String Error Rate Research Articles

Hidden Markov Modeling-Based Decision-Making Using Short-Length Sensor Time Series

Real-time decision-making (e.g., monitoring and active control of dynamical systems) often requires feature extraction and pattern classification from short-length time series of sensor data. An example is thermoacoustic instabilities (TAI) in combustion systems, caused by spontaneous excitation of one or more natural modes of acoustic waves. The TAI are typically manifested by large-amplitude self-sustained pressure oscillations in time scales of milliseconds, which need to be mitigated by fast actuation of the control signals, requiring early detection of the forthcoming TAI. This issue is addressed in this technical brief by hidden Markov modeling (HMM) and symbolic time series analysis (STSA) for near-real-time recognition of anomalous patterns from short-length time series of sensor data. An STSA technique is first proposed, which utilizes a novel HMM-based partitioning method to symbolize the time series by using the Viterbi algorithm. Given the observed time series and a hidden Markov model, the algorithm generates a symbol string with maximum posterior probability. This symbol string is optimal in the sense of minimizing string error rates in the HMM framework. Then, an HMM likelihood-based detection algorithm is formulated and its performance is evaluated by comparison with the proposed STSA-based algorithm as a benchmark. The algorithms have been validated on a laboratory-scale experimental apparatus. The following conclusions are drawn from the experimental results: (1) superiority of the proposed STSA method over standard methods in STSA for capturing the dynamical behavior of the underlying process, based on short-length time series and (2) superiority of the proposed HMM likelihood-based algorithm over the proposed STSA method for different lengths of sensor time series.

Robust endpoint detection and energy normalization for real-time speech and speaker recognition

When automatic speech recognition (ASR) and speaker verification (SV) are applied in adverse acoustic environments, endpoint detection and energy normalization can be crucial to the functioning of both systems. In low signal-to-noise ratio (SNR) and nonstationary environments, conventional approaches to endpoint detection and energy normalization often fail and ASR performances usually degrade dramatically. The purpose of this paper is to address the endpoint problem. For ASR, we propose a real-time approach. It uses an optimal filter plus a three-state transition diagram for endpoint detection. The filter is designed utilizing several criteria to ensure accuracy and robustness. It has almost invariant response at various background noise levels. The detected endpoints are then applied to energy normalization sequentially. Evaluation results show that the proposed algorithm significantly reduces the string error rates in low SNR situations. The reduction rates even exceed 50% in several evaluated databases. For SV, we propose a batch-mode approach. It uses the optimal filter plus a two-mixture energy model for endpoint detection. The experiments show that the batch-mode algorithm can detect endpoints as accurately as using HMM forced alignment while the proposed one has much less computational complexity.

Integrated bias removal techniques for robust speech recognition

In this paper, we present a family of maximum likelihood (ML) techniques that aim at reducing an acoustic mismatch between the training and testing conditions of hidden Markov model (HMM)-based automatic speech recognition (ASR) systems. Our study is conducted in two phases. In the first phase, we evaluate two classes of robustness techniques; those that represent the acoustic mismatch for the entire utterance as a single additive bias and those that represent the mismatch as a non-stationary bias. In the second phase, we propose a codebook-based stochastic matching (CBSM) approach for bias removal both at the feature level and at the model level. CBSM associates each bias with an ensemble of HMM mixture components that share similar acoustic characteristics. It is integrated with hierarchical signal bias removal and further extended to account for n -best candidates. Experimental results on connected digits, recorded over a cellular network, shows that incorporating bias removal reduces both the word and string error rates by about 12% and 16%, respectively, when using a global bias, and 36% and 31%, respectively, when using a non-stationary bias.

Formant estimation for speech recognition

This paper presents a new method for estimating formant frequencies. The formant model is based on a digital resonator. Each resonator represents a segment of the short-time power spectrum. The complete spectrum is modeled by a set of digital resonators connected in parallel. An algorithm based on dynamic programming produces both the model parameters and the segment boundaries that optimally match the spectrum. We used this method in experimental tests that were carried out on the TI digit string data base. The main results of the experimental tests are: (1) the presented approach produces reliable estimates of formant frequencies across a wide range of sounds and speakers; and (2) the estimated formant frequencies were used in a number of variants for recognition. The best set-up resulted in a string error rate of 4.2% on the adult corpus of the TI digit string data base.

Discriminative utterance verification for connected digits recognition

Utterance verification represents an important technology in the design of user-friendly speech recognition systems. It involves the recognition of keyword strings and the rejection of nonkeyword strings. This paper describes a hidden Markov model-based (HMM-based) utterance verification system using the framework of statistical hypothesis testing. The two major issues on how to design keyword and string scoring criteria are addressed. For keyword verification, different alternative hypotheses are proposed based on the scores of antikeyword models and a general acoustic filler model. For string verification, different measures are proposed with the objective of detecting nonvocabulary word strings and possibly erroneous strings (so-called putative errors). This paper also motivates the need for discriminative hypothesis testing in verification. One such approach based on minimum classification error training is investigated in detail. When the proposed verification technique was integrated into a state-of-the-art connected digit recognition system, the string error rate for valid digit strings was found to decrease by 57% when setting the rejection rate to 5%. Furthermore, the system was able to correctly reject over 99.9% of nonvocabulary word strings.

