Abstract: Speaker independent recognition of small vocabularies, spoken over the long distance telephone network, is achieved using two types of models, one type for defined vocabulary words (e.g., collect, calling-card, person, third-number and operator), and one type for extraneous input which ranges from non-speech sounds to groups of non-vocabulary words (e.g. `I want to make a collect call please`). For this type of key word spotting, modifications are made to a connected word speech recognition algorithm based on state-transitional (hidden Markov) models which allow it to recognize words from a pre-defined vocabulary list spoken in an unconstrained fashion. Statistical models of both the actual vocabulary words and the extraneous speech and background noises are created. A syntax-driven connected word recognition system is then used to find the best sequence of extraneous input and vocabulary word models for matching the actual input speech.

Abstract: An arrangement and a method for recognizing a speech pattern employs a quick initial processing based upon subpatterns of speech features from which time-sequence information has been removed is followed by more conventional processing of a plurality of the best candidates. The initial processing, or pre-selection, eliminates large numbers of unlikely candidates, the best candidates being retained by setting an empirically determined threshold for similarity, as measured by a suitable distance measurement. To facilitate retrieving the complete speech feature patterns including time-sequence information for the subsequent processing, the code books of the subpatterns also contain associated indices pointing to the corresponding complete patterns.

Abstract: An arrangement for endpoint detection improves speech recognition accuracy where the input signal includes nonstationary noise. Energy pulses are found by looking for local energy level peaks, then analyzing surrounding energy levels to determine pulse boundaries. Energy pulses are combined according to predetermined criteria to form longer pulses corresponding to words or phrases in the input signal.

Abstract: Markov model speech pattern templates are formed for speech analysis systems by analyzing identified speech patterns to generate frame sequences of acoustic feature signals representative thereof. The speech pattern template is produced by iteratively generating succeeding Markov model signal sets starting with an initial Markov model signal set. Each iteration includes forming a set of signals representative of the current iteration Markov model of the identified speech pattern responsive to said frame sequences of acoustic feature signals and one of the previous Markov model signal sets and comparing the current iteration Markov model signal set with said previous Markov model signal set to generate a signal corresponding to the similarity therebetween. The iterations are terminated when said similarity signal is equal to or smaller than a predetermined value and the last formed Markov model signal set is selected as a reference template for said identified speech pattern.

Abstract: In a speech recognition arrangement, a speech pattern is recognized as one of a plurality of reference patterns for which acoustic feature signal templates are stored. Each template includes a time frame (e.g., 10 millisecond) sequence of spectral features e.g., LPC and nonspectral e.g., acoustic energy (E) normalized to the peak energy over an utterance interval. LPC and normalized energy feature signal sequences are produced for an unknown speech pattern. For each time frame, the correspondence between the LPC features of the speech pattern and each reference is measured as well as the correspodence between the energy (E) features. In comparing the unknown speech features to those of the reference templates, the dynamic time warp distance DT=D.sub.LPC +.alpha.D.sub.E is used where .alpha. is a weighting factor selected to minimize the probability of erroneous recognition.

Abstract: In a speech recognition system, dynamic time warp (DTW) calculations are reduced by use of a search strategy for an optimal non-linear search warp function path: signals predictive of the similarity are generated responsive to the preceding and current correspondence signals of the input word to reference warped words.

Abstract: A system for generating speech pattern templates for use with either speech recognition or speech synthesis. Reference demisyllable templates are first generated from a reference first speaker using both manual and automatic analysis. The analysis for a second speaker is simplified and automated by comparing with the first speaker's templates. The second speaker speaks the same words at a rate time-warped to match the first speakers rate and template. We define a demisyllable as each of the two halves of a syllable, assuming a syllable starts and ends with a noisy consonant, and the syllable is split at its vowel center, thereby simplifying concatenation and comparison. Key features of the invention include generating a set of signals representative of the time alignment between the first and second speaker's templates, and the time-of-occurence boundaries of each syllable in a word.

