Reliably detecting salient phonetic-acoustic cues plays an important role in speech recognition based on speech landmarks. Once these speech landmarks are located, not only phone recognition can be performed but some other useful information can be derived as well. This paper focuses on the topic of detecting burst onset landmark, an important phonetic characteristic in stops and affricates. The proposed burst onset detector is based on random forest, a learning algorithm renowned for its high accuracy and efficiency in classification. By appending intermediate detection results to MFCCs, the expanded feature can bring benefit to the recognition of stop and affricate consonants in continuous speech.

It had been shown that a segment of pitch contour represented by a set of Legendre polynomial coefficients was successful to the pair-wise language identification task. Feature vectors comprising these polynomial coefficients were formerly modeled by a Gaussian mixture model (GMM) for each language. However, the static model like GMM does not take advantage of the temporal information across several pitch contours. It is intuitive that the temporal information of prosodic features should be used for capturing the characteristics of a specific language. In this paper, a novel dynamic model in ergodic topology is proposed. The experiments show that the proposed method significantly improves the identification rate, even for stress-timed and syllable-timed languages.

An approach to automatic language identification (LID) using pitch contour information is proposed. A segment of pitch contour is approximated by a set of Legendre polynomials so that coefficients of the polynomials form a feature vector to represent this pitch contour. A Gaussian mixture model (GMM) method based on feature vectors extracted from pitch contours is suggested for the LID. Our experiments show that only two or three coefficients are necessary to obtain reasonably good identification rates. We also find that the length of the segmented pitch contour is another useful feature for LID, so that it is included to improve the performance further. Pair-wise language identification experiments on the OGI-TS corpus show that our proposed approach is very promising We also find that tonal languages and pitch accent languages achieve better performance in our system.

This paper presents a robust approach for speaker identification when the speech signal is corrupted by additive noise and channel distortion. Robust features are derived by assuming that the corrupting noise is stationary and the channel effect is fixed during an utterance. A two-step temporal filtering procedure on the autocorrelation sequence is proposed to minimize the effect of additive and convolutional noises. The first step applies a temporal filtering procedure in autocorrelation domain to remove the additive noise, and the second step is to perform the mean subtraction on the filtered autocorrelation sequence in logarithmic spectrum domain to remove the channel effect. No prior knowledge of noise characteristic is necessary. The additive noise can be a colored noise. Then the proposed robust feature is combined with the projection measure technique to gain further improvement in recognition accuracy. Experimental results show that the proposed method can significantly improve the performance of speaker identification task in noisy environment.

When a speech signal is encoded in some low bit-rate coding formats, it becomes more difficult to distinguish speaker identities. The paper investigates the codec effect on acoustic and prosodic features. A new representation of prosodic features based on the piecewise fitting of the pitch contour is introduced. A method for including residual features based on the LDA (linear discriminant analysis) algorithm is suggested. By combining prosodic features with acoustic features, we can improve the performance of a speaker recognition system. A series of experiments is performed with coded speech affected by G.729A and GSM codec processes to demonstrate the effectiveness of our proposed method.

Although the noise masking threshold (NMT) has been applied in adapting speech enhancement systems, it may suffer from underestimation in low-SNR condition. We derive a robust weighting factor for each wavelet subband. The function of the robust weighting factor is to keep the energy of residual noise lower than the NMT and the speech distortion smaller than the residual noise. If the energy of residual noise is greater than the NMT, the wavelet coefficients (WCs) of noisy speech are suppressed to remove more residual noise. If the energy of residual noise is smaller than the NMT, the weighting factor is set to one to retain speech quality. It results in a lower bound of NMT for preventing the underestimation of weighting factors. Experimental results show that the proposed method can improve the naturalness of enhanced speech.

This paper proposes a set of hierarchical harmonicities to analyze the harmonic structure of a speech signal. The high redundant harmonic structure in voiced speech makes the perception of speech more robust in noisy environments. It is possible to recover a distorted harmonic structure. The systematic information of harmonic structure is represented by a set of harmonicities, including grid, temporal, spectral and segmental harmonicities. By using this harmonic measure, the speech quality after performing the enhancement can be evaluated. It gives an indicator of whether the further enhancement is necessary. A post speech enhancement based on the reconstruction of harmonic structure is proposed to enhance the quality of voiced speech. The experiment shows the effectiveness of the proposed method.

