An accelerated failure time (AFT) model has been widely used for the analysis of censored survival or failure time data. However, the AFT imposes the restrictive log-linear relation between the survival time and the explanatory variables. In this paper, we introduce a relevance vector machine survival (RVMS) model based on Weibull AFT model that enables the use of kernel framework to automatically learn the possible nonlinear effects of the input explanatory variables on target survival times. We take advantage of the Bayesian inference technique in order to estimate the model parameters. We also introduce two approaches to accelerate the RVMS training. In the first approach, an efficient smooth prior is employed that improves the degree of sparsity. In the second approach, a fast marginal likelihood maximization procedure is used for obtaining a sparse solution of survival analysis task by sequential addition and deletion of candidate basis functions. These two approaches, denoted by smooth RVMS and fast RVMS, typically use fewer basis functions than RVMS and improve the RVMS training time; however, they cause a slight degradation in the RVMS performance. We compare the RVMS and the two accelerated approaches with the previous sparse kernel survival analysis method on a synthetic data set as well as six real-world data sets. The proposed kernel survival analysis models have been discovered to be more accurate in prediction, although they benefit from extra sparsity. The main advantages of our proposed models are: 1) extra sparsity that leads to a better generalization and avoids overfitting; 2) automatic relevance sample determination based on data that provide more accuracy, in particular for highly censored survival data; and 3) flexibility to utilize arbitrary number and types of kernel functions (e.g., non-Mercer kernels and multikernel learning).

Sparse kernel methods have been widely used in regression and classification applications. The performance and the sparsity of these methods are dependent on the appropriate choice of the corresponding kernel functions and their parameters. Typically, the kernel parameters are selected using a cross-validation approach. In this paper, a learning method that is an extension of the relevance vector machine (RVM) is presented. The proposed method can find the optimal values of the kernel parameters during the training procedure. This algorithm uses an expectation-maximization approach for updating kernel parameters as well as other model parameters; therefore, the speed of convergence and computational complexity of the proposed method are the same as the standard RVM. To control the convergence of this fully parameterized model, the optimization with respect to the kernel parameters is performed using a constraint on these parameters. The proposed method is compared with the typical RVM and other competing methods to analyze the performance. The experimental results on the commonly used synthetic data, as well as benchmark data sets, demonstrate the effectiveness of the proposed method in reducing the performance dependency on the initial choice of the kernel parameters.

Speech segmentation to covariance‐stationary regions is of great interest in various fields of speech processing. This study presents a novel sequential segmentation algorithm relying on Random Matrix Theory (RMT). The proposed approach utilises an RMT‐derived test statistic checking stationarity of sample covariance matrices. Utilising the statistical properties of the sample eigenvalues recently investigated in the RMT literature, a new expression for the decision threshold is derived in the non‐asymptotic case where the number of samples is comparable with the dimension. This derivation is of great help to detect very short non‐stationary intervals where the performance of most segmentation methods based on Autoregressive models degrades. Furthermore, based on the results of previous step, a new segmentation procedure is proposed. This procedure is applied to both synthetic and real‐world speech data. Comparing the obtained simulation results with the state‐of‐the‐art segmentation algorithms, the proposed approach has demonstrated good performance with extremely lower computational cost.

This paper introduces a novel sparse Bayesian machine-learning algorithm for embedded feature selection in classification tasks. Our proposed algorithm, called the relevance sample feature machine (RSFM), is able to simultaneously choose the relevance samples and also the relevance features for regression or classification problems. We propose a separable model in feature and sample domains. Adopting a Bayesian approach and using Gaussian priors, the learned model by RSFM is sparse in both sample and feature domains. The proposed algorithm is an extension of the standard RVM algorithm, which only opts for sparsity in the sample domain. Experimental comparisons on synthetic as well as benchmark data sets show that RSFM is successful in both feature selection (eliminating the irrelevant features) and accurate classification. The main advantages of our proposed algorithm are: less system complexity, better generalization and avoiding overfitting, and less computational cost during the testing stage.

Single Channel Speech Separation has been the objective of extensive research in recent years. In this paper, we propose a hybrid system of coherent and incoherent modulation filtering for separation of the target speech from the interference. In the proposed system, subband envelopes are determined using a coherently detected subband carrier based on the time-dependent spectral Center-Of-Gravity (COG) demodulation and then the interference signal is eliminated by applying the adaptive Affine Projection (AP) filter to the subband envelop. The reference signal for the adaptive AP filter is provided by a parallel incoherent system of modulation filtering. Our evaluations, based on several objective measures, indicate that the proposed system extracts the majority of target speech signal segments with minimal interference, outperforming previous systems in voiced speech separation.