Abstract: This speech recognizer concatenates a string of reference isolated-words for comparison with the unknown string of connected-words. The invention includes a level-building (LB) algorithm, "level" implying a location in a sequence of words. A constrained endpoint dynamic-time-warp algorithm, in which the slope of the warping function is restricted between 1/2 and 2, is used to find the best alignment between an unknown continuous-word test pattern, and a concatenated sequence of L reference patterns. Properties of the LB algorithm include: modification of the references; back-track decision logic; heuristic selection of multiple candidates, and syntax constraints. As a result, the processing required is less than two-level dynamic-program-matching and sampling algorithms.

Abstract: An arrangement for endpoint detection improves speech recognition accuracy and lowers rejection rates by developing an ordered list of endpoint candidates. A triple thresholding technique defines energy signal pulses. The energy pulses are combined according to predetermined criteria to form the endpoint candidates.

Abstract: Recognition of continuous speech by comparison with prestored isolated words may be confused by the merging together of spoken adjacent words (coarticulation). Improved recognition is attained by generating overlap-words, e.g., words whose first phoneme is the end phoneme of the preceding word in a string of words. The reference candidate series of overlap-words is transformed under dynamic time warping so as to time-match the utterance series of overlap-words.

Abstract: A speech controlled dialing circuit identifies input utterances which may be a command word (mode select), repertory word (dialing name or number), or non-recognized ("Other"). Responsive to the identification of each occurring input utterance, a set of predetermined templates are selected to identify the next occuring utterance. A programmed microprocessor system is described to implement the main controller function.

Abstract: A speech analyzer for recognizing an unknown utterance as one of a set of reference words is adapted to generate a feature signal set for each utterance of every reference word. At least one template signal is produced for each reference word which template signal is representative of a group of feature signal sets. Responsive to a feature signal set formed from the unknown utterance and each reference word template signal, a signal representative of the similarity between the unknown utterance and the template signal is generated. A plurality of similarity signals for each reference word is selected and a signal corresponding to the average of said selected similarity signals is formed. The average similarity signals are compared to identify the unknown utterance as the most similar reference word.

Abstract: A speech analyzer for recognizing an unknown utterance as one of a set of reference words is adapted to generate a feature signal set for each utterance of every reference word. At least one template signal is produced for each reference word which template signal is representative of a group of feature signal sets. Responsive to a feature signal set formed from the unknown utterance and each reference word template signal, a signal representative of the similarity between the unknown utterance and the template signal is generated. A plurality of similarity signals for each reference word is selected and a signal corresponding to the average of said selected similarity signals is formed. The average similarity signals are compared to identify the unknown utterance as the most similar reference word.

Abstract: A speech controlled dialing circuit identifies input utterances which may be a command word (mode select), repertory word (dialing name or number), or nonrecognized ("Other"). Responsive to the identification of each occurring input utterance, a set of predetermined templates are selected to identify the next occuring utterance. A programmed microprocessor system is described to implement the main controller function.

Abstract: An arrangement for endpoint detection improves speech recognition accuracy and lowers rejection rates by developing an ordered list of endpoint candidates. A triple thresholding technique defines energy signal pulses. The energy pulses are combined according to predetermined criteria to form the endpoint candidates.

Abstract: A speech recognizer includes a plurality of stored constrained hidden Markov model reference templates and a set of stored signals representative of prescribed acoustic features of the said plurality of reference patterns. The Markov model template includes a set of N state signals. The number of states is preselected to be independent of the reference pattern acoustic features and preferably substantially smaller than the number of acoustic feature frames of the reference patterns. An input utterance is analyzed to form a sequence of said prescribed feature signals representative of the utterance. The utterance representative prescribed feature signal sequence is combined with the N state constrained hidden Markov model template signals to form a signal representative of the probability of the utterance being each reference pattern. The input speech pattern is identified as one of the reference patterns responsive to the probability representative signals.