String-based minimum verification error (SB-MVE) training for speech recognition

In recent years, we have experienced an increasing demand for speech recognition technology to be utilized in various real-world applications, such as name dialling, message retrieval, etc. During this process, we have learned that the performance of speech recognition systems under laboratory environment cannot be duplicated in the actual service. Two major causes have been identified to this problem. The first is the lack of robustnesswhen the acoustic conditions in testing are different from those in training. The second is the lack of flexibilitywhen handling spontaneous speech input which often contains extraneous speech in addition to the desired speech segments of key phrases. This paper focuses on one aspect of achieving flexible speech recognition, namely, improving the ability to cope with naturally spoken utterances through discriminative utterance verification. We propose an algorithm for training utterance verification systems based on the minimum verification error (MVE) training framework. Experimental results on speaker-independent telephone-based connected digits show a significant improvement in verification accuracy when the discriminant function used in MVE training is made consistent with the confidence measure used in utterance verification. At a 10% rejection rate, for example, the new proposed method reduces the string error rate by a further 22·7% over our previously reported results in which the MVE-based discriminative training was not incorporated.

Robust parametric modeling of durations in hidden Markov models

A major weakness of conventional hidden Markov models is that they implicitly model state durations by a geometric distribution, which is usually inappropriate. This paper presents a modified Viterbi algorithm that, by incorporating proper state and word duration modeling, significantly reduces the string error rate of the conventional Viterbi algorithm for a speaker-independent, connected-digit string task. The algorithm has essentially the same computational requirements of the conventional Viterbi algorithm.

High-performance connected digit recognition using maximum mutual information estimation

Hidden markov models (HMM's) are one of the most powerful speech recognition tools available today. Even so, the inadequacies of HMM's as a "correct" modeling framework for speech are well known. In this context, it is argued in this paper that the maximum mutual information estimation (MMIE) formulation for training is more appropriate than maximum likelihood estimation (MLE) for reducing the error rate. Corrective MMIE training is introduced. It is a very efficient new training algorithm which uses a modified version of a discrete reestimation formula recently proposed by Gopalakrishnan et al.( see IEEE Trans. Inform. Theory, Jan. 1991). Reestimation formulas are proposed for the case of diagonal Gaussian densities and their convergence properties are experimentally demonstrated. A description of how these formulas are integrated into our training algorithm is given. Using the MMIE framework for training, it is shown how weighting the contribution of different parameter sets in the computation of output probabilities introduces substantial recognition improvements. Using the TIDIGITS connected digit corpus, a large number of experiments are performed with the ideas, techniques, and algorithms presented in this paper. These experiments show that MMIE systematically provides substantial error rate reductions with respect to MLE alone and that, thanks to the new training techniques, these results can be obtained at an acceptable computational cost. The best results obtained in the experiments were 0.29% word error rate and 0.89% string error rate on the adult portion of the corpus.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

An N-best candidates-based discriminative training for speech recognition applications

The authors propose an N-best candidates-based discriminative training procedure for constructing high-performance HMM speech recognizers. The algorithm has two distinct features: N-best hypotheses are used for training discriminative models; and a new frame-level loss function is minimized to improve the separation between the correct and incorrect hypotheses. The N-best candidates are decoded based on their recently proposed tree-trellis fast search algorithm. The new frame-level loss function, which is defined as a halfwave rectified log-likelihood difference between the correct and competing hypotheses, is minimized over all training tokens. The minimization is carried out by adjusting the HMM parameters along a gradient descent direction. Two speech recognition applications have been tested, including a speaker independent, small vocabulary (ten Mandarin Chinese digits), continuous speech recognition, and a speaker-trained, large vocabulary (5000 commonly used Chinese words), isolated word recognition. Significant performance improvement over the traditional maximum likelihood trained HMMs has been obtained. In the connected Chinese digit recognition experiment, the string error rate is reduced from 17.0 to 10.8% for unknown length decoding and from 8.2 to 5.2% for known length decoding. In the large vocabulary, isolated word recognition experiment, the recognition error rate is reduced from 7.2 to 3.8%. Additionally, they have found that using more relaxed decoding constraints in preparing N-best hypotheses yields better recognition results. >

Context‐dependent acoustic subword modeling for connected digit recognition

Accurate and robust connected digit recognition is essential for a wide range of telecommunication services. Based on training and testing using only clean network digit data, and using the same whole‐word model architecture as in the TI/NIST connected digit testing, the string error rate increased from less than 1% to more than 5%. The performance degraded even further when evaluated on data collected with different network conditions. Most of the observed errors were caused by changing channel characteristics, highly variable digit pronunciations, and inadequate modeling of cross‐digit coarticulation. Results are presented for a number of context‐dependent whole‐word and subword modeling techniques developed to overcome some of the above problems. The most effective one is a new acoustic subword modeling approach that assumes that each digit model consists of three parts, namely, head, body, and tail subword units. Multiple heads and tails are also allowed, one for each of the 11 possible preceding and following digits and the background. Cross‐digit coarticulation is modeled by connecting the pair of digits through the corresponding tail and head units. Testing on about 12 000 digit strings, collected from five regions, this new model architecture reduced the string error rate to under 2%.

