robustness in noisy environments and the performance of speech recognition systems. The algorithm measures the long-term spectral divergence (LTSD) between speech and noise and formulates the speech/ non-speech decision rule by comparing the long-term spectral envelope to the average noise spectrum, thus yielding a high discriminating decision rule and minimizing the average number of decision errors. The decision threshold is adapted to the measured noise energy while a controlled hang-over is activated only when the observed signal-to-noise ratio is low. It is shown by conducting an analysis of the speech/non-speech LTSD distributions that using long-term information about speech signals is beneficial for VAD. The proposed algorithm is compared to the most commonly used VADs in the field, in terms of speech/non-speech discrimination and in terms of recognition performance when the VAD is used for an automatic speech recognition system. Experimental results demonstrate a sustained advantage over standard VADs such as G.729 and adaptive multi-rate (AMR) which were used as a reference, and over the VADs of the advanced front-end for distributed speech recognition. Ó 2003 Elsevier B.V. All rights reserved. Keywords: Speech/non-speech detection; Speech enhancement; Speech recognition; Long-term spectral envelope; Long-term spectral divergence 1. Introduction * Corresponding author. Tel.: +34-958243271; fax: +34958243230. E-mail addresses: javierrp@ugr.es (J. Ram
ırez), segura@ ugr.es (J.C. Segura), carmen@ugr.es (C. Ben
ıtez), atv@ugr.es
de la Torre), rubio@ugr.es (A. Rubio). (A. 1 Tel.: +34-958243283; fax: +34-958243230. 2 Tel.: +34-958243193; fax: +34-958243230. An important problem in many areas of speech processing is the determination of presence of speech periods in a given signal. This task can be identified as a statistical hypothesis problem and its purpose is the determination to which category or class a given signal belongs. The decision is made based on an observation vector, frequently 0167-6393/$ - see front matter Ó 2003 Elsevier B.V. All rights reserved. doi:10.1016/j.specom.2003.10.002 272 J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 Nomenclature AFE AMR DSR DTX ETSI advanced front end adaptive multi-rate distributed speech recognition discontinuous transmission European Telecommunication Standards Institute FAR0 speech false alarm rate FAR1 non-speech false alarm rate FD frame-dropping GSM global system for mobile communications HR0 non-speech hit-rate HR1 speech hit-rate HM high mismatch training/test mode HMM hidden Markov model called feature vector, which serves as the input to a decision rule that assigns a sample vector to one of the given classes. The classification task is often not as trivial as it appears since the increasing level of background noise degrades the classifier effectiveness, thus leading to numerous detection errors. The emerging applications of speech technologies (particularly in mobile communications, robust speech recognition or digital hearing aid devices) often require a noise reduction scheme working in combination with a precise voice activity detector (VAD) (Bouquin-Jeannes and Faucon, 1994, 1995). During the last decade numerous researchers have studied different strategies for detecting speech in noise and the influence of the VAD decision on speech processing systems (Freeman et al., 1989; ITU, 1996; Sohn and Sung, 1998; ETSI, 1999; Marzinzik and Kollmeier, 2002; Sangwan et al., 2002; Karray and Martin, 2003). Most authors reporting on noise reduction refer to speech pause detection when dealing with the problem of noise estimation. The non-speech detection algorithm is an important and sensitive part of most of the existing singlemicrophone noise reduction schemes. There exist well known noise suppression algorithms (Berouti et al., 1979; Boll, 1979), such as Wiener filtering HTK ITU LPC LTSE LTSD MM ROC SDC SND SNR VAD WF WAcc WM ZCR hidden Markov model toolkit International Telecommunication Union linear prediction coding coefficients long-term spectral estimation long-term spectral divergence medium mismatch training/test mode receiver operating characteristic SpeechDat-Car speech/non-speech detection signal-to-noise ratio voice activity detector (detection) wiener filtering word accuracy well matched training/test mode zero-crossing rate (WF) or spectral subtraction, that are widely used for robust speech recognition, and for which, the VAD is critical in attaining a high level of performance. These techniques estimate the noise spectrum during non-speech periods in order to compensate its harmful effect on the speech signal. Thus, the VAD is more critical for non-stationary noise environments since it is needed to update the constantly varying noise statistics affecting a misclassification error strongly to the system performance. In order to palliate the importance of the VAD in a noise suppression systems Martin proposed an algorithm (Martin, 1993) that continually updated the noise spectrum in order to prevent a misclassification of the speech signal causes a degradation of the enhanced signal. These techniques are faster in updating the noise but usually capture signal energy during speech periods, thus degrading the quality of the compensated speech signal. In this way, it is clearly better using an efficient VAD for most of the noise suppression systems and applications. VADs are employed in many areas of speech processing. Recently, various voice activity detection procedures have been described in the literature for several applications including mobile communication services (Freeman et al., 1989), real-time speech transmission on the Internet J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 (Sangwan et al., 2002) or noise reduction for digital hearing aid devices (Itoh and Mizushima, 1997). Interest of research has focused on the development of robust algorithms, with special attention being paid to the study and derivation of noise robust features and decision rules. Sohn and Sung (1998) presented an algorithm that uses a novel noise spectrum adaptation employing soft decision techniques. The decision rule was derived from the generalized likelihood ratio test by assuming that the noise statistics are known a priori. An enhanced version (Sohn et al., 1999) of the original VAD was derived with the addition of a hang-over scheme which considers the previous observations of a first-order Markov process modeling speech occurrences. The algorithm outperformed or at least was comparable to the G.729B VAD (ITU, 1996) in terms of speech detection and false-alarm probabilities. Other researchers presented improvements over the algorithm proposed by Sohn et al. (1999). Cho et al. (2001a); Cho and Kondoz (2001) presented a smoothed likelihood ratio test to alleviate the detection errors, yielding better results than G.729B and comparable performance to adaptive multi-rate (AMR) option 2. Cho et al. (2001b) also proposed a mixed decision-based noise adaptation yielding better results than the soft decision noise adaptation technique reported by Sohn and Sung (1998). Recently, a new standard incorporating noise suppression methods has been approved