Multichannel blind deconvolution (MCBD) algorithms are known to suffer from an extensive computational complexity problem, which makes them impractical for blind source separation (BSS) of speech and audio signals. This problem is even more serious with noncausal MCBD algorithms that must be used in many frequently occurring BSS setups. In this paper, we propose a novel frequency domain algorithm for the efficient implementation of noncausal multichannel blind deconvolution. A block-wise formulation is first developed for filtering and adaptation of filter coefficients. Based on this formulation, we present a modified overlap-save procedure for noncausal filtering in the frequency domain. We also derive update equations for training both causal and anti-causal filters in the frequency domain. Our evaluations indicate that the proposed frequency domain implementation reduces the computational requirements of the algorithm by a factor of more than 100 for typical filter lengths used in blind speech separation. The algorithm is employed successfully for the separation of speech mixtures in a reverberant room. Simulation results demonstrate the superior performance of the proposed algorithm over causal MCBD algorithms in many potential source and microphone positions. It is shown that in BSS problems, causal MCBD algorithms with center-spike initialization do not always converge to a delayed form of the desired noncausal solution, further revealing the need for an efficient noncausal MCBD algorithm.

This paper presents a method based on higher order statistics (HOS), namely the normalized third-order moment (skewness), for blind estimation of the inverse filter of the room impulse response (RIR). Skewness is used as a measure of asymmetry, and a comprehensive comparison with the commonly used metric (kurtosis) is presented. It is shown that a sufficiently long linear predictive (LP) residual of the speech signal has an asymmetric pdf with sufficient skewness to be used as a score function for the HOS-based approach. The proposed algorithm is optimized for the inverse filter estimation problem. This optimization includes an efficient initialization for high reverberation intensities, enabling the method to be employed in highly reverberant rooms. The direct-to-reverberation ratio (DRR) as well as the equalized impulse response clearly show that our method can estimate the inverse filter even in highly reverberant environments. In addition, performance results using recorded background noise and in time-varying environments illustrate that our approach is applicable in real world situations. The proposed method is shown to be superior to the method by Wu and Wang, particularly in terms of reducing the coloration effect. Experiments under different acoustic conditions confirm the effectiveness of the proposed method for time delay estimation (TDE). Finally, the proposed algorithm is used as the first-stage of monaural segregation, and it is shown to improve the performance under different conditions.

In this paper, we suggest applying tree structure on the sinusoidal parameters. The suggested sinusoidal coder is targeted to find the coded sinusoidal parameters obtained by minimizing a likelihood function in a least square (LS) sense. From a rate-distortion standpoint, we address the problem of how to allocate available bits among different frequency bands to code sinusoids at each frame. For further analyzing the quantization behavior of the proposed method, we assess the quantization performance with respect to other methods: the short-time Fourier transform (STFT) based coder commonly used for speech enhancement or separation, and the line spectral frequency (LSF) coder used in speech coding. Through extensive simulations, we show that the proposed quantizer leads to less spectral distortion as well as higher perceived quality for the re-synthesized signals based on the coded parameters in a model-based approach with respect to previous STFT-based methods. The proposed method lowers the complexity, and, due to its tree-structure, leads to a rapid search capability. It provides flexibility for use in many speaker-independent applications by finding the most likely frequency vectors selected from a list of frequency candidates. Therefore, the proposed quantizer can be considered an attractive candidate for model-based speech applications in both speaker-dependent and speaker-independent scenarios.

Single-channel separation (SCS) is a challenging scenario where the objective is to segregate speaker signals from their mixture with high accuracy. In this research a novel framework called subband perceptually weighted transformation (SPWT) is developed to offer a perceptually relevant feature to replace the commonly used magnitude of the short-time Fourier transform (STFT). The main objectives of the proposed SPWT are to lower the spectral distortion (SD) and to improve the ideal separation quality. The performance of the SPWT is compared to those obtained using mixmax and Wiener filter methods. A comprehensive statistical analysis is conducted to compare the SPWT quantization performance as well as the ideal separation quality with other features of log-spectrum and magnitude spectrum. Our evaluations show that the SPWT provides lower SD values and a more compact distribution of SD, leading to more acceptable subjective separation quality as evaluated using the mean opinion score.

