There are several reasons why conventional speech recognition systems modeled on adult data fail to perform satisfactorily on children's speech input. For instance, children's vocal characteristics differ significantly from those of adults. In addition, their choices of vocabulary and sentence construction modalities usually do not conform to adult patterns. We describe comparative studies demonstrating the performance gain realized by adopting to children's acoustic and language model data to construct a children's speech recognition system.

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%.

Conventional speech recognition systems such as the IBM Tangora tend to be adversely influenced by external factors such as background noise and microphone characteristics. This paper discusses some adaptation techniques to counteract such influences. We also consider ways to reduce computation so that the methods may be implemented for real time Tangora operation. The adaptation strategy utilized two sets of VQ codebooks derived from the training data, one representing the ambience and the other characterizing the speech domain. A speech versus ambience decision was made by examining several factors, such as a comparison of the overall energy in a frame with a percentile point of a running energy histogram. We include results to demonstrate the effectiveness of our approach.>

The authors describe a series of experiments in which the phonetic baseform is deduced automatically for new words by utilizing actual utterances of the new word in conjunction with a set of automatically derived spelling-to-sound rules. Recognition performance was evaluated on new words spoken by two different speakers when the phonetic baseforms were extracted via the above approach. The error rates on these new words were found to be comparable to or better than when the phonetic baseforms were derived by hand, thus validating the basic approach.>

Given sufficient samples of data tagged with their class identities, three techniques for constructing supervised prototypes to represent these classes are examined. The first method consists of averaging the tokens of each class separately to obtain the prototypes. In the second approach, several tokens, picked uniformly from each class, are designated as prototypes. The third technique involves a systematic search procedure to select effective prototypes and discard obsolete ones. Approximately two hours of continuous speech data from each of two speakers were used for experimentation. Each centisecond frame of speech was labeled with one of 200 phonetic subunit names utilizing hidden Markov model training and Viterbi alignment procedures. Prototypes were determined from the first part of the data, whereas the last part served to measure the classification performance. Average accuracies ranged from 24.2% with 200 prototypes in the first, to 31.5% with 32000 prototypes in the second, to 38.5% with 2258 prototypes in the third method.>

A description is presented of the authors' current research on automatic speech recognition of continuously read sentences from a naturally-occurring corpus: office correspondence. The recognition system combines features from their current isolated-word recognition system and from their previously developed continuous-speech recognition system. It consists of an acoustic processor, an acoustic channel model, a language model, and a linguistic decoder. Some new features in the recognizer relative to the isolated-word speech recognition system include the use of a fast match to prune rapidly to a manageable number the candidates considered by the detailed match, multiple pronunciations of all function words, and modeling of interphone coarticulatory behavior. The authors recorded training and test data from a set of ten male talkers. The perplexity of the test sentences was found to be 93; none of sentences was part of the data used to generate the language model. Preliminary (speaker-dependent) recognition results on these talkers yielded an average word error rate of 11.0%.>

The Speech Recognition Group at IBM Research in Yorktown Heights has designed a real-time, isolated-utterance speech recognizer for natural language with a 5,000-word vocabulary based on the IBM Personal Computer (PC) AT model and two IBM Signal Processors realized in VLSI technology. The enrollment period for a new user is approximately 20 minutes. The basic vocabulary is chosen from the most common words in several collections of documents such as office memoranda and business letters. The system supports spelling and interactive personalization to augment this vocabulary. Signal processing, vector quantization, and acoustic matching algorithms are programmed on the IBM Signal Processors which fit into the PC AT chassis. The PC AT controls the Processors and implements the decoder stack search and the language model, as well as the application-specific interface. The modular architecture of the design is expandable to a 20,000-word vocabulary system by the addition of two more IBM Signal Processors housed in a PC Expansion Unit.

This paper deals with two experiments with a large vocabulary isolated word recognizer. The first compares word error rates for 1) meaningful sentences belonging to actual documents and 2) random word lists from the same vocabulary. The error rate is considerably lower for random word lists. The second experiment investigates the performance of the recognition system on sentences containing words outside the vocabulary of the recognizer. Sentences from a 5000 word vocabulary task are recognized with a recognizer limited to a 2000 word subvocabulary. The error rate is only slightly higher than it would be if recognition of the full 5000 word vocabulary was allowed.

The IBM continuous speech recognition system has relied usually on 80- or 81-dimensional frequency-domain output every centisecond from the signal processing stages to generate its training and test data. On a controlled one-speaker data-base, a performance measure of 90.7 percent word-level accuracy was achieved using this type of data. For reasons of computational advantages, an investigation was carried out to determine a suitable method for reducing this dimensionality to 30 with minimal loss in accuracy. In one study, eigenvectors derived from the covariance matrix of 81- dimensional data were utilized to optimally rotate the data down to 30 dimensions. Two different variations of this experiment, the speaker-dependent and the speaker-independent cases, were attempted. In the other study, a more traditional approach of dividing the relevant frequency domain into 30 separate bands was investigated. The results of these studies indicated that the latter approach was marginally superior in performance to either of the two eigenvector techniques, and, in fact, accomplished the desired data reduction with no loss in accuracy.

The discrete cosine transform (DCT) algorithm and an adaptive transform coder (ATC) for speech are implemented in a 16-bit fixed point arithmetic microprocessor. Appropriate normalization of the data at various stages of processing is shown to be important for proper operation of the transform and coder algorithms. Other departures from conventional procedures are outlined. Several male and female speakers take part in the evaluation of the speech coder. Recordings representative of the coder performance are presented.

This paper is concerned with the following three aspects of the discrete utterance recognition problem: utterance normalization, dynamic programming algorithm implementation, and boundary error effects. Performance sensitivity as a function of each aspect of the problem is comparatively studied utilizing several available alternatives and significant conclusions are drawn regarding each of them. The concept of proportional normalizing is introduced as an effective method of handling the utterance normalization problem. A database consisting of the utterances of the alpha-digit vocabulary produced by several male and female speakers is used to conduct all the experiments.

A 1-bit dynamic programming algorithm is introduced where the path representing the warping function is coded as one bit per sample point. This algorithm is then applied to the task of speech coding. The resulting coded data consist primarily of two types of blocks -- the reference block and the dynamic programming path block. The data in the reference block represent a section of speech coded by a conventional adaptive differential pulse code modulation (ADPCM) technique. The data in the dynamic programming path block correspond to a 1-bit per sample representation of the warping function resulting from matching a subsequent section of the speech with the ADPCM encoded section. A side benefit of the coding procedure is the realization of a method of pitch tracking.

Some inequalities are derived between the Kullback divergence, the Bhattacharyya coefficient, the Matusita distance, and the Kolmogorov variational distance.

A modification of conventional training techniques for linear threshold dichotomous pattern recognition is described. The modification is capable of convergence in some cases where the classical training techniques are not. The utility of this technique for signal detection problems is discussed.

Features which are linearly dependent on other features do not contribute toward pattern classification by linear techniques. In order to detect the linearly dependent features, a measure of linear dependence is proposed. This measure is used as an aid in feature selection. Speaker verification experiments demonstrate the usefulness of employing such a measure in practice.

A method of decision making which leads to the formation of a gain vector by utilizing a design set of patterns from only the class to be recognized is proposed. Some comparisons of this method with the adaline technique have been made and the method has been applied to the practical problem of speaker verification.