by the European Telecommunication Standards Institute (ETSI) for feature extraction and distributed speech recognition (DSR). The so-called advanced front-end (AFE) (ETSI, 2002) incorporates an energy-based VAD (WF AFE VAD) for estimating the noise spectrum in Wiener filtering speech enhancement, and a different VAD for nonspeech frame dropping (FD AFE VAD). On the other hand, a VAD achieves silence compression in modern mobile telecommunication systems reducing the average bit rate by using the discontinuous transmission (DTX) mode. Many practical applications, such as the global system for mobile communications (GSM) telephony, use silence detection and comfort noise injection for higher coding efficiency. The International Telecommunication Union (ITU) adopted a toll- 273 quality speech coding algorithm known as G.729 to work in combination with a VAD module in DTX mode. The recommendation G.729 Annex B (ITU, 1996) uses a feature vector consisting of the linear prediction (LP) spectrum, the full-band energy, the low-band (0–1 KHz) energy and the zerocrossing rate (ZCR). The standard was developed with the collaboration of researchers from France Telecom, the University of Sherbrooke, NTT and AT&T Bell Labs and the effectiveness of the VAD was evaluated in terms of subjective speech quality and bit rate savings (Benyassine et al., 1997). Objective performance tests were also conducted by hand-labelling a large speech database and assessing the correct identification of voiced, unvoiced, silence and transition periods. Another standard for DTX is the ETSI adaptive multi-rate speech coder (ETSI, 1999) developed by the special mobile group for the GSM system. The standard specifies two options for the VAD to be used within the digital cellular telecommunications system. In option 1, the signal is passed through a filterbank and the level of signal in each band is calculated. A measure of the signal-to-signal ratio (SNR) is used to make the VAD decision together with the output of a pitch detector, a tone detector and the correlated complex signal analysis module. An enhanced version of the original VAD is the AMR option 2 VAD. It uses parameters of the speech encoder being more robust against environmental noise than AMR1 and G.729. These VADs have been used extensively in the open literature as a reference for assessing the performance of new algorithms. Marzinzik and Kollmeier (2002) proposed a new VAD algorithm for noise spectrum estimation based on tracking the power envelope dynamics. The algorithm was compared to the G.729 VAD by means of the receiver operating characteristic (ROC) curves showing a reduction in the non-speech false alarm rate together with an increase of the non-speech hit rate for a representative set of noises and conditions. Beritelli et al. (1998) proposed a fuzzy VAD with a pattern matching block consisting of a set of six fuzzy rules. The comparison was made using objective, psychoacoustic, and subjective parameters being G.729 and AMR VADs used as a reference (Beritelli et al., 2002). Nemer et al. (2001) 274 J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 presented a robust algorithm based on higher order statistics (HOS) in the linear prediction coding coefficients (LPC) residual domain. Its performance was compared to the ITU-T G.729 B VAD in various noise conditions, and quantified using the probability of correct and false classifications. The selection of an adequate feature vector for signal detection and a robust decision rule is a challenging problem that affects the performance of VADs working under noise conditions. Most algorithms are effective in numerous applications but often cause detection errors mainly due to the loss of discriminating power of the decision rule at low SNR levels (ITU, 1996; ETSI, 1999). For example, a simple energy level detector can work satisfactorily in high signal-to-noise ratio conditions, but would fail significantly when the SNR drops. Several algorithms have been proposed in order to palliate these drawbacks by means of the definition of more robust decision rules. This paper explores a new alternative towards improving speech detection robustness in adverse environments and the performance of speech recognition systems. A new technique for speech/ non-speech detection (SND) using long-term information about the speech signal is studied. The algorithm is evaluated in the context of the AURORA project (Hirsch and Pearce, 2000; ETSI, 2000), and the recently approved Advanced Front-end standard (ETSI, 2002) for distributed speech recognition. The quantifiable benefits of this approach are assessed by means of an exhaustive performance analysis conducted on the AURORA TIdigits (Hirsch and Pearce, 2000) and SpeechDat-Car (SDC) (Moreno et al., 2000; Nokia, 2000; Texas Instruments, 2001) databases, with standard VADs such as the ITU G.729 (ITU, 1996), ETSI AMR (ETSI, 1999) and AFE (ETSI, 2002) used as a reference. and Sung, 1998; Nemer et al., 2001). Typically, these features represent the variations in energy levels or spectral difference between noise and speech. The most discriminating parameters in speech detection are the signal energy, zero-crossing rates, periodicity measures, the entropy, or linear predictive coding coefficients. The proposed speech/non-speech detection algorithm assumes that the most significant information for detecting voice activity on a noisy speech signal remains on the time-varying signal spectrum magnitude. It uses a long-term speech window instead of instantaneous values of the spectrum to track the spectral envelope and is based on the estimation of the so-called long-term spectral envelope (LTSE). The decision rule is then formulated in terms of the long-term spectral divergence (LTSD) between speech and noise. The motivations for the proposed strategy will be clarified by studying the distributions of the LTSD as a function of the long-term window length and the misclassification errors of speech and non-speech segments. 2.1. Definitions of the LTSE and LTSD Let xðnÞ be a noisy speech signal that is segmented into overlapped frames and, X ðk; lÞ its amplitude spectrum for the k band at frame l. The N -order long-term spectral envelope is defined as LTSEN ðk; lÞ ¼ maxfX ðk; l þ jÞgj¼þN j¼N The N -order long-term spectral divergence between speech and noise is defined as the deviation of the LTSE respect to the average noise spectrum magnitude N ðkÞ for the k band, k ¼ 0; 1; . . . ; NFFT  1, and is given by LTSDN ðlÞ ¼ 10 log10 2. VAD based on the long-term spectral divergence VADs are generally characterized by the feature selection, noise estimation and classification methods. Various features and combinations of features have been proposed to be used in VAD algorithms (ITU, 1996; Beritelli et al., 1998; Sohn ð1Þ NFFT1 X LTSE2 ðk; lÞ 1 NFFT k¼0 N 2 ðkÞ ! ð2Þ It will be shown in the rest of the paper that the LTSD is a robust feature defined as a long-term spectral distance measure between speech and noise. It will also be demonstrated that using longterm speech information increases the speech detection robustness in adverse environments and, 275 J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 when compared to VAD algorithms based on instantaneous measures of the SNR level, it will enable formulating noise robust decision rules with improved speech/non-speech discrimination. 