Input Speech Segment Research Articles

A fast hierarchical search algorithm for discriminative keyword spotting

A keyword spotter can be considered as a binary classifier which classifies a set of uttered sentences into two groups on the basis of whether they contain target keywords or not. For this classification task, the keyword spotter needs to identify the target keywords locations based on a fast and accurate search algorithm. In our previous works, we exploited a modified Viterbi (M-Viterbi) search algorithm which has two known drawbacks. First, to locate the target keywords, it runs an exhaustive search through all possible segments of input speech. Second, while computing the start and end time-frames of each new phone, it makes the keyword spotter to trace-back and re-evaluate the timing alignments of all previous one(s), despite the fact that very limited amount of data -if any- would get updated as a result. These two pitfalls cause a dramatically enlarged search space as well as a significant increase in computational complexity. In this paper, we propose a Hierarchical Search (H-Search) algorithm which allows the system to ignore some segments of input speech at each level of hierarchy, according to their lower likelihood of containing the target keywords. In addition, unlike the M-Viterbi algorithm, the H-Search algorithm does not demand repeated evaluations when computing the current phone alignment which, in turn, results in a narrowed-down search space (O(TP) versus O(TPLmax) – where T is number of frames, P is number of keyword phones and Lmax is the maximum phone duration) as well as a decreased computational complexity (O(TPLmax) versus O(TPLmax3)) compared to those of the M-Viterbi algorithm. We applied the H-Search algorithm to the classification part of an Evolutionary Discriminative Keyword Spotting (EDKWS) system introduced in our previous works. The experimental results indicate that the H-Search algorithm is executed 100 times faster than the M-Viterbi algorithm while the performance of the EDKWS system degrades no more than two percent compared to that of the M-Viterbi algorithm.

Speaker Clustering Using Decision Tree-Based Phone Cluster Models With Multi-Space Probability Distributions

This paper presents an approach to speaker clustering using decision tree-based phone cluster models (DT-PCMs). In this approach, phone clustering is first applied to construct the universal phone cluster models to accommodate acoustic characteristics from different speakers. Since pitch feature is highly speaker-related and beneficial for speaker identification, the decision trees based on multi-space probability distributions (MSDs), useful to model both pitch and cepstral features for voiced and unvoiced speech simultaneously, are constructed. In speaker clustering based on DT-PCMs, contextual, phonetic, and prosodic features of each input speech segment is used to select the speaker-related MSDs from the MSD decision trees to construct the initial phone cluster models. The maximum-likelihood linear regression (MLLR) method is then employed to adapt the initial models to the speaker-adapted phone cluster models according to the input speech segment. Finally, the agglomerative clustering algorithm is applied on all speaker-adapted phone cluster models, each representing one input speech segment, for speaker clustering. In addition, an efficient estimation method for phone model merging is proposed for model parameter combination. Experimental results show that the MSD-based DT-PCMs outperform the conventional GMM- and HMM-based approaches for speaker clustering on the RT09 tasks.

Development of a 4 kbit/s hybrid sinusoidal/CELP speech coder

A comprehensive performance analysis of sinusoidal and code excited linear prediction (CELP) speech coding is given around 4 kbit/s, using both subjective and objective measurements. Based on the observations made, justification for the multi-modal hybrid coding approach employing both sinusoidal and CELP coding is given, and an implementation of such a coder is described. This 4 kbit/s sinusoidal/CELP speech coder utilizes four modes to classify the input speech segment: voiced, jittery-voiced, plosive and unvoiced. For voiced segments sinusoidal coding is used whereas different CELP versions are employed for the other modes. The quality of the implemented 4 kbit/s sinusoidal/CELP speech coder in clean speech conditions is finally verified by a listening test. In the test, the 4 kbit/s coder performed almost as well as the high-quality references used, but it still needs improvements to be classified as a high-quality 4 kbit/s speech coder.

Apparatus and methods for speech recognition including individual or speaker class dependent decoding history caches for fast word acceptance or rejection

A method and an apparatus are provided for performing speech recognition on speech segments frequently input by a user. The method and the apparatus include use of keyword scoring in connection with a speech recognition vocabulary, a temporary score, and a predetermined margin to determine an appropriate output as being representative of the input speech segment.