Harmonic structure can be easily recognized in the timefrequency representation of speech signals even in the diverse environment. The harmonicity is a measure of the completeness of harmonic structure. This paper extends the use of conventional harmonicity measure to the tasks of pitch estimation and voice activity detection. A set of hierarchical harmonicities, including grid, temporal, spectral and segmental harmonicities, is derived for this purpose. A series of experiments are conducted to show the effectiveness of using harmonicities in speech processing.

A model-based framework of classification error rate estimation is proposed for speech and speaker recognition. It aims at predicting the run-time performance of a hidden Markov model (HMM) based recognition system for a given task vocabulary and grammar without the need of running recognition experiments using a separate set of testing samples. This is highly desirable both in theory and in practice. However, the error rate expression in HMM-based speech recognition systems has no closed form solution due to the complexity of the multi-class comparison process and the need for dynamic time warping to handle various speech patterns. To alleviate the difficulty, we propose a one-dimensional model-based misclassification measure to evaluate the distance between a particular model of interest and a combination of many of its competing models. The error rate for a class characterized by the HMM is then the value of a smoothed zero-one error function given the misclassification measure. The overall error rate of the task vocabulary could then be computed as a function of all the available class error rates. The key here is to evaluate the misclassification measure in terms of the parameters of environmental-matched models without running recognition experiments, where the models are adapted by very limited data that could be just the testing utterance itself. In this paper, we show how the misclassification measure could be approximated by first computing the distance between two mixture Gaussian densities, then between two HMMs with mixture Gaussian state observation densities and finally between two sequences of HMMs. The misclassification measure is then converted into classification error rate. When comparing the error rate obtained in running actual experiments and that of the new framework, the proposed algorithm accurately estimates the classification error rate for many types of speech and speaker recognition problems. Based on the same framework, it is also demonstrated that the error rate of a recognition system in a noisy environment could also be predicted.

Speech model combination with the background noise has been shown effective to improve the pattern classification rate of noisy speech. The model combination can be performed by the addition of the spectral statistics such as the means and the variances. Since the speech feature for pattern classification has to be expressed in the cepstral domain, the combined spectral statistics have to be transferred into the cepstral domain for speech recognition. In our previous study, we have proposed a direct adaptation scheme of the cepstral variance that is without the mapping from the spectral domain to the cepstral domain. In this paper, an improved version to perform the adaptation is proposed. From the study, it is observed that the adapted variance can be expressed as a linear interpolation of the speech and the noise variances to obtain a comparable recognition rate that is obtained with the mapping process. Due to the direct adaptation of the variances, a lot of computation can be reduced to perform the environmental adaptation.

In this paper, we discuss the use of weighted filter bank analysis (WFBA) to increase the discriminating ability of mel frequency cepstral coefficients (MFCCs). The WFBA emphasizes the peak structure of the log filter bank energies (LFBEs) obtained from filter bank analysis while attenuating the components with lower energy in a simple, direct, and effective way. Experimental results for recognition of continuous Mandarin telephone speech indicate that the WFBA-based cepstral features are more robust than those derived by employing the standard filter bank analysis and some widely used cepstral liftering and frequency filtering schemes both in channel-distorted and noisy conditions.

In this paper, an signal-to-noise ratio (SNR)-incremental stochastic matching (SISM) algorithm is proposed for robust speech recognition in noisy environments. The SISM algorithm is an extension of Sankar and Lee's (1996) stochastic matching (SM) for dealing with the distortion due to additive noise. We address two issues concerning the original maximum likelihood-based SM techniques. One concern is that the initial condition of the expectation-maximization (EM) algorithm has to be set carefully if the mismatch between training and testing is large. The other is that the performance is often limited by the newly adapted model in noise compensation instead of reaching the higher level of accuracy often obtained in clean environments. Our proposed SISM algorithm attempts to improve the initial condition and to relax the performance bound. First, the SISM algorithm provides a good initial condition making use of a set of environment-matched models. The second is a recursive operation, i.e., the reference model in each recursion is changed along the direction of SNR increment in order to push the recognition performance to that obtained at higher SNR levels. Experimental results show that the SISM algorithm provides further improvement after the best environment-matched performance has been reached, and can therefore obtain an additional discriminative power through using the speech models with higher SNR instead of retraining process.