2.2. LTSD distributions of speech and silence In this section we study the distributions of the LTSD as a function of the long-term window length (N ) in order to clarify the motivations for the algorithm proposed. A hand-labelled version of the Spanish SDC database was used in the analysis. This database contains recordings from close-talking and distant microphones at different driving conditions: (a) stopped car, motor running, (b) town traffic, low speed, rough road and (c) high speed, good road. The most unfavourable noise environment (i.e. high speed, good road) was selected and recordings from the distant microphone were considered. Thus, the N -order LTSD was measured during speech and non-speech periods, and the histogram and probability distri- butions were built. The 8 kHz input signal was decomposed into overlapping frames with a 10-ms window shift. Fig. 1 shows the LTSD distributions of speech and noise for N ¼ 0, 3, 6 and 9. It is derived from Fig. 1 that speech and noise distributions are better separated when increasing the order of the long-term window. The noise is highly confined and exhibits a reduced variance, thus leading to high non-speech hit rates. This fact can be corroborated by calculating the classification error of speech and noise for an optimal Bayes classifier. Fig. 2 shows the classification errors as a function of the window length N . The speech classification error is approximately reduced by half from 22% to 9% when the order of the VAD is increased from 0 to 6 frames. This is motivated by the separation of the LTSD distributions that takes place when N is increased as shown in Fig. 1. On the other hand, the increased speech detection robustness is only prejudiced by a moderate increase in the speech detection error. According to Fig. 2, the optimal value of the order of the VAD N= 0 N= 3 Noise Speech 0.2 0.15 0.15 0.1 0.05 0.05 0 10 20 0 30 N= 6 N= 9 Noise Speech P(LTSD) 0.05 0.05 LTSD Noise Speech 0.15 0.1 20 30 30 0.2 0.1 10 20 0.25 0.15 0 10 LTSD 0.2 0 0 LTSD 0.25 P(LTSD) 0.2 0.1 0 Noise Speech 0.25 P(LTSD) P(LTSD) 0.25 0 0 10 20 LTSD Fig. 1. Effect of the window length on the speech/non-speech LTSD distributions. 30 276 J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 Classification error of Speech and NonSpeech periods 0.22 Initialization 0.2 - Noise spectrum estimation - Optimal threshold calculation NonSpeech Detection Error Speech Detection Error Error 0.18 Signal segmentation 0.16 0.14 Estimation of the LTSE 0.12 0.1 Computation of the LTSD 0.08 VAD= 1 0 1 2 3 4 5 6 7 8 9 10 LTSD N < 0.06 Decision rule LTSD< Fig. 2. Speech and non-speech detection error as a function of the window length. VAD= 1 would be N ¼ 6. As a conclusion, the use of longterm spectral divergence is beneficial for VAD since it reduces importantly misclassification errors. Hang-over scheme Noise spectrum update VAD= 0 2.3. Definition of the LTSD VAD algorithm A flowchart diagram of the proposed VAD algorithm is shown in Fig. 3. The algorithm can be described as follows. During a short initialization period, the mean noise spectrum N ðkÞ (k ¼ 0; 1; . . . ; NFFT  1) is estimated by averaging the noise spectrum magnitude. After the initialization period, the LTSE VAD algorithm decomposes the input utterance into overlapped frames being their spectrum, namely X ðk; lÞ, processed by means of a ð2N þ 1Þ-frame window. The LTSD is obtained by computing the LTSE by means of Eq. (1). The VAD decision rule is based on the LTSD calculated using Eq. (2) as the deviation of the LTSE with respect to the noise spectrum. Thus, the algorithm has an N -frame delay since it makes a decision for the l-th frame using a (2N þ 1)-frame window around the l-th frame. On the other hand, the first N frames of each utterance are assumed to be non-speech periods being used for the initialization of the algorithm. The LTSD defined by Eq. (2) is a biased magnitude and needs to be compensated by a given offset. This value depends on the noise spectral Fig. 3. Flowchart diagram of the proposed LTSE algorithm for voice activity detection. variance and the order of the VAD and can be estimated during the initialization period or assumed to take a fixed value. The VAD makes the SND by comparing the unbiased LTSD to an adaptive threshold c. The detection threshold is adapted to the observed noise energy E. It is assumed that the system will work at different noisy conditions characterized by the energy of the background noise. Optimal thresholds c0 and c1 can be determined for the system working in the cleanest and noisiest conditions. These thresholds define a linear VAD calibration curve that is used during the initialization period for selecting an adequate threshold c as a function of the noise energy E: 8 E 6 E0 > < c0 c¼ c0 c1 E E > : 0 1 c1 c0 c1 E þ c0  1E 1 =E0 E0 < E < E 1 E P E1 ð3Þ where E0 and E1 are the energies of the background noise for the cleanest and noisiest condi- 277 J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 tions that can be determined examining the speech databases being used. A high speech/non-speech discrimination is ensured with this model since silence detection is improved at high and medium SNR levels while maintaining a high precision detecting speech periods under high noise conditions. The VAD is defined to be adaptive to timevarying noise environments with the following algorithm for updating the noise spectrum N ðkÞ during non-speech periods being used: 8 aN ðk; l  1Þ þ ð1  aÞNK ðkÞ > > < if speech pause is detected ð4Þ N ðk; lÞ ¼ N ðk; l  1Þ > > : otherwise where NK is the average spectrum magnitude over a K-frame neighbourhood: NK ðkÞ ¼ 1 2K þ 1 K X X ðk; l þ jÞ ð5Þ j¼K Finally, a hangover was found to be beneficial to maintain a high accuracy detecting speech periods at low SNR levels. Thus, the VAD delays the speech to non-speech transition in order to prevent low-energy word endings being misclassified as silence. On the other hand, if the LTSD achieves a given threshold LTSD0 the hangover mechanism is turned off to improve non-speech detection when the noise level is low. Thus, the LTSE VAD yields an excellent classification of speech and pause periods. Examples of the operation of the LTSE VAD on an utterance of the Spanish SDC database are shown in Fig. 4a (N ¼ 6) and Fig. 4b (N ¼ 0). The use of a long-term window for formulating the decision rule reports quantifiable benefits in speech/non-speech detection. It can be seen that using a 6-frame window reduces the variability of the LTSD in the absence of speech, thus yielding to reduced noise variance and better speech/non-speech discrimination. Speech detection is not affected by the smoothing process involved in the long-term spectral estimation algorithm and maintains good margins that correctly separate speech and pauses. On the other hand, the inherent anticipation of the VAD decision contributes to reduce speech clipping errors. 