Parallel pattern verifier with dynamic time warping

A speech recognition system is disclosed which employs a network of elementary local decision modules for matching an observed time-varying speech pattern against all possible time warpings of the stored prototype patterns. For each elementary speech segment, an elementary recognizer provides a score indicating the degree of correlation of the input speech segment with stored spectral patterns. Each local decision module receives the results of the elementary recognizer and, at the same time, receives an input from selected ones of the other local decision modules. Each local decision module specializes in a particular node in the network wherein each node matches the probability of how well the input segment of speech matches the particular sound segments in the sounds of the words spoken. Each local decision module takes the prior decisions of all preceding sound segments which are input from the other local decision modules and makes a selection of the locally optimum time warping to be permitted. By this selection technique, each speech segment is stretched or compressed by an arbitrary, nonlinear function based on the control of the interconnections of the other local decision modules to a particular local decision module. Each local decision module includes an accumulator memory which stores the logarithmic probabilities of the current observation which is conditional upon the internal event specified by a word to be matched or identifier of the particular pattern that corresponds to the subject node for that particular pattern. For each observation, these probabilities are computed and loaded into the accumulator memory of all the modules and, the result of the locally optimum time warping representing the accumulated score or network path to a node for the word with the highest probability is chosen.

Very low bit rate speech coding based on joint segmentation and variable length segment quantizer

This paper presents a very low bit rate LPC vocoder based on a joint segmentation and quantization method using spectral segments having variable time length. The method exploits the nonuniform distribution of speech characteristics in the time and spectral domains. A measure of the spectral distance between a variable‐length input speech segment and a fixed‐length segment template is introduced based on linear time warping. The optimum segment boundaries and templates for a spectral sequence are efficiently determined using a dynamic programming technique so that the total spectral distortion in a voice interval is minimized. The segment templates are obtained by a sub‐optimum pattern learning method, which guarantees a monotonic decrease in distortion, using a combined segmentation and clustering technique. Experimental results for a single male speaker show that this method reduces the initial distortion by 20% and yields a sound articulation score of 78%.

Signal-dependent matching for isolated word speech recognition systems

A new approach to the design of IWSR systems is proposed in this paper. This involves a dynamic matching strategy based on the nature of the input speech segment. This is called signal-dependent matching. The computational complexity in the implementation of the proposed algorithm is significantly reduced by adopting a two stage approach in matching. In the first stage, the warping path between the test utterance and a reference utterance is determined. In the second stage, the distance between the utterances is computed along the path. There will be a slight degradation in the performance of a two stage approach as compared to the single stage approach, but this can be tolerated in view of the significant computational advantage. The performance degradation is more than compensated by the signal-dependent matching strategy in the second stage. To measure the improvement in the recognition performance, a new index of performance is defined, that reflects the characteristics of the distance matrix for a given vocabulary, rather than the characteristics of the confusion matrix. The performance of the signal-dependent matching algorithm is significantly better than the standard dynamic time warping matching algorithm for confusable as well as nonconfusable vocabulary. We also develop a signal-dependent matching algorithm, which takes into account some distortions in the input speech. As an example we offer the agorithm twice the same test utterance, once undistorted, once after a distortion. Our research until now indicates a improvement in automatic isolated word speech recognition systems while using signal-dependent parameter measuring and signal dependent matching.

A synthesis-based method for pitch extraction

A synthesis-based method for pitch extraction of the speech signal is proposed. The method synthesizes a number of log power spectra for different values of fundamental frequency and compares them with the log power spectrum of the input speech segment. The average magnitude (AM) difference between the two spectra is used for comparison. The value of fundamental frequency that gives the minimum AM difference between the synthesized spectrum and the input spectrum is chosen as the estimated value of fundamental frequency. The voiced/unvoiced decision is made on the basis of the value of the AM difference at the minimum. For synthesizing the log power spectrum, the speech signal is assumed to be the output of an all-pole filter. The transfer function of the all-pole filter is estimated from the input speech segment by using the autocorrelation method of linear prediction. The synthesis-based method is tried out on real speech data and the results are discussed.