The output log energies from the band-pass filters of a filter bank are very sensitive to environmental variations. Therefore, inclusion of filter bank analysis (FBA) in the derivation of speech features often makes spectral comparisons unreliable. To remedy this problem, a fuzzy membership function is presented in this paper to perform an equalization of the variances of cepstral coefficients by weighting the frequency sequence of log filter-bank energies (LFBE) in a simple, direct but effective way. Experimental results for recognition of continuous Mandarin telephone speech showed that the proposed fuzzy filter bank analysis (FFBA) offers a significant improvement in the discrimination capability of cepstral features while maintaining a low computation cost.

Line spectrum frequency (LSF) is an efficient and reliable representation for speech signals. This property makes LSFs popular in low bit-rate speech coding. However, the use of LSF parameters in speech recognition still faces the problem of performance degradation in a noisy environment. This paper investigates the noise effect in the line spectrum domain and presents a model adaptation method for the improvement of noisy speech recognition. It comes out an interpolation mechanism in the line spectrum domain so that clean speech models can be adjusted according to the LSF of the noise signal. An experiment on Mandarin digit recognition has demonstrated the effectiveness and robustness of our proposed method.

Within the context of continuous-density HMM speech recognition in noise, we report on imputation of missing time-frequency regions using emission state probability distributions. Spectral subtraction and local signal--to-- noise estimation based criteria are used to separate the present from the missing components. We consider two approaches to the problem of classification with missing data: marginalization and data imputation. A formalism for data imputation based on the probability distributions of individual Hidden Markov model states is presented. We report on recognition experiments comparing state based data imputation to marginalization in the context of connected digit recognition of speech mixed with factory noise at various global signal-to-noise ratios, and wideband restoration of speech. Potential advantages of the approach are that it can be followed by conventional techniques like cepstral features or artificial neural networks for speech recognition. 1. INTRODUCTION T...

This paper presents a novel method using robust features for speech recognition when the speech signal is corrupted by additive and convolutional noises. This method is conceptually simple and easy to be implemented. The additive noise and the convolutional noise are removed by temporal trajectory filtering in the autocorrelation domain and cepstral domain, respectively. No prior information of noise corruption is necessary. A task of multi-speaker isolated digit recognition is conducted to demonstrate the effectiveness of using these robust features. The cases of the channel filtered speech signal corrupted by additive white noise and color noise are tested. Experimental results show that significant improvements can be achieved as compared with some traditional features.

We propose the phone-dependent channel compensated hidden Markov model (PDCC-HMM) for telephone speech recognition. The PDCC-HMM is derived by modifying the conventional hidden Markov model (HMM) with the phone-dependent channel compensation vectors. The telephone speech is recognized efficiently by using the derived PDCC-HMM. Experiments demonstrate the robustness of PDCC-HMM in speech recognition and show the significant reduction of recognition error rate by 50% compared to the conventional HMM method.

We propose an improved maximum a posteriori (MAP) learning algorithm of continuous-density hidden Markov model (CDHMM) parameters for speaker adaptation. The algorithm is developed by sequentially combining three adaptation approaches. First, the clusters of speaker-independent HMM parameters are locally transformed through a group of transformation functions. Then, the transformed HMM parameters are globally smoothed via the MAP adaptation. Within the MAP adaptation, the parameters of unseen units in adaptation data are further adapted by employing the transfer vector interpolation scheme. Experiments show that the combined algorithm converges rapidly and outperforms those other adaptation methods.

