Speech activity detection (SAD) is an important component for various speech processing applications and has been researched extensively recently. The pitch continuity, a significant characteristic of speech, however, has not successfully played a role in existing SAD methods. In this work, we propose a novel way to integrate the pitch continuity with pitch-related features. Practice is carried out through the Combo-SAD approach: We examine three consecutive frames and assume that they all have the same pitch as the center frame due to pitch continuity. Corresponding feature values are recomputed at the adjusted pitch location and then used in the final expression. The new combo feature is evaluated with various types of additive noise at different signal-to-noise ratios (SNR). The results show that the new feature leads to better SAD performance (with an up to 39.3% relative improvement on miss rate compared to Combo-SAD). We also introduce a novel variant of the underlying autocorrelation function and illustrate how it can improve the accuracy of pitch detection.

Conventional speech recognition engines extract Mel Frequency Cepstral Coefficients (MFCC) features from incoming speech. This paper presents a novel approach for feature extraction in which speech is processed according to the Auditory Image Model, a model of human psychoacoustics. We fist describe the proposed frontend, then we present recognition results obtained with the TIMIT database. Comparing with previously published results on the same task, the new approach achieves a 10% improvement in recognition accuracy.

In many speech recognition applications such as name dialing, it is necessary to have the ability to know when a recognition error has occurred so that undesired or unpredicted system behavior can be minimized. Con dence measure is usually used for detection of probable errors. In this paper, a new method for measuring condence is presented. The method is based on use of recursive mumble models. During a regular decoding from which word hypotheses and word boundaries are known, the score of recursive mumblemodels is then determined. The (weighted) di erence between the word detail-match score and the mumble score is used as the con dence measure. It is next compared to a prede ned threshold to decide whether the decoded result is con dently correct or not. The method has been evaluated with two di erent databases. The results show that the new method outperforms our previous method solely based on the word detail-match scores. In particular, the results show that the new method is able to reduce the equal error rate from 32% to 23% and that it rejects far more (78% versus 35%) out-of-domain sentences at the xed 5% false rejection rate.

This paper describes a new con dence measure based on rank-ordering subphone likelihood scores. The approach consists of three major steps. The rst step is standard decoding which hypothesizes a word string given an utterance. Then forced Viterbi alignment is made at the second step. (This step is of course not needed for a Viterbi decoder.) From the aligned sentence, the third step computes likelihood scores of the hypothesized subphone and all other competing subphones to generate a list of subphones in the descending order of the likelihood scores. A rank is assigned to the hypothesized subphone according to its positioning in the list. The rank value is then merged to obtain the corresponding rank at the phone level. After the merge, selective weighting is applied such that contribution of phones having large acoustic variations is de-emphasized. Additional upper-bound limiting is also made to guarantee a rank computation not to be contaminated by a very bad segment. The new con dence measure has been favorably evaluated on word conrmation/rejection experiments with a small vocabulary of many confusable and short words. More speci cally, experimental results show that the new approach outperforms other measures such as whole-word scores by reducing the equal error rate from 32% to 20%.

This paper presents a speaker recognition system based on hidden Markov models (HMM). The system utilizes concatenated phoneme HMMs and works in a text-prompted mode. Each registered speaker has a separate set of HMMs which are trained using the Baum-Welch algorithm. The speaker recognition system has been evaluated with the YOHO voice verification corpus in terms of both speaker verification and closed-set speaker identification. It is shown that by using 10 seconds of testing speech, an error rate of 0.09% for male and 0.29% for female are obtained for speaker identification with a total population of 138 talkers. For speaker verification, under the 0% false rejection condition, the system achieves a false acceptance rate of 0.09% for male and 0% for female. This paper also studies effects of various factors (such as the mixture number and cohort selection) on the performance of speaker recognition.

Most contemporary speech recognizers are designed to operate with close-talking speech and they work best in a quiet laboratory condition. There is an apparent need to render environment robustness to these systems. The objective of the paper is to explore utility of existing speech recognition technology in adverse "real-world" environments for distant-talking applications. A synergistic system consisting of microphone array and neural network (MANN) is utilized to mitigate environmental interference introduced by reverberation, ambient noise, and channel mismatch between training and testing conditions. The MANN system is evaluated with experiments on continuous distant-talking speech recognition. The results show that the MANN system elevates the word recognition accuracy to a level which is competitive with a retrained speech recognizer and that the neural network compensation performs better than some previously researched techniques.