A coarticulation model for continuous digit recognition

Between‐word coarticulation is one of the major problems in continuous speech recognition since it modifies the acoustic characteristics at the boundaries of words. With large vocabulary speech recognition the problem can be solved by introducing the concept of interword units. If the vocabulary is small enough (e.g., digits) all possible coarticulations between all words of the vocabulary can be modeled. In this study every digit is represented by three segments, namely, a core segment that can be assumed reasonably insensitive to any coarticulation effect and head and tail segments that represent, respectively, the initial and the final part of every work spoken in isolation. In addition, a set of juncture segments is defined that represent the junction between every possible pair of words. The recognition process is driven by a regular grammar that represents all the allowed segments sequences, namely, digits spoken in isolation as well as sequences of digits spoken continuously or with pauses between words. Every segment is represented by a mixture density HMM. Since the vocabulary is composed of 11 words (digits 0 to 9 and oh) the overall number of models is 154 (11 cores, 11 heads, 11 tails, and 121 junctures). A number of experiments were carried out using the TI connected digits database, which consists of a set of digit strings. Recognition was performed with different experimental conditions, i.e., varying the number of mixture components, the number of models per segment, and the structure of the connection between words. String error rates of 3.27% and 1.97% were obtained for the best and the two best hypothesized strings. These results compare favorably with those obtained with whole word modeling.

High performance connected digit recognition using hidden Markov models

The authors use an enhanced analysis feature set consisting of both instantaneous and transitional spectral information and test the hidden-Markov-model (HMM)-based connected-digit recognizer in speaker-trained, multispeaker, and speaker-independent modes. For the evaluation, both a 50-talker connected-digit database recorded over local, dialed-up telephone lines, and the Texas Instruments, 225-adult-talker, connected-digits database are used. Using these databases, the performance achieved was 0.35, 1.65, and 1.75% string error rates for known-length strings, for speaker-trained, multispeaker, and speaker-independent modes, respectively, and 0.78, 2.85, and 2.94% string error rates for unknown-length strings of up to seven digits in length for the three modes. Several experiments were carried out to determine the best set of conditions (e.g., training, recognition, parameters, etc.) for recognition of digits. The results and the interpretation of these experiments are described. >

Connected digit recognition using a level-building DTW algorithm

In this paper we present a novel method for recognizing a string of connected digits based upon the use of a recently proposed level-building dynamic time warping (DTW) algorithm. The recognition system attempts to build up the string, level-by-level (i.e., digit-by-digit), by comparing portions of the test string to isolated digit reference patterns. A backtracking procedure is used to find the best string (i.e., minimum accumulated distance) as well as a set of reasonable alternative candidates. The system was tested on a number of talkers speaking variable length digit strings (from two to five digits) over dialed up telephone lines. String error rates of 4.8 percent and 4.6 percent were obtained for speaker-trained and speaker-independent systems. Word error rates of 0.7 percent (for speaker-trained tests) and 0.9 percent (for speaker-independant tests) were obtained. The digit reference templates were obtained from autocorrelation averaging of a pair of isolated word templates for each digit of the speaker-trained system, and from a clustering analysis of isolated words for the speaker-independent system.

Application of dynamic time warping to connected digit recognition

A connected digit recognizer is proposed in which a set of isolated word templates is used as reference patterns and an unconstrained dynamic time warping (DTW) algorithm is used to literally spot the digits in the string. Segmentation boundaries between digits are obtained as the termination point of the dynamic path from the previous time warp. A region around the boundary is searched for the optimum starting point for the succeeding digit. At each stage the recognizer keeps track of a set of candidate digit strings for each test string. The string with the smallest accumulated distance is used as the preliminary string estimate. To help improve the recognition accuracy, two post-correction techniques were applied to the entire set of hypothesized digit strings. One technique creates a reference string by concatenating reference contours of the digits of the string, and comparing this to the test string using a constrained dynamic time warping algorithm. The second technique performs a similar comparison using voiced-unvoiced-silence contours instead of the measured features. Small but consistent improvements in recognition accuracy have been obtained using these techniques for both speaker-trained and speaker-independent systems with digit strings recorded over dialed-up telephone lines. For variable length digit strings of from 2 to 5 digits (where the recognizer was not told the length of the string), word error rates of about 2-3 percent and string error rates on the order of 8 percent were obtained for both speaker-dependent and speaker-independent systems.