An acoustic signal processing circuit extracts input speech pattern data and subsidiary feature data from an input speech signal. The input speech pattern data comprise frequency spectra, whereas the subsidiary feature data comprise phoneme and acoustic features. These data are then stored in a data buffer memory. The similarity measures between the input speech pattern data stored in the data buffer memory and reference speech pattern data stored in a dictionary memory are computed by a similarity computation circuit. When the largest similarity measure exceeds a first threshold value and when the difference between the largest similarity measure and the second largest measure exceeds a second threshold value, category data of the reference pattern which gives the largest similarity measure is produced by a control circuit to correspond to an input speech. When recognition cannot be performed, the categories of the reference speech patterns which respectively give the largest to mth similarity measures are respectively compared with the subsidiary feature data. In this manner, subsidiary feature recognition of the input voice is performed by a subsidiary feature recognition section.

Speech recognition system

A method for recognizing spoken words of a speech includes extracting feature vectors from a speech signal which corresponds to a spoken phrase, and segmenting and classifying the successive extracted feature vectors into syllable oriented word subunits by means of a stored supply of word subunits to form a set of hypotheses. The set of hypotheses is used to generate, by three dimensional time dynamic comparision, a set of word hypotheses by comparing the segmented and classified word subunits with standard pronunciations and pronunciation variants of a plurality of words stored in a reference pattern vocabulary. The generated set of word hypotheses are then subjected to syntactic analysis to determine the spoken phrase.

Method for speaker-adaptive speech recognition

A speech recognition system prompts a user to provide a first utterance, which is recorded. Speech recognition is performed on the first user utterance to yield a recognition result. The user is prompted to provide a second user utterance, which is recorded, processed and compared to the first utterance to detect at least one acoustic difference. The acoustic model used by the speech recognition engine is modified as a function of the acoustic difference.

Ein solches Verfahren soll sowohl zur Erkennung einzelner Worter als auch von kontinuierlich gesprochener Sprache geeignet sein. Es soll sich durch Robustheit der Wortmustererkennung bei fehlerhafter Silbensegmentierung und bei variabler Aussprache, z.B. bei einem Verschlucken von Silben, auszeichnen. Auserdem mus es eine schnelle Adaption des Systems an einen neuen Sprecher und eine prinzipiell beliebige Generierung und Erweiterung des Wortschatzes aus geschriebenem Text ohne ein explizites Systemtraining durch Vorsprechen ermoglichen. Eine echtzeitnahe Erkennung von Wortern und Wortfolgen soll auch bei sehr umfangreichen Wortschatzen moglich sein. Bekannte Verfahren zur Spracherkennung benotigen ein sehr aufwendiges Trainingsverfahren. Auserdem wird bei kontinuierlich gesprochener Sprache und groserem Vokabular bereits bei mittleren Vokabulargrosen eine unuberschaubare Hypothesenflut erzeugt. Erfindungsgemas werden die extrahierten Mekrmalsvektoren mit Hilfe eines Wortuntereinheiten-Inventars in einen Hypothesensatz aus silbenorientierten Wortuntereinheiten segmentiert und klassifiziert. Aus dem Hypothesensatz wird, durch ein dreidimensionaler zeitdynamischer Vergleich mit Aussprachevarianten aus einem Referenzmuster-Wortschatz, eir Worthypothesensatz erzeugt, der einer syntaktischen Analyse zur Ermittlung der gesprochenen Phrase unterworfen wird.

Verfahren zur sprecheradaptiven Erkennung von Sprache

The word-based approach to automatic speech recognition is one which has received attention from many researchers and has been exploited in various practical applications. A typical recognition system has a set of stored reference patterns, one or more for each word in the vocabulary to be recognised. These reference patterns are formed from training utterances supplied before a recognition session begins, either by the intended user of the system or, for a speaker-independent system, by a representative set of speakers. When the system is used for recognition, each new input utterance is compared with the stored patterns and is recognised as the word (or sequence of words) for which the minimal value of a distance (dissimilarity) measure, or equivalently the maximal likelihood, is obtained. The comparison of the input with the reference patterns is typically accomplished by an algorithm incorporating dynamic programming, which finds the optimal alignment of input and reference patterns and the corresponding distance or likelihood. This approach to recognition, in its basic form, retains the same reference patterns unchanged throughout the recognition of any sequence of input utterances. Thus the recognition system has no capability of learning from the new utterances presented during a recognition session. If a recognition system can be made to adapt its reference patterns during its operation, to incorporate information from the recognised utterances, then this may be expected to allow progressive improvement of the modelling of the words (as pronounced by the current speaker), and hence enhancement of the accuracy of recognition provided that the adaptation of incorrect words' reference patterns in cases of misrecognition can be prevented or kept to a sufficiently low level. By adaptation, speaker-specific initial reference patterns can be made more reliably representative of the speaker's typical pronunciations, by the use of data from additional utterances of the words; and speaker-independent reference patterns can be made speaker-specific through the incorporation of information from utterances by the speaker currently using the recognition system. Adaptation can also permit the dynamic adjustment of reference patterns to track any gradual drift, or systematic difference from one occasion to another, in the speaker's voice or pronunciations or in the level and characteristics of background noise.

Adaptation of reference patterns in word-based speech recognition

In this speaker-dependent recognition system, recognition is based on syllable template matching and each syllable has several templates. In the initial training for each speaker, 590 templates for 111 syllables are made, each including various contextual variations. The authors studied a learning method in which the syllable templates are automatically optimized. It is judged whether or not an input syllable should be learned according to the recent recognition condition. If it should be learned, the input syllable pattern replaces the template that contributes the least to recognition in the templates segmented from the same context and in the same syllable category. Automatic learning was evaluated on recognition of speech data obtained by reading Japanese sentences at a rate of about 4 to 5 syllables per second. The results over eight speakers showed an average syllable recognition accuracy of 71.0% without and 82.5% with automatic learning. Further, by increasing the maximum number of templates to 1024, it rose to 84.8%.

Voice-activated word processor with automatic learning for dynamic optimization of syllable-templates

A voice-activated word processor has been realized which provides text input by articulated, phrase-by-phrase Japanese speech, using an automatic learning algorithm to improve syllable recognition accuracy. A speaker-dependent, continuous-phrase speech recognizer incorporated into the word processor is capable of real-time recognition using 111 monosyllables as the basic recognition units. The recognizer is trained by new user to construct 590 reference syllable templates based on uttering a set of words, necessary for syllable template-matching process. Syllable templates are continually updated during use of the word processor through automatic learning. The automatic learning algorithm replaces low-accuracy syllable templates with new patterns extracted from the input speech. The replacement is sensitive to syllable context and is carried out based on recent and longer term history of the recognition accuracy of each syllable. Such learning information is derived from comparison between syllable recognition results and a user-confirmed character string of input phrase as correct, and is accumulated on a phrase-by-phrase basis. The automatic learning algorithm was tested in experiments using Japanese sentences read with pauses between phrases at approximately 4 to 5 syllables per second. The results, using eight speakers, show average syllable recognition accuracy of 82.5% with automatic learning, compared to 71.0% achieved without the learning. Further, the recognition accuracy is increased to 86.5% when the maximum number of syllable templates is increased to 2, 048; both template replacement and template addition are carried out until the maximum number of templates is reached.

https://www.jstage.jst.go.jp/article/ast1980/10/3/10_3_133/_pdf

Torn Ueda

Papers

Voice-activated word processor with automatic learning for dynamic optimization of syllable-templates

Voice-activated word processor with automatic learning for dynamic optimization of syllable-templates