3. Experimental framework Several experiments are commonly conducted to evaluate the performance of VAD algorithms. The analysis is normally focused on the determination of misclassification errors at different SNR levels (Beritelli et al., 2002; Marzinzik and Kollmeier, 2002), and the influence of the VAD zdecision on speech processing systems (BouquinJeannes and Faucon, 1995; Karray and Martin, 2003). The experimental framework and the objective performance tests conducted to evaluate the proposed algorithm are described in this section. 3.1. Speech/non-speech discrimination analysis First, the proposed VAD was evaluated in terms of the ability to discriminate between speech and pause periods at different SNR levels. The original AURORA-2 database (Hirsch and Pearce, 2000) was used in this analysis since it uses the clean TIdigits database consisting of sequences of up to seven connected digits spoken by American English talkers as source speech, and a selection of eight different real-world noises that have been artificially added to the speech at SNRs of 20, 15, 10, 5, 0 and )5 dB. These noisy signals have been recorded at different places (suburban train, crowd of people (babble), car, exhibition hall, restaurant, street, airport and train station), and were selected to represent the most probable application scenarios for telecommunication terminals. In the discrimination analysis, the clean TIdigits database was used to manually label each utterance as speech or non-speech frames for reference. Detection performance as a function of the SNR was assessed in terms of the non-speech hitrate (HR0) and the speech hit-rate (HR1) defined as the fraction of all actual pause or speech frames that are correctly detected as pause or speech frames, respectively: HR0 ¼ N0;0 N0ref HR1 ¼ N1;1 N1ref ð6Þ where N0ref and N1ref are the number of real nonspeech and speech frames in the whole database, 278 J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 30 LTSD 25 20 15 10 5 0 0 200 400 600 800 1000 1200 frame 4 3 x 10 2 x(n) 1 0 –1 –2 0 1 2 3 4 5 6 7 8 9 4 n (a) x 10 30 LTSD 20 10 0 –10 0 200 400 600 800 1000 1200 frame 4 3 x 10 x(n) 2 1 0 –1 –2 0 1 2 3 4 (b) 5 n 6 7 8 9 4 x 10 Fig. 4. VAD output for an utterance of the Spanish SpeechDat-Car database (recording conditions: high speed, good road, distant microphone). (a) N ¼ 6, (b) N ¼ 0. respectively, while N0;0 and N1;1 are the number of non-speech and speech frames correctly classified. The LTSE VAD decomposes the input signal sample at 8 kHz into overlapping frames with a 10-ms shift. Thus, a 13-frame long-term window and NFFT ¼ 256 was found to be good choices for the noise conditions being studied. Optimal detection threshold c0 ¼ 6 dB and c1 ¼ 2:5 dB J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 were determined for clean and noisy conditions, respectively, while the threshold calibration curve was defined between E0 ¼ 30 dB (low noise energy) and E1 ¼ 50 dB (high noise energy). The hangover mechanism delays the speech to non-speech VAD transition during 8 frames while it is deactivated when the LTSD exceeds 25 dB. The offset is fixed and equal to 5 dB. Finally, it is used a forgotten factor a ¼ 0:95, and a 3-frame neighbourhood (K ¼ 3) for the noise update algorithm. Fig. 5 provides the results of this analysis and compares the proposed LTSE VAD algorithm to standard G.729, AMR and AFE VADs in terms of non-speech hit-rate (Fig. 5a) and speech hit-rate (Fig. 5b) for clean conditions and SNR levels ranging from 20 to )5 dB. Note that results for the two VADs defined in the AFE DSR standard (ETSI, 2002) for estimating the noise spectrum in the Wiener filtering stage and non-speech framedropping are provided. Note that the results shown in Fig. 5 are averaged values for the entire set of noises. Thus, the following conclusions can 100 LTSE G.729 AMR1 AMR2 AFE (FD) AFE (WF) 90 HR0 (%) 80 70 60 50 40 30 20 10 (a) 0 Clean 20 dB 15 dB 10 dB 5 dB 0 dB -5 dB SNR 100 95 HR1 (%) 90 85 80 LTSE G.729 AMR1 AMR2 AFE (FD) AFE (WF) 75 70 65 60 55 (b) 50 Clean 20 dB 15 dB 10 dB 5 dB 0 dB -5 dB SNR Fig. 5. Speech/non-speech discrimination analysis: (a) nonspeech hit-rate (HR0), (b) speech hit rate (HR1). 279 be derived from Fig. 5 about the behaviour of the different VADs analysed: (i) G.729 VAD suffers poor speech detection accuracy with the increasing noise level while nonspeech detection is good in clean conditions (85%) and poor (20%) in noisy conditions. (ii) AMR1 yields an extreme conservative behaviour with high speech detection accuracy for the whole range of SNR levels but very poor non-speech detection results at increasing noise levels. Although AMR1 seems to be well suited for speech detection at unfavourable noise conditions, its extremely conservative behaviour degrades its non-speech detection accuracy being HR0 less than 10% below 10 dB, making it less useful in a practical speech processing system. (iii) AMR2 leads to considerable improvements over G.729 and AMR1 yielding better nonspeech detection accuracy while still suffering fast degradation of the speech detection ability at unfavourable noisy conditions. (iv) The VAD used in the AFE standard for estimating the noise spectrum in the Wiener filtering stage is based in the full energy band and yields a poor speech detection performance with a fast decay of the speech hit-rate at low SNR values. On the other hand, the VAD used in the AFE for frame-dropping achieves a high accuracy in speech detection but moderate results in non-speech detection. (v) LTSE achieves the best compromise among the different VADs tested. It obtains a good behaviour in detecting non-speech periods as well as exhibits a slow decay in performance at unfavourable noise conditions in speech detection. Table 1 summarizes the advantages provided by the LTSE-based VAD over the different VAD methods being evaluated by comparing them in terms of the average speech/non-speech hit-rates. LTSE yields a 47.28% HR0 average value, while the G.729, AMR1, AMR2, WF and FD AFE VADs yield 31.77%, 31.31%, 42.77%, 57.68% and 28.74%, respectively. On the other hand, LTSE attains a 98.15% HR1 average value in speech 280 J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 Table 1 Average speech/non-speech hit rates for SNR levels ranging from clean conditions to )5 dB VAD G.729 AMR1 AMR2 AFE (WF) AFE (FD) LTSE HR0 (%) HR1 (%) 31.77 93.00 31.31 98.18 42.77 93.76 57.68 88.72 28.74 97.70 47.28 98.15 detection while G.729, AMR1, AMR2, WF and FD AFE VADs provide 93.00%, 98.18%, 93.76%, 88.72% and 97.70%, respectively. Frequently VADs avoid losing speech periods leading to an extremely conservative behaviour in detecting speech pauses (for instance, the AMR1 VAD). Thus, in order to correctly describe the VAD performance, both parameters have to be considered. Thus, considering together speech and nonspeech hit-rates, the proposed VAD yielded the best results when compared to the most representative VADs analysed. 