Fast adaptive algorithms targeted for low-resource implementation on oversampled filterbanks are discussed. First, simplifications to the pseudo-affine projection (PAP) are proposed that decrease the computation cost without degrading the performance. Next, the sequential update PAP algorithm is proposed to further decrease the computational complexity. It is shown that, unlike the partial update fast affine projection, increasing the decimation rate of the partial update algorithm does not lead to more aliasing in the sequential update PAP. Moreover, it is experimentally observed that with proper regularization or step-size scaling, the convergence rate of the sequential PAP is identical for various decimation factors of the partial update. The proposed methods are implemented and evaluated on a low-resource oversampled filterbank platform.

Partial update algorithms such as sequential LMS have been proposed to reduce computation cost. To further decrease the computation cost for low-resource real-time echo cancellation by oversampled subband adaptive filters (OS-SAFs), we propose a new hybrid partial update algorithm. This algorithm, based on a polyphase perspective, selectively prunes (zeros out) polyphase components of the adaptive filter. When employed on OS-SAFs, the proposed algorithm achieves significant complexity reduction. This algorithm is applied to sequential versions of LMS and pseudo-affine projection algorithms employed to adapt OS-SAFs. Presented performance evaluations prove that considerable computation cost reduction (typically more than four times) is feasible with practically insignificant degradations in the steady-state performance of the OS-SAF system. Moreover, the algorithm does not affect initial convergence properties.

The affine projection algorithm (APA) has been shown to improve the performance of oversampled subband adaptive filters (OS-SAFs) compared to classical normalized least mean square (NLMS) algorithms. Because of the complexity of APA, however, only low-order APAs are practical for real-time implementation. Thus, we propose a reduced-complexity version of the Gauss-Seidel fast APA (GSFAPA) for adapting the subband filters in OS-SAF systems. We propose modifying the GSFAPA with a complexity reduction method based on partial filter update, and also with a low-cost method for combined regularization and step size control. We show the advantage of the new algorithm - termed low-cost Gauss-Seidel fast affine projection - compared to the APA in a subband echo canceller application.

This paper describes a low-power implementation of the Bluetooth subband codec (SBC) for high-fidelity wireless audio. The design uses a configurable weighted overlap-add (WOLA) filterbank coprocessor to implement the analysis and synthesis filterbanks. A new method to convert the two-times over-sampled, complex WOLA subband signals to equivalent critically sampled, real-valued SBC subband signals is presented. The WOLA coprocessor allows for an efficient parallel implementation of the filterbank and quantization portions of the SBC algorithm. Details of the overall system design are also presented, including measurements of power consumption and resource requirements. The final real-time, fixed-point implementation is compared to an off-line floating-point reference and found to produce no audible difference in decoded signal quality.

The log energy parameter has long been used as an extension to the basic cepstral feature vector in speech recognition. The use of a normalization technique for the log energy parameter has also been widely accepted. In this paper, a simple energy normalization scheme is introduced that allows direct use of the frame energy parameter in speech recognition and performs well in the presence of noise. Its combination with traditional cepstral mean and variance normalizations has led to error rate improvements of up to 55% on the Aurora 2 task, in comparison to the baseline clean-trained system using feature set including the log energy parameter. This achievement has been obtained with neither complicated programming nor computation expensive routines. The performance of this scheme on an utterance-wide basis has been close to that of the off-line speaker-wide normalization, which makes it a good candidate for practical systems.

The phenomenon of acoustic shock occurs when a headset user is subjected to an audio disturbance at an uncomfortable or unsafe level. In this paper, a new acoustic shock limiting algorithm is proposed which reduces the output level by utilizing both time and frequency domains limiting features. During acoustic shocks, the algorithm improves the intelligibility of the underlying speech and minimizes the artifacts by reducing the shock in the subbands based on the narrowband and broadband characteristics of the input signal. The algorithm is implemented on a low power DSP system where the input data is analyzed in both time and frequency domains. The method is tested for sinusoidal and speech inputs. The results demonstrate shock compression and limiting, with reduced speech distortion during shock onset and good overall speech quality. 1.

Performance of adaptive noise cancellation (ANC) degrades severely when uncorrelated noise components are present at the two inputs. Thus, practical background diffuse noises pose a serious problem for ANC systems. In this letter, we propose a new hybrid system that integrates subband adaptive filters (SAFs) and a Wiener filter. The hybrid system is implemented on an oversampled DFT filterbank that efficiently integrates the SAF and the Wiener filter components in the frequency-domain. Performance evaluation of the hybrid system in presence of diffuse noise interference shows that the proposed system is superior to both the Wiener filter and the SAF subsystems.