Environment-independent continuous speech recognition is important for the successful development of speech recognizers in real world applications. Linear compensation methods do not work well if the mismatches between training; and testing environments are not linear. In this paper, a neural network compensation technique is explored to mitigate the distortion resulting from additive noise, distant-talking, or telephone channels. The advantage of the neural network compensation method is that retraining of a speech recognizer for each particular application is avoided. Furthermore, since neural networks are trained to transform distorted speech feature vectors to those corresponding to clean speech, it may outperform a retrained speech recognizer trained on distorted speech. Three experiments are conducted to evaluate the capability of the neural network compensation method; recognition of additive noisy speech, distant-talking speech, and telephone speech.

We overall vocal-tract length differs between individuals. The difference is more apparent between males and females. In this work, a new method is presented which normalizes the overall vocal-tract length in the cepstrum domain. By properly selecting the upper frequency of a FFT spectrum, the derived cepstrum coefficients exhibit a pattern which is less susceptible to the length variations. The proposed method is evaluated in isolated-word speech recognition experiments where a recognizer was exclusively trained on male speech and tested on female speech, and vice versa. The method elevated the word recognition accuracy from 57.0 to 78.8% and from 70.5 to 86.1%, respectively.>

This paper describes a fast algorithm that computes the impulse response of the vocal tract from its transfer function. First, numerical methods for computing the transfer function of a given vocal-tract configuration are briefly outlined. These methods include techniques (1) to decompose the numerator and denominator of the transfer function and (2) to efficiently determine the resonance modes of the vocal tract. Next, is a description of how to calculate residues at the poles and how to express the vocal-tract transfer function as a partial fraction expansion series. Each term in the expansion corresponds to an elementary formant generator, and the additive terms correspond to a parallel formant architecture. A second-order digital filter is derived for each formant generator. The impulse response of the vocal tract can therefore be specified compactly by a set of such filters. Good agreement is observed between the directly calculated transfer function and the one synthesized by the proposed algorithm. The algorithm is being used in the articulatory speech synthesizer under development both at Rutgers University and at the Royal Institute of Technology, Sweden. An ambitious goal is to incorporate the method into a text-to-speech synthesizer and/or an adaptive voice mimic system.

Hands-free operation of speech processing equipment is sometimes desired so that the user is unencumbered by hand-held or body-worn microphones. This paper explores the use of array microphones to capture speech under adverse acoustic conditions, and provide input to a system for automatic speaker identification. The system is evaluated using reverberated speech signals, generated by a computer model of room acoustics and transduced by different simulated microphone-arrays. For comparison, the system is also evaluated using close-talking microphone input. It is found that 2-D matched-filter microphone arrays are capable of producing high speaker identification scores in a hostile acoustic environment, such as multipath distortion and competing noise sources. The paper also explores the influence of vector quantization techniques, codebook size, and order of cepstrum coefficients on the performance of the speaker identification system.>

An articulation-based speech synthesis system is briefly described. The system first carries out a frequency-domain analysis of the vocal tract and then derives the counterpart in the time domain in terms of partial fraction expansion of the transfer function. The advantages of the approach are that the vocal-tract acoustics simulation is accurate and effective and that the synthesis computation is efficient. Furthermore, the system presents a conceptual compatibility with conventional formants synthesizers. The emphasis is on two detailed aspects of the synthesizer: the parametrization of the vocal tract in the area-function domain and optimal formant connection, to maximally preserve resonance continuity, when two contiguous segments have different numbers of formants.>

Voice source parameters are generally derived from time-domain inverse filtering. An effective system of parameterization is the LF model, which has been adopted for the authors' studies of voice source variations in connected speech. The effects of variations of the different LF parameters on the voice source spectrum are demonstrated. Dynamic variations within a linguistic frame and with respect to different speaker types are exemplified. The LF-model has been implemented in a text-to-speech system, and work on development of rules is progressing.>