3.2. Receiver operating characteristic curves An additional test was conducted to compare speech detection performance by means of the ROC curves (Madisetti and Williams, 1999), a frequently used methodology in communications based on the hit and error detection probabilities (Marzinzik and Kollmeier, 2002), that completely describes the VAD error rate. The AURORA subset of the original Spanish SDC database (Moreno et al., 2000) was used in this analysis. This database contains 4914 recordings using close-talking and distant microphones from more than 160 speakers. As in the whole SDC database, the files are categorized into three noisy conditions: quiet, low noisy and highly noisy conditions, which represent different driving conditions and average SNR values of 12, 9 and 5 dB. The non-speech hit rate (HR0) and the false alarm rate (FAR0 ¼ 100 ) HR1) were determined in each noise condition for the proposed LTSE VAD and the G.729, AMR1, AMR2, and AFE VADs, which were used as a reference. For the calculation of the false-alarm rate as well as the hit rate, the ‘‘real’’ speech frames and ‘‘real’’ speech pauses were determined by hand-labelling the database on the close-talking microphone. The non-speech hit rate (HR0) as a function of the false alarm rate (FAR0 ¼ 100 ) HR1) for 0 < c 6 10 dB is shown in Fig. 6 for recordings from the distant microphone in quiet, low and high noisy conditions. The working point of the adaptive LTSE, G.729, AMR and the recently approved AFE VADs (ETSI, 2002) are also included. It can be derived from these plots that: (i) The working point of the G.729 VAD shifts to the right in the ROC space with decreasing SNR, while the proposed algorithm is less affected by the increasing level of background noise. (ii) AMR1 VAD works on a low false alarm rate point of the ROC space but it exhibits poor non-speech hit rate. (iii) AMR2 VAD yields clear advantages over G.729 and AMR1 exhibiting important reduction in the false alarm rate when compared to G.729 and increase in the non-speech hit rate over AMR1. (iv) WF AFE VAD yields good non-speech detection accuracy but works on a high false alarm rate point on the ROC space. It suffers rapid performance degradation when the driving conditions get noisier. On the other hand, FD AFE VAD has been planned to be conservative since it is only used in the DSR standard for frame-dropping. Thus, it exhibits poor non-speech detection accuracy working on a low false alarm rate point of the ROC space. (v) LTSE VAD yields the lowest false alarm rate for a fixed non-speech hit rate and also, the highest non-speech hit rate for a given false alarm rate. The ability of the adaptive LTSE VAD to tune the detection threshold by means the algorithm described in Eq. (3) enables working on the optimal point of the ROC curve for different noisy conditions. Thus, the algorithm automatically selects the J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 G.729 VAD. The label ½ð30; 8Þð50; 3:25Þ indicates adequate points ½ðE0 ; c0 Þ; ðE1 ; c1 Þ describing the VAD linear threshold tuning in Eq. (3). The proposed VAD yields the best speech pause detection accuracy, important reduction of the false alarm rate when compared to G.729, and comparable speech detection accuracy when compared to AMR VADs. In general, the false alarm rates can be decreased by changing threshold criteria in the algorithmÕs decision rules with the corresponding decrease of the hit rates. PAUSE HIT RATE (HR0) 100 80 60 LTSE VAD G.729 AMR1 AMR2 ADAPT. LTSE VAD [(30,8) (50,3.25)] AFE (FD) AFE (WF) 40 20 0 0 5 (a) 10 15 20 FALSE ALARM RATE (FAR0) 3.3. Influence of the VAD on a speech recognition system PAUSE HIT RATE (HR0) 100 80 60 40 LTSE VAD G.729 AMR1 AMR2 ADAPT. LTSE VAD [(30,8) (50,3.25)] AFE (FD) AFE (WF) 20 0 0 10 (b) 20 30 40 FALSE ALARM RATE (FAR0) PAUSE HIT RATE (HR0) 100 80 60 LTSE VAD G.729 AMR1 AMR2 ADAPT. LTSE VAD [(30,8) (50,3.25)] AFE (FD) AFE (WF) 40 20 0 0 (c) 10 20 30 40 281 50 60 FALSE ALARM RATE (FAR0) Fig. 6. ROC curves: (a) stopped car, motor running, (b) town traffic, low speed rough road, (c) high speed, good road. appropriate decision threshold for a given noisy condition in a similar way as it is carried out in the AMR (option 1) standard. Thus, the adaptive LTSE VAD provides a sustained improvement in both speech pause hit rate and false alarm rate over G.729 and AMR VAD being the gains especially important over the Although the discrimination analysis or the ROC curves are effective to evaluate a given algorithm, the influence of the VAD in a speech recognition system was also studied. Many authors claim that VADs are well compared by evaluating speech recognition performance (Woo et al., 2000) since non-efficient SND is an important source of the degradation of recognition performance in noisy environments (Karray and Martin, 2003). There are two clear motivations for that: (i) noise parameters such as its spectrum are updated during non-speech periods being the speech enhancement system strongly influenced by the quality of the noise estimation, and (ii) framedropping, a frequently used technique in speech recognition to reduce the number of insertion errors caused by the noise, is based on the VAD decision and speech misclassification errors lead to loss of speech, thus causing irrecoverable deletion errors. The reference framework (Base) is the ETSI AURORA project for distributed speech recognition (ETSI, 2000) while the recognizer is based on the hidden Markov model toolkit software package (Young et al., 2001). The task consists on recognizing connected digits which are modelled as whole word hidden Markov models with the following parameters: 16 states per word, simple left-to-right models, mixture of 3 Gaussians per state and only the variances of all acoustic coefficients (no full covariance matrix) while speech pause models consist of three states with a mixture of 6 Gaussians per state. The 282 J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 39-parameter feature vector consists of 12 cepstral coefficients (without the zero-order coefficient), the logarithmic frame energy plus the corresponding delta and acceleration coefficients. Two training modes are defined for the experiments conducted on the AURORA-2 database: (i) training on clean data only (clean training), and (ii) training on clean and noisy data (multi-condition training). For the AURORA-3 SpeechDat-Car databases, the so-called well-matched (WM), medium-mismatch (MM) and high-mismatch (HM) conditions are used. These databases contain recordings from the close-talking and distant microphones. In WM condition, both close-talking and hands-free microphones are used for training and testing. In MM condition, both training and testing are performed using the hands-free microphone recordings. In HM condition, training is done using close-talking microphone material from all driving conditions while testing is done using hands-free microphone material taken for low noise and high noise driving conditions. Finally, recognition performance is assessed in terms of the word accuracy (WAcc) that considers deletion, substitution and insertion errors. The influence of the VAD decision on the performance of different feature extraction schemes was studied. The first approach (shown in Fig. 7b) incorporates Wiener filtering (WF) to the Base system as noise suppression method. The second feature extraction algorithm that was evaluated uses Wiener filtering and non-speech frame dropping as shown in Fig. 7c. In this preliminary set of experiments, a simple noise reduction algorithm based on Wiener filtering