AT&T Labs Research — Leading Invention, Driving Innovation

The present disclosure relates to systems, methods, and computer-readable media for generating a lexicon for use with speech recognition. The method includes receiving symbolic input as labeled speech data, overgenerating potential pronunciations based on the symbolic input, identifying potential pronunciations in a speech recognition context, and storing the identified potential pronunciations in a lexicon. Overgenerating potential pronunciations can include establishing a set of conversion rules for short sequences of letters, converting portions of the symbolic input into a number of possible lexical pronunciation variants based on the set of conversion rules, modeling the possible lexical pronunciation variants in one of a weighted network and a list of phoneme lists, and iteratively retraining the set of conversion rules based on improved pronunciations. Symbolic input can include multiple examples of a same spoken word. Speech data can be labeled explicitly or implicitly and can include words as text and recorded audio.

Systems, computer-implemented methods, and tangible computer-readable media for generating a pronunciation model. The method includes identifying a generic model of speech composed of phonemes, identifying a family of interchangeable phonemic alternatives for a phoneme in the generic model of speech, labeling the family of interchangeable phonemic alternatives as referring to the same phoneme, and generating a pronunciation model which substitutes each family for each respective phoneme. In one aspect, the generic model of speech is a vocal tract length normalized acoustic model. Interchangeable phonemic alternatives can represent a same phoneme for different dialectal classes. An interchangeable phonemic alternative can include a string of phonemes.

Disclosed are systems, methods and computer readable media for applying a multi-state barge-in acoustic model in a spoken dialogue system comprising the steps of (1) presenting a prompt to a user from the spoken dialog system. (2) receiving an audio speech input from the user during the presentation of the prompt, (3) accumulating the audio speech input from the user, (4) applying a non-speech component having at least two one-state Hidden Markov Models (HMMs) to the audio speech input from the user, (5) applying a speech component having at least five three-state HMMs to the audio speech input from the user, in which each of the five three-state HMMs represents a different phonetic category, (6) determining whether the audio speech input is a barge-in-speech input from the user, and (7) if the audio speech input is determined to be the barge-in-speech input from the user, terminating the presentation of the prompt.

Systems and methods are provided for recognizing speech in a spoken dialogue system. The method includes receiving input speech having a pre-vocalic consonant or a post-vocalic consonant, generating at least one output lattice that calculates a first score by comparing the input speech to a training model to provide a result and distinguishing between the pre-vocalic consonant and the post-vocalic consonant in the input speech. A second score is calculated by measuring a similarity between the pre-vocalic consonant or the post vocalic consonant in the input speech and the first score. At least one category is determined for the pre-vocalic match or mismatch or the post-vocalic match or mismatch by using the second score and the results of the an automated speech recognition (ASR) system are refined by using the at least one category for the pre-vocalic match or mismatch or the post-vocalic match or mismatch.

Disclosed are systems, methods and computer readable media for training acoustic models for an automatic speech recognition systems (ASR) system. The method includes receiving a speech signal, defining at least one syllable boundary position in the received speech signal, based on the at least one syllable boundary position, generating for each consonant in a consonant phoneme inventory a pre-vocalic position label and a post-vocalic position label to expand the consonant phoneme inventory, reformulating a lexicon to reflect an expanded consonant phoneme inventory, and training a language model for an automated speech recognition (ASR) system based on the reformulated lexicon.

Disclosed are systems and methods for training a barge-in-model for speech processing in a spoken dialogue system comprising the steps of (1) receiving an input having at least one speech segment and at least one non-speech segment, (2) establishing a restriction of recognizing only speech states during speech segments of the input and non-speech states during non-speech segments of the input, (2) generating a hypothesis lattice by allowing any sequence of speech Hidden Markov Models (HMMs) and non-speech HMMs, (4) generating a reference lattice by only allowing speech HMMs for at least one speech segment and non-speech HMMs for at least one non-speech segment, wherein different iterations of training generates at least one different reference lattice and at least one reference transcription, and (5) employing the generated reference lattice as the barge-in-model for speech processing.

A method and apparatus for performing speech recognition are provided. A Vocal Tract Length Normalized acoustic model for a speaker is generated from training data. Speech recognition is performed on a first recognition input to determine a first best hypothesis. A first Vocal Tract Length Normalization factor is estimated based on the first best hypothesis. Speech recognition is performed on a second recognition input using the Vocal Tract Length Normalized acoustic model to determine an other best hypothesis. An other Vocal Tract Length Normalization factor is estimated based on the other best hypothesis and at least one previous best hypothesis.