The performance of the normalized least mean square (NLMS) algorithm for adaptive filtering is dependent on the spectral flatness of the reference input. Thus, the standard NLMS algorithm does not perform well in over-sampled subband adaptive filters (OS-SAF) because colored subband signals are generated even for white input signals. Thus we propose the use of the affine projection algorithm (APA) to adapt the individual subband filters in OS-SAF systems. The OS-SAF using APA for adaptation is implemented on a fast, low-resource oversampled filterbank. Through both theoretical and experimental analyses, it is demonstrated that a low order APA significantly improves the convergence behavior, offering a low computational complexity compared to the recursive least squares (RLS) method. We employ a recursive method of calculating the correlation matrix to further decrease the computation cost without affecting the performance.

The ETSI AMR-2 VAD is rigorously evaluated in clean and noisy conditions. The VAD is then simplified and optimized for porting to an ultra low-resource DSP system using a fast oversampled DFT filterbank. The parameters of the low-resource VAD are optimized using two speakers and 6 types of noise at SNR from -10 to 20 dB. The VAD is then tested by employing sentences from two other speakers and 12 different types of noise. Results show that the low-resource VAD offers a performance comparable to that of the ETSI VAD in both clean and noisy conditions. When deployed on a custom DSP running at a clock speed of 1.28 MHz and consuming less than 1 milliwatt of power, the low-resource VAD uses less than 30% of the available system resources.

The performance of the normalized least mean square (NLMS) algorithm for adaptive filtering is dependent on the spectral flatness of the reference input. Thus, the standard NLMS algorithm does not perform well in over-sampled subband adaptive filters (OS-SAFs) because colored subband signals are generated even for white input signals. Thus we propose the use of the affine projection algorithm (APA) to adapt the individual subband filters in OS-SAP systems. The OS-SAF using APA for adaptation is implemented on a fast, low-resource over-sampled filterbank. Through both theoretical and experimental analyses, it is demonstrated that a low order APA will significantly improve the convergence behavior, offering a low computational complexity compared to the recursive least squares (RLS) method. We employ a recursive method of calculating the correlation matrix to further decrease the computation cost without affecting the performance.

The ETSI AMR-2 VAD is rigorously evaluated in clean and noisy conditions. The VAD is then simplified and optimized for porting to an ultra low-resource DSP system using a fast oversampled DFT filterbank. The parameters of the low-resource VAD are optimized using two speakers and 6 types of noise at SNRs from -10 to 20 dB. The VAD is then tested by employing sentences from two other speakers and 12 different types of noise. Results show that the low-resource VAD offers a performance comparable to that of the ETSI VAD in both clean and noisy conditions. When deployed on a custom DSP running at a clock speed of 1.28 MHz and consuming less than 1 milliWatt of power, the low-resource VAD uses less than 30% of the available system resources.

The convergence rate of the Least Mean Square (LMS) algorithm is dependent on the eigenvalue distribution of the reference input correlation matrix. When adaptive filters are employed in low-delay over-sampled subband structures, colored subband signals considerably decelerate the convergence speed. Here, we propose and implement two promising techniques for improving the convergence rate based on: 1) Spectral emphasis and 2) Decimation of the subband signals. We analyze the effects of the proposed methods based on theoretical relationships between eigenvalue distribution and convergence characteristics. We also propose a combined decimation and spectral emphasis whitening technique that exploits the advantages of both methods to dramatically improve the convergence rate. Moreover, through decimation the combined whitening approach reduces the overall computation cost compared to subband LMS with no preprocessing. Presented theoretical and simulation results confirm the effectiveness of the proposed convergence improvement methods. 1.

Acoustic Shock describes a condition where sudden loud acoustic signals in communication equipment causes hearing damage and discomfort to the users. To combat this problem, a subband-based acoustic shock limiting (ASL) algorithm is proposed and implemented on an ultra low-power DSP system with an input-output latency of 6.5 msec. This algorithm processes the input signal in both the time and frequency domains. This approach allows the algorithm to detect sudden increases in sound level (time-domain), as well as frequencyselectively suppressing shock disturbances in frequency domain. The unaffected portion of the sound spectrum is thus preserved as much as possible. A simple ASL algorithm calibration procedure is proposed to satisfy different sound pressure level (SPL) limit requirements for various communication equipment. Acoustic test results show that the ASL algorithm limits acoustic shock signals to below specified SPL limits while preserving speech quality. 1.