was used in order to clearly show the influence of the VAD on the system performance. The algorithm has been implemented as described for the first stage of the Wiener filtering in the AFE (ETSI, 2002). No other mismatch reduction techniques already present in the AFE standard (waveform processing or blind equalization) have been considered since they are not affected by the VAD decision and can mask the impact of the VAD precision on the overall system performance. In the next section, results for a full version of the AFE will be presented. Noisy speech Feature Extraction (Base) Recognizer (HTK ) (a) Noisy speech Wiener Filtering Recognizer (HTK) VAD (noise estimation) (b) Noisy speech Wiener Filtering Frame dropping VAD VAD (noise estimation) (frame dropping) Recognizer (HTK) (c) Fig. 7. Speech recognition systems used to evaluate the proposed VAD. (a) Reference recognition system. (b) Enhanced speech recognition system incorporating Wiener filtering as noise suppression method. The VAD is used for estimating the noise spectrum during non-speech periods. (c) Enhanced speech recognition system incorporating Wiener filtering as noise suppression method and frame dropping. The VADs are used for noise spectrum estimation in the Wiener filtering stage and for frame dropping. Table 2 shows the AURORA-2 recognition results as a function of the SNR for the system shown in Fig. 7a (Base), as well as for the speech enhancement systems shown in Fig. 7b (Base + WF) and Fig. 7c (Base + WF + FD), when G.729, AMR, AFE, and LTSE are used as VAD algorithms. These results were averaged over the three test sets of the AURORA-2 recognition experiments. An estimation of the 95% confidence interval (CI) is also provided. Notice that, particularly, for the recognition experiments based on the AFE VADs, we have used the same configuration used in the standard (ETSI, 2002) with different VADs for WF and FD. The same feature extraction scheme was used for training and testing. Only exact speech periods are kept in the FD stage and consequently, all the frames classified by the VAD as non-speech are discarded. FD has impact on the training of silence models since less 283 J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 Table 2 Average word accuracy for the AURORA-2 database System Base Base + WF VAD used None G.729 AMR1 AMR2 AFE LTSE G.729 AMR1 AMR2 AFE LTSE (a) Clean training Clean 99.03 20 dB 94.19 15 dB 85.41 10 dB 66.19 5 dB 39.28 0 dB 17.38 )5 dB 8.65 Average 60.49 CI (95%) ±0.24 98.81 87.70 75.23 59.01 40.30 23.43 13.05 57.13 ±0.24 98.80 97.09 92.05 74.24 44.29 23.82 12.09 66.30 ±0.23 98.81 97.23 94.61 87.50 71.01 41.28 13.65 78.33 ±0.20 98.77 97.68 95.19 87.29 66.05 30.31 4.97 75.30 ±0.21 98.84 97.48 95.17 88.71 72.38 42.51 14.78 79.25 ±0.20 98.41 83.46 71.76 59.05 43.52 27.63 14.94 57.08 ±0.24 97.87 96.83 92.03 71.65 40.66 23.88 14.05 65.01 ±0.23 98.63 96.72 93.76 86.36 70.97 44.58 18.87 78.48 ±0.20 98.78 97.82 95.28 88.67 71.55 41.78 16.23 79.02 ±0.20 99.12 98.14 96.39 91.45 77.06 48.37 20.40 82.28 ±0.19 (b) Multi-condition training Clean 98.48 98.16 20 dB 97.39 93.96 15 dB 96.34 89.51 10 dB 93.88 81.69 5 dB 85.70 68.44 0 dB 59.02 42.58 )5 dB 24.47 18.54 Average 86.47 75.24 CI (95%) ±0.17 ±0.21 98.30 97.04 95.18 91.90 80.77 53.29 23.47 83.64 ±0.18 98.51 97.86 96.97 94.43 87.27 65.45 30.31 88.40 ±0.16 97.86 97.60 96.56 93.98 86.41 64.63 28.78 87.84 ±0.16 98.43 97.94 97.10 94.64 87.52 66.32 31.33 88.70 ±0.15 97.50 96.05 94.82 91.23 81.14 54.50 23.73 83.55 ±0.18 96.67 96.90 95.52 91.76 80.24 53.36 23.29 83.56 ±0.18 98.12 97.57 96.58 93.80 85.72 62.81 27.92 87.29 ±0.16 98.39 97.98 96.94 93.63 85.32 63.89 30.80 87.55 ±0.16 98.78 98.50 97.67 95.53 88.40 67.09 32.68 89.44 ±0.15 Base + WF + FD non-speech frames are available for training. However, if FD is effective enough, few nonspeech periods will be handled by the recognizer in testing and consequently, little influence will have the silence models on the speech recognition performance. The proposed VAD outperforms the standard G.729, AMR1, AMR2 and AFE VADs when used for WF and also, when the VAD is used for removing non-speech frames. Note that the VAD decision is used in the WF stage for estimating the noise spectrum during non-speech periods, and a good estimation of the SNR is critical for an efficient application of the noise reduction algorithm. In this way, the energy-based WF AFE VAD suffers fast performance degradation in speech detection as shown in Fig. 5b, thus leading to numerous recognition errors and the corresponding increase of the word error rate, as shown in Table 2a. On the other hand, FD is strongly influenced by the performance of the VAD and an efficient VAD for robust speech recognition needs a compromise between speech and non-speech detection accuracy. When the VAD suffers a rapid performance degradation under severe noise conditions it losses too many speech frames and leads to numerous deletion errors; if the VAD does not correctly identify nonspeech periods it causes numerous insertion errors the corresponding FD performance degradation. The best recognition performance is obtained when the proposed LTSE VAD is used for WF and FD. Thus, in clean training (Table 2a) the reductions of the word error rate were 58.71%, 49.36%, 17.66% and 15.54% over G.729, AMR1, AMR2 and AFE VADs, respectively, while in multi-condition training (Table 2b) the reductions were of up to 35.81%, 35.77%, 16.92% and 15.18%. Note that FD yields better results for the speech recognition system trained on clean speech. Thus, the reference recognizer yields a 60.49% average WAcc while the enhanced speech recognizer based on the proposed LTSE VAD obtains an 82.28% average value. This is motivated by the fact that models trained using clean speech does not adequately model noise processes, and normally cause insertion errors during non-speech periods. Thus, removing efficiently speech pauses will lead to a significant reduction of this error source. On the 284 J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 standard VADs obtaining the best results followed by AFE, AMR2, AMR1 and G.729. Similar improvements were obtained for the experiments conducted on the Spanish (Moreno et al., 2000), German (Texas Instruments, 2001) and Finnish (Nokia, 2000) SDC databases for the three defined training/test modes. When the VAD is used for WF and FD, the LTSE VAD provided the best recognition results with 52.73%, 44.27%, 7.74% and 25.84% average improvements over G.729, AMR1, AMR2 and AFE, respectively, for the different training/test modes and databases (Table 4). other hand, noise is well modelled when models are trained using noisy speech and the speech recognition system tends itself to reduce the number of insertion errors in multi-condition training as shown in Table 2a. Concretely, the reference recognizer yields an 86.47% average WAcc while the enhanced speech recognizer based on the proposed LTSE VAD obtains an 89.44% average value. On the other hand, since in the worst case CI is 0.24% we can conclude that our VAD method provides better results than the VADs examined and that the improvements are especially important in low SNR conditions. Finally, Table 3 compares the word accuracy