When a speech signal is contaminated by additive noise, its cepstral coefficients are assumed to be the functions of the noise power. By using Taylor series expansion with respect to the noise power, the cepstral vector can be approximated by a nominal vector plus the first derivative term. The nominal cepstrum corresponds to the clean speech signal and the first derivative term is a quantity used to adapt the speech feature to a noisy environment. A deviation vector is introduced to estimate the derivative term. The experiments show that the feature adaptation based on the deviation vectors is superior to those of projection based methods.

This paper proposes a Bayesian affine transformation of hidden Markov model (HMM) parameters for reducing the acoustic mismatch problem in telephone speech recognition. Our purpose is to transform the existing HMM parameters into its new version of specific telephone environment using affine function so as to improve the recognition rate. The maximum a posteriori (MAP) estimation which merges the prior statistics into transformation is applied for estimating the transformation parameters. Experiments demonstrate that the proposed Bayesian affine transformation is effective for instantaneous adaptation and supervised adaptation in telephone speech recognition. Model transformation using MAP estimation performs better than that using maximum-likelihood (ML) estimation.

We present a hybrid algorithm for adapting a set of speaker-independent hidden Markov models (HMMs) to a new speaker based on a combination of maximum a posteriori (MAP) parameter transformation and adaptation. The algorithm is developed by first transforming clusters of HMM parameters through a class of transformation functions. Then, the transformed HMM parameters are further smoothed via Bayesian adaptation. The proposed transformation/adaptation process can be iterated for any given amount of adaptation data, and it converges rapidly in terms of likelihood improvement. The algorithm also gives a better speech recognition performance than that obtained using transformation or adaptation alone for almost any practical amount of adaptation data.

Because the norm of testing cepstral vector was shrinked in a noisy environment, the model parameters, i.e., mean vector and covariance matrix, should be adapted simultaneously. We propose a method called variance adapted likelihood measure (VALM) which adapts the mean vector using a projection-based scale factor and adapts the covariance matrix using a variance reduction function estimated from the training database. The variance reduction function can be obtained according to various phonetic units. In the hidden Markov model based experiments, the speech recognition performance is greatly improved by applying VALM. The most significant improvement is achieved when the variance reduction function is separately estimated for different state parameters.

In this study, we propose a maximum a posterior (MAP) estimation of channel bias to compensate the channel mismatch in telephone speech recognition.For a telephone speech, the channel bias is estimated by maximizing a posterior probability.Because a posterior probability is composed of a likelihood function and a prior density, we introduce a scale factor to evaluate their weights in MAP estimation.To further improve the performance, a prior channel statistics is extended to multiple components and the channel mismatch is separately compensated for different segments.Besides, a rapid MAP estimation applied in feature domain is also proposed for reducing the computational complexity.Experiments show that proposed method can significantly improve recognition rates and computational complexity.

An extended Levinson-Durbin algorithm is proposed to recursively calculate the linear prediction coefficients and their derivatives with respect to the noise power for a noisy auto regressive process. This algorithm would be useful in some applications or analysis related to an autoregressive (AR) process that is contaminated by additive white noise.

Abstruct- In this correspondence, a two-stage approach based on Karhunen-Loeve transform and 2-D prediction is proposed for efficient quantization of line spectrum pair (LSP) parameters of speech. Besides, a switched classifier is incorporated with this approach to reduce the outlier frames (spectral distortion greater than 2 dB) down to about 0.27% and to eliminate frames with spectral distortion greater than 4 dB at an average bit-rate below 19 b/frame.

Absh-act-In this correspondence, we propose a new scoring method for speaker verification called the normalized log-likelihood score. This method is derived from the Bayes testfor minimum risk by the assumption of two hypotheses-the actual speaker is the claimed speaker or the actural speaker is an impostor-to attain the objective of king the probability of error. The performance of this new scoring fnnction used in speaker verification is examined by a series of experiments. For a 1Mspeaker database of isolated single digits, the equal error rate obtained by the normalized log-likelihood scoring method can be significantly decreased from 11.65 to 3.65% for the closed set test and from 12.30 to 8.22% for the open set test, as compared with those obtained by the conventional scoring method.