This paper presents the design and implementation of a subband cardioid beamformer on an ultra low-power miniature DSP platform, using a 2-microphone endfire array. The subband beamformer extends the classical time-domain, narrow-band algorithm to a frequency-domain, broadband implementation, so it is suitable for general speech and audio applications. An oversampled, weighted overlap-add filterbank is used to allow wide gain and phase adjustments for low power, low group delay requirements. A subband IIR filter is proposed to overcome the non-zero bandwidth of the frequency bands, and to introduce a nearly linear phase adjustment across the bands. The subband implementation allows the flexibility of integrating the beamformer with additional algorithms at different frequency ranges. The beamformer has been implemented in real-time on Dspfactory's Toccata platform, which has been specifically designed for ultra low-power, miniature, head-mounted audio devices. At 1.25 Volts with a 5 MIPS DSP core, the Toccata consumes only about 800 micro Watts without microphones and receivers.

A real-time HMM-based isolated word recognition system is implemented on an ultra low-power miniature DSP system. The DSP system consumes less than 1 milliWatt, much less than what is considered today as “low-resource”. It has a very small footprint and requires only a single hearing aid sized 1 volt battery. The efficient implementation of HMM and MFCC feature extraction algorithms is accomplished through the use of three processing units running concurrently. In addition to the DSP core, an input/output processor creates frames of input speech signals, and a WOLA fi1terbank unit performs windowing, FFT and vector multiplications. A system evaluation using a vocabulary of 18 words shows a success rate of more than 99%.

An efficient implementation of a time-domain speech synthesis system on an ultra low-power, miniature, programmable block-floating-point DSP system is introduced. The DSP system, operating at a clock rate as low as 1.28 MHz, is well suited for speech and audio processing applications. Similar to the MBR-PSOLA technique, this time-domain synthesis method uses a normalized speech database generated by a high-quality harmonic synthesis. To reduce the memory usage and communication bandwidth, the normalized database is compressed using a block-adaptive, ADPCM approach. Listening tests comparing the synthetic speech quality on the DSP system and the same method implemented on a high-resource computer system show no degradations due to the memory, register length, or other low-resource limitations on the DSP system. The system consumes less than 1 mW at 1 volt.

The purpose of this work is to demonstrate that distributed speech recognition front-ends can be deployed in environments which provide for very little power and CPU resources, with possibly no degradation of speech recognition quality when compared to standard floatingpoint implementations. The ETSI distributed speech recognition front-end standard is implemented on an ultra low-power miniature DSP system. The efficient implementation of the ETSI algorithm components, i.e. feature extraction, feature compression and multi-framing, is accomplished through the use of three processing units running concurrently. In addition to a DSP core, an input/output processor creates frames of input speech signals, and a weighted overlap-add (WOLA) filterbank unit performs windowing, FFT and vector multiplications. System evaluation using the TI digits database shows that the performance of the ultra low-power DSP system is equivalent to the reference implementation provided by ETSI.

A real-time subband adaptive noise cancellation system on an ultra low-power miniature DSP system is implemented. The system is targeted at personal communication devices where the speaker may be in a noisy environment. The system is implemented on an ultra low-power DSP system that incorporates a DSP core and an oversampled WOLA filterbank. Pre-emphasis filters are used to increase the convergence rate of a leaky LMS algorithm in the oversampled subband implementation. System performance is also improved relative to a fullband implementation due to benefits arising from using subband adaptive filters instead of fullband filters. A 10 dB reduction of noise power is achieved in tests using various noise conditions. The entire DSP system consumes 2.1 mW and can be realized in a package size of 6.5 x 3.5 x 2.5 mm. 1.

The problem of speech enhancement using wavelet thresholding algorithm is considered. Major problems in applying the basic algorithm are discussed and modifications are proposed to improve the method. First, we propose the use of different thresholds for different wavelet bands. Next, by employing a pause detection algorithm, noise profile is estimated and the thresholds are adapted. This enables the modified enhancement system to handle colored and nonstationary noises. Finally, a wavelet-based voiced/unvoiced classification is proposed and implemented that can further improve the performance of the enhancement system. To evaluate the system performance, we have used real-life noise types such as multi-talker babble and low-pass noises. Subjective and objective evaluations show that the proposed system improves the performance the wavelet thresholding algorithm.