results averaged for clean and multi-condition style training modes to the performance of the recognition system using the hand-labelling database. These results show that the performance of the proposed algorithm is very close to that of the manually tagged database. In all test sets, the proposed VAD algorithm is observed to outperform 3.4. Recognition results for the LTSE VAD replacing the AFE VADs In order to compare the proposed method to the best available results, the VADs of the full AFE standard (ETSI, 2002; Macho et al., 2002) (including both the noise estimation VAD and Table 3 AURORA 2 recognition result summary VAD used G.729 AMR1 AMR2 AFE LTSE Hand-labelling Base + WF Base + WF + FD 66.19 70.32 74.97 74.29 83.37 82.89 81.57 83.29 83.98 85.86 84.69 86.86 Table 4 Average word accuracy for the SpeechDat-Car databases System Base Base + WF VAD used None G.729 AMR1 AMR2 AFE LTSE G.729 AMR1 AMR2 AFE LTSE Base + WF + FD Finnish WM MM HM Average 92.74 80.51 40.53 71.26 93.27 75.99 50.81 73.36 93.66 78.93 40.95 71.18 95.52 75.51 55.41 75.48 94.28 78.52 55.05 75.95 95.34 75.10 56.68 75.71 88.62 67.99 65.80 74.14 94.57 81.60 77.14 84.44 95.52 79.55 80.21 85.09 94.25 82.42 56.89 77.85 94.99 80.51 80.60 85.37 Spanish WM MM HM Average 92.94 83.31 51.55 75.93 89.83 79.62 66.59 78.68 85.48 79.31 56.39 73.73 91.24 81.44 70.14 80.94 89.71 76.12 68.84 78.22 91.34 84.35 65.59 80.43 88.62 72.84 65.50 75.65 94.65 80.59 62.41 74.33 95.67 90.91 85.77 90.78 95.28 90.23 77.53 87.68 96.55 91.28 87.07 91.63 German WM MM HM Average 91.20 81.04 73.17 81.80 90.60 82.94 78.40 83.98 90.20 77.67 70.40 79.42 93.13 86.02 83.07 87.41 91.48 84.11 82.01 85.87 92.85 85.65 83.58 87.36 87.20 68.52 72.48 76.07 90.36 78.48 66.23 78.36 92.79 83.87 81.77 86.14 93.03 85.43 83.16 87.21 93.77 86.68 83.40 87.95 76.33 ±0.44 78.67 ±0.42 74.78 ±0.45 81.28 ±0.40 80.01 ±0.41 81.17 ±0.40 75.29 ±0.44 79.04 ±0.42 87.34 ±0.34 84.25 ±0.37 88.32 ±0.33 Average CI (95%) 285 J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 frame dropping VAD) were replaced by the proposed LTSE VAD and the AURORA recognition experiments were conducted. Tables 5 and 6 show the word error rates obtained for the full AFE standard and the modified AFE incorporating the proposed VAD for noise estimation in Wiener filtering and frame dropping. We can observe that a significant improvement is achieved in both databases. In AURORA 2, the word error rate was reduced from 8.14% to 7.82% for the multicondition training experiments and from 13.07% to 11.87% for the clean training experiments. In AURORA 3, the improvements were especially important in high mismatch experiments being the word error rate reduced from 13.27% to 10.54%. Note that these improvements are only achieved by replacing the VAD of the full AFE and not introducing new improvements in the feature extraction algorithm. On the other hand, if the AURORA complex Back-End using digit models with 20 Gaussians per state and a silence model with 36 Gaussians per state is considered, the AURORA 2 word error rate is reduced from 12.04% to 11.29% for the clean training experiments and from 6.57% to 6.13% for the multi-condition training experiments when the VADs of the original AFE are replaced by the proposed LTSE VAD. The efficiency of a VAD for speech recognition depends on a number of factors. If speech pauses Table 5 Average word error rates for the AURORA 2 databases: (a) full AFE standard, (b) Full AFE with LTSE as VAD for noise estimation and frame-dropping Table 6 Average word error raes for the AURORA 3 databases: (a) full AFE standard, (b) full AFE with LTSE as VAD for noise estimation and frame-dropping AURORA 3 word error rate (%) Finnish Spanish German Danish Average 3.96 19.49 14.77 12.10 ±0.58 3.39 6.21 9.23 5.84 ±0.36 4.87 10.40 8.70 7.76 ±0.56 6.02 22.49 20.39 15.38 ±0.75 4.56 14.65 13.27 10.27 ±0.28 (b) AFE + LTSE Well 3.72 Mid 18.06 High 9.36 Overall 10.15 CI (95%) ±0.54 3.07 7.11 7.07 5.48 ±0.35 4.53 10.61 8.33 7.61 ±0.56 6.10 22.27 17.39 14.58 ±0.73 4.35 14.51 10.54 9.46 ±0.27 (a) AFE Well Mid High Overall CI (95%) are very long and are more frequent than speech periods, insertion errors would be an important error source. On the contrary, if pauses are short, maintaining a high speech hit-rate can be beneficial to reduce the number of deletion errors without being significantly important insertion errors. The mismatch between training and test conditions also affects the importance of the VAD in a speech recognition system. When the system suffers a high mismatch between training and test, more important and effective can be a VAD for increasing the performance of speech recognizers. This fact is mainly motivated by the efficiency of the frame-dropping stage in such conditions and the efficient application of the noise suppression algorithms. AURORA 2 word error rate (%) Set A Set B Set C Overall (a) AFE Multi Clean Average CI (95%) 7.71 12.49 10.10 ±0.16 7.90 12.94 10.42 ±0.17 9.49 14.58 11.99 ±0.25 8.14 13.07 10.61 ±0.11 (b) AFE + LTSE Multi Clean Average CI (95%) 7.36 11.44 9.40 ±0.16 7.56 11.56 9.56 ±0.16 9.27 13.35 11.31 ±0.24 7.82 11.87 9.85 ±0.10 4. Conclusion This paper presented a new VAD algorithm for improving speech detection robustness in noisy environments and the performance of speech recognition systems. The VAD is based on the estimation of the long-term spectral envelope and the measure of the spectral divergence between speech and noise. The decision threshold is adapted to the measured noise energy while a controlled hangover is activated only when the observed SNR is 286 J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 low. It was shown by analysing the distribution probabilities of the long-term spectral divergence that, using long-term information about the speech signal is beneficial for improving speech detection accuracy and minimizing misclassification errors. A discrimination analysis using the AURORA2 speech database was conducted to assess the performance of the proposed algorithm and to compare it to standard VADs such as ITU G.729B, AMR, and the recently approved advanced front-end standard for distributed speech recognition. The LTSE-based VAD obtained the best behaviour in detecting non-speech periods and was the most precise in detecting speech periods exhibiting slow performance degradation at unfavourable noise conditions. An analysis of the VAD ROC curves was also conducted using the Spanish SDC database. The ability of the adaptive LTSE VAD to tune the detection threshold enabled working on the optimal point of the ROC curve for different noisy conditions. The adaptive LTSE VAD provided a sustained improvement in both non-speech hit rate and false alarm rate over G.729 and AMR VAD being the gains particularly important over the G.729 VAD. The working point of the G.729 algorithm shifts to the right in the ROC space with decreasing SNR, while the proposed algorithm is less affected by the increasing level of background noise. On the other hand, it has been shown that the AFE VADs yield poor speech/non-speech discrimination. Particularly, WF AFE VAD yields good non-speech detection accuracy but works on a high false alarm rate point on the ROC space, thus suffering rapid