A segmental probabilistic model based on an orthogonal polynomial representation of speech signals is proposed. Unlike the conventional frame based probabilistic model, this segment based model concatenates the similar acoustic characteristics of consecutive frames into an acoustic segment and represents the segment by an orthogonal polynomial function. An algorithm which iteratively performs recognition and segmentation processes is proposed for estimating the parameters of the segment model. This segment model is applied in the text independent speaker verification. For a 20-speaker database, the experimental results show that the performance by using segment models is better than that by using the conventional frame based probabilistic model. The equal error rate can be reduced by 3.6% when the models are represented by 64-mixture density functions.

This correspondence proposes a new CELP coding method which embeds speech classification in adaptive codebook search. This approach can retain the synthesized speech quality at bit-rates below 4 kb/s. A pitch analyzer is designed to classify each frame by its periodicity, and with a finite-state machine, one of four states is determined. Then the adaptive codebook search scheme is switched according to the state. Simulation results show that higher SEGSNR and lower computation complexity can be achieved, and the pitch contour of the synthesized speech is smoother than that produced by conventional CELP coders.>

This paper presents a speech recognition method based on the distribution of two-dimensional cepstral (TDC) coefficients. For each recognition unit, a TDC matrix is calculated. A set of selected TDC coefficients forms a pattern to represent this speech segment. By assuming the Gaussian distribution of the TDC coefficients, a statistical model for a class of speech patterns is generated. The recognition process is to evaluate the probability of a TDC pattern belonging to a specific pattern class and to find the model which gives the highest probability. This method is applied to the recognition of Mandarin syllables. The experimental result shows that the proposed method is very promising for syllabic languages such as Mandarin.

An efficient search algorithm for generating syllable hypotheses in continuous Mandarin speech recognizer is proposed. The fast syllable hypothesizing algorithm consists of a forward, time-synchronous, modified Viterbi search and a parallel backtracking procedure. This algorithm can generate a lattice of syllable hypotheses with very low time delay. The superiority of the new algorithm to other direct sentence hypothesis algorithms is also experimentally demonstrated in terms of search efficiency and accuracy.>

Presents a fast algorithm for spotting a large number of keywords in unconstrained, continuous speech using an HMM-based continuous speech recognizer. This fast algorithm is based on a two stage scheme. In the first stage, the forward backward search is performed for detecting N most likely common subwords. In the second stage, the tree-trellis search is carried out to determine the optimum keyword by traversing the tree-structural vocabulary in an effective way. Compared with the conventional whole-word based keyword spotting algorithm, the proposed fast algorithm can drastically reduce the computational cost.>

A novel spectral coding method, two-dimensional differential line spectra pair coding (2DdLSP), is proposed. Taking advantage of the strong inter-frame, and intra-frame correlation of LSP parameters, a two-dimensional linear prediction technique is used to reduce the variance of the parameters to be quantized. One scalar quantization and two vector quantization schemes are designed to quantize the 2-D prediction residuals. Without further buffering delay, the spectral distortion of 1 dB/sup 2/ can be achieved at 19 b/frame when the frame period is 10 ms. Both within- and out-of-training tests show the robustness of the method to speech data variance.>

Four varieties of line spectrum pair (LSP) frequencies for speaker recognition applications and their corresponding performances are studied. The four varieties are all-LSP frequencies, odd or even line spectrum frequencies, the mean of adjacent LSP frequencies, and the difference of adjacent LSP frequencies. A speaker-based vector quantization approach is used in the evaluation. The results show that each variety of LSP frequencies has a very high accurate rate.>

This study purposes a method for recognizing the lexical tones in Mandarin speech. The method is based on Vector Quantization (VQ) and Hidden Markov Models (HMM). The pitch periods are extracted to derive the feature vectors which represent pitch height and pitch contour slope. One HMM is trained by the feature vectors of monosyllables for each tone. Then the HMMs are used to recognize the tone of monosyllables and disyllables. For the monosyllables, the accuracy rate can be 93.75% for speaker-independent cases. For the disyllables, the accuracy rates are 93% for the first syllables and 90% for the second syllables. It shows that the tone of the second syllable may be affected by the preceding syllable. This degradation also reveals the fact of tone variation in Mandarin speech.