In this paper we describe a speech enhancement system based on spectral subtraction and a proposed AR-HMM-based pause detection algorithm. When trained off-line with limited types of environmental noises, the system performs well under almost all types of environmental noises. We show that the proposed pause detection method is at least one order of magnitude faster than the traditional correlation-based algorithms, and at the same time, has a better performance in locating long pauses in noisy speech. A method of musical noise reduction is also applied in this work. The hybrid speech enhancement system is implemented on a Pentium-166 PC. Using a standard sound card or digital audio interface, the system works in real-time with a negligible solid delay. The speech enhancement experiments using the hybrid system and the conventional spectral subtraction method clearly demonstrate the superiority of the proposed pause detection approach.

This paper presents a general overview of the current research activities in the European PICASSO project on speaker verification for telephone applications. First, the general formalism used by the project is described. Then the scientific issues under focus are discussed in detail. Finally, the paper briefly describes the Picassoft research platform. Along the article, entry points to more specific work also published in the Eurospeech99 proceedings are given.

An improved hidden Markov model-based (HMM-based) speech enhancement system designed using the minimum mean square error principle is implemented and compared with a conventional spectral subtraction system. The improvements to the system are: (1) incorporation of mixture components in the HMM for noise in order to handle noise nonstationarity in a more flexible manner, (2) two efficient methods in the speech enhancement system design that make the system real-time implementable, and (3) an adaptation method to the noise type in order to accommodate a wide variety of noise expected under the enhancement system's operating environment. The results of the experiments designed to evaluate the performance of the HMM-based speech enhancement systems in comparison with spectral subtraction are reported. Three types of noise-white noise, simulated helicopter noise, and multitalker (cocktail party) noise-were used to corrupt the test speech signals. Both objective (global SNR) and subjective mean opinion score (MOS) evaluations demonstrate consistent superiority of the HMM-based enhancement systems that incorporate the innovations described in this paper over the conventional spectral subtraction method.

A modeling approach to auditory speech analysis and recognition is proposed and evaluated, where a composite auditory model is used to generate parallel sets of auditory-nerve instantaneous firing rates (IFRs) along the spatial dimension, followed by a processing stage that constructs from the IFRs the interval statistics in a form called the interpeak interval histogram (IPIH). A speech preprocessor is designed that performs transformation on the auditory IPIHs and interfaces the IPIH-based auditory representation with a hidden Markov model-based (HMM-based) speech recognizer. The results demonstrate that the new preprocessor consistently outperforms the conventional mel frequency cepstral coefficient-based (MFCC-based) preprocessor for the signal-to-noise ratio (SNR) level up to at least 16 dB.

We describe our recent work on the real-time implementation of a state-of-the-art HMM-based MMSE speech enhancement algorithm, where our earlier published algorithm has been approximated, optimized, and simplified. The key innovations enabling the enhancement system to run in real-time are: (1) an algorithm for automatic selection of the noise model, (2) pruning the MMSE forward calculations, (3) a new pause detection method useable for a SNR down to 0 dB, and (4) task partitioning of the entire system. A preliminary version of the real-time enhancement system is simulated on the IBM-PC 486DX2/66 and at the same time implemented on a DSP platform based on dual TMS320C30 DSP chips using single precision floating point arithmetic.

We have investigated effectiveness of several HMM-based speech enhancement strategies in the hearing aid context and compared their performance to a traditional method based on spectral subtraction (also called Wiener filtering). The performance of the various systems has been evaluated using several types of noise likely to be encountered by hearing impaired listeners. Both the objective and subjective evaluations demonstrate consistent superiority of the HMM-based enhancement methods over the spectral subtraction method. The superiority is more evident for sentences contaminated by nonstationary noises at low input SNRs.>

The authors describe a novel approach to speech recognition by directly modeling the statistical characteristics of the speech waveforms. This approach allows them to remove the need for using speech preprocessors, which conventionally serve a role of converting speech waveforms into frame-based speech data subject to a subsequent modeling process. Central to their method is the representation of the speech waveforms as the output of a time-varying filter excited by a Gaussian source time-varying in its power. In order to formulate a speech recognition algorithm based on this representation, the time variation in the characteristics of the filter and of the excitation source is described in a compact and parametric form of the Markov chain. They analyze in detail the comparative roles played by the filter modeling and by the source modeling in speech recognition performance. Based on the result of the analysis, they propose and evaluate a normalization procedure intended to remove the sensitivity of speech recognition accuracy to often uncontrollable speech power variations. The effectiveness of the proposed speech-waveform modeling approach is demonstrated in a speaker-dependent, discrete-utterance speech recognition task involving 18 highly confusable stop consonant-vowel syllables. The high accuracy obtained shows promising potentials of the proposed time-domain waveform modeling technique for speech recognition.