performance degradation when the driving conditions get noisier. On the other hand, FD AFE VAD is only used in the DSR standard for frame dropping and has been planned to be conservative exhibiting poor nonspeech detection accuracy and working on a low false alarm rate point of the ROC space. It was also studied the influence of the VAD in a speech recognition system. The proposed VAD outperformed the standard G.729, AMR1, AMR2 and AFE VADs when the VAD decision is used for estimating the noise spectrum in Wiener fil- tering and when the VAD is employed for nonspeech frame dropping. Particularly, when the feature extraction algorithm was based on Wiener filtering and frame-drooping, and the models were trained using clean speech, the proposed LTSE VAD leaded to word error rate reductions of up to 58.71%, 49.36%, 17.66% and 15.54% over G.729, AMR1, AMR2 and AFE VADs, respectively, while the advantages were of up to 35.81%, 35.77%, 16.92% and 15.18% when the models were trained using noisy speech. Similar improvements were obtained for the experiments conducted on the SpeechDat-Car databases. It was also shown that the performance of the proposed algorithm is very close to that of the manually tagged database. The proposed VAD was also compared to the AFE VADs using the full AFE standard as feature extraction algorithm. When the proposed VAD replaced the VADs of the AFE standard, a significant reduction of the word error rate was obtained in both clean and multi-condition training experiments and also for the SpeechDat-Car databases. Acknowledgement This work has been supported by the Spanish Government under TIC2001-3323 research project. References Benyassine, A., Shlomot, E., Su, H., Massaloux, D., Lamblin, C., Petit, J., 1997. ITU-T recommendation G.729 Annex B: a silence compression scheme for use with G.729 optimized for V.70 digital simultaneous voice and data applications. IEEE Comm. Magazine 35 (9), 64–73. Berouti, M., Schwartz, R., Makhoul, J., 1979. Enhancement of speech corrupted by acoustic noise. In: Internat. Conf. on Acoust. Speech Signal Process., pp. 208–211. Beritelli, F., Casale, S., Cavallaro, A., 1998. A robust voice activity detector for wireless communications using soft computing. IEEE J. Select. Areas Comm. 16 (9), 1818– 1829. Beritelli, F., Casale, S., Rugeri, G., Serrano, S., 2002. Performance evaluation and comparison of G.729/AMR/fuzzy voice activity detectors. IEEE Signal Process. Lett. 9 (3), 85–88. J. Ram
ırez et al. / Speech Communication 42 (2004) 271–287 Boll, S.F., 1979. Suppression of acoustic noise in speech using spectral subtraction. IEEE Trans. Acoust. Speech Signal Process. 27, 113–120. Bouquin-Jeannes, R.L., Faucon, G., 1994. Proposal of a voice activity detector for noise reduction. Electron. Lett. 30 (12), 930–932. Bouquin-Jeannes, R.L., Faucon, G., 1995. Study of voice activity detector and its influence on a noise reduction system. Speech Comm. 16, 245–254. Cho, Y.D., Kondoz, A., 2001. Analysis and improvement of a statistical model-based voice activity detector. IEEE Signal Process. Lett. 8 (10), 276–278. Cho, Y.D., Al-Naimi, K., Kondoz, A., 2001a. Improved voice activity detection based on a smoothed statistical likelihood ratio. In: Internat. Conf. on Acoust. Speech Signal Process., Vol. 2, pp. 737–740. Cho, Y.D., Al-Naimi, K., Kondoz, A., 2001b. Mixed decisionbased noise adaptation for speech enhancement. Electron. Lett. 37 (8), 540–542. ETSI EN 301 708 recommendation, 1999. Voice activity detector (VAD) for adaptive multi-rate (AMR) speech traffic channels. ETSI ES 201 108 recommendation, 2000. Speech processing, transmission and quality aspects (STQ); distributed speech recognition; front-end feature extraction algorithm; compression algorithms. ETSI ES 202 050 recommendation, 2002. Speech processing, transmission and quality aspects (STQ); distributed speech recognition; advanced front-end feature extraction algorithm; compression algorithms. Freeman, D.K., Cosier, G., Southcott, C.B., Boyd, I., 1989. The voice activity detector for the PAN-European digital cellular mobile telephone service. In: Internat. Conf. on Acoust. Speech Signal Process., Vol. 1, pp. 369–372. Hirsch, H.G., Pearce, D., 2000. The AURORA experimental framework for the performance evaluation of speech recognition systems under noise conditions. In: ISCA ITRW ASR2000: Automatic Speech Recognition: Challenges for the Next Millennium. Itoh, K., Mizushima, M., 1997. Environmental noise reduction based on speech/non-speech identification for hearing aids. In: Internat. Conf. on Acoust. Speech Signal Process., Vol. 1, pp. 419–422. ITU-T recommendation G.729-Annex B, 1996. A silence compression scheme for G.729 optimized for terminals conforming to recommendation V.70. Karray, L., Martin, A., 2003. Towards improving speech detection robustness for speech recognition in adverse environment. Speech Comm. 40 (3), 261–276. 287 Macho, D., Mauuary, L., Noe, B., Cheng, Y.M., Ealey, D., Jouvet, D., Kelleher, H., Pearce, D., Saadoun, F., 2002. Evaluation of a noise-robust DSR front-end on AURORA databases. In: Proc. 7th Internat. Conf. on Spoken Language Process. (ICSLP 2002), pp. 17–20. Madisetti, V., Williams, D.B., 1999. Digital Signal Processing Handbook. CRC/IEEE Press. Martin, R., 1993. An efficient algorithm to estimate the instantaneous SNR of speech signals. In: Eurospeech, Vol. 1, pp. 1093–1096. Marzinzik, M., Kollmeier, B., 2002. Speech pause detection for noise spectrum estimation by tracking power envelope dynamics. IEEE Trans. Speech Audio Process. 10 (2), 109–118. Moreno, A., Borge, L., Christoph, D., Gael, R., Khalid, C., Stephan, E., Jeffrey, A., 2000. SpeechDat-Car: a large speech database for automotive environments. In: Proc. II LREC. Nemer, E., Goubran, R., Mahmoud, S., 2001. Robust voice activity detection using higher-order statistics in the LPC residual domain. IEEE Trans. Speech Audio Process. 9 (3), 217–231. Nokia, 2000. Baseline results for subset of SpeechDat-Car Finnish Database for ETSI STQ WI008 advanced front-end evaluation, http://icslp2002.colorado.edu/special_sessions/ aurora/references/aurora_ref3.pdf. Sangwan, A., Chiranth, M.C., Jamadagni, H.S., Sah, R., Prasad, R.V., Gaurav, V., 2002. VAD techniques for realtime speech transmission on the Internet. In: IEEE Internat. Conf. on High-Speed Networks and Multimedia Comm., pp. 46–50. Sohn, J., Sung, W., 1998. A voice activity detector employing soft decision based noise spectrum adaptation. In: Internat. Conf. on Acoust. Speech Signal Process., Vol. 1, pp. 365– 368. Sohn, J., Kim, N.S., Sung, W., 1999. A statistical model-based voice activity detection. IEEE Signal Process. Lett. 6 (1), 1–3. Texas Instruments, 2001. Description and baseline results for the subset of the SpeechDat-Car German Database used for ETSI STQ AURORA WI008 advanced DSR front-end evaluation, http://icslp2002.colorado.edu/special_sessions/ aurora/references/aurora_ref6.pdf. Woo, K., Yang, T., Park, K., Lee, C., 2000. Robust voice activity detection algorithm for estimating noise spectrum. Electron. Lett. 36 (2), 180–181. Young, S., Evermann, G., Kershaw, D., Moore, G., Odell, J., Ollason, D., Valtchev, V., Woodland, P., 2001. The HTK Book (for HTK Version 3.1).