Abstract: Aspects of the present disclosure enable humanly-specified relationships to contribute to a mapping that enables compression of the output structure of a machine-learned model. An exponential model such as a maximum entropy model can leverage a machine-learned embedding and the mapping to produce a classification output. In such fashion, the feature discovery capabilities of machine-learned models (e.g., deep networks) can be synergistically combined with relationships developed based on human understanding of the structural nature of the problem to be solved, thereby enabling compression of model output structures without significant loss of accuracy. These compressed models provide improved applicability to “on device” or other resource-constrained scenarios.

Abstract: Aspects of the present disclosure enable humanly-specified relationships to contribute to a mapping that enables compression of the output structure of a machine-learned model. An exponential model such as a maximum entropy model can leverage a machine-learned embedding and the mapping to produce a classification output. In such fashion, the feature discovery capabilities of machine-learned models (e.g., deep networks) can be synergistically combined with relationships developed based on human understanding of the structural nature of the problem to be solved, thereby enabling compression of model output structures without significant loss of accuracy. These compressed models provide improved applicability to “on device” or other resource-constrained scenarios.

Abstract: Aspects of the present disclosure enable humanly-specified relationships to contribute to a mapping that enables compression of the output structure of a machine-learned model. An exponential model such as a maximum entropy model can leverage a machine-learned embedding and the mapping to produce a classification output. In such fashion, the feature discovery capabilities of machine-learned models (e.g., deep networks) can be synergistically combined with relationships developed based on human understanding of the structural nature of the problem to be solved, thereby enabling compression of model output structures without significant loss of accuracy. These compressed models provide improved applicability to “on device” or other resource-constrained scenarios.

Abstract: Aspects of the present disclosure enable humanly-specified relationships to contribute to a mapping that enables compression of the output structure of a machine-learned model. An exponential model such as a maximum entropy model can leverage a machine-learned embedding and the mapping to produce a classification output. In such fashion, the feature discovery capabilities of machine-learned models (e.g., deep networks) can be synergistically combined with relationships developed based on human understanding of the structural nature of the problem to be solved, thereby enabling compression of model output structures without significant loss of accuracy. These compressed models provide improved applicability to “on device” or other resource-constrained scenarios.

Abstract: Recognition techniques may include the following. On a first processing entity, a first recognition process is performed on a first element, where the first recognition process includes: in a first state machine having M (M>1) states, determining a first best path cost in at least a subset of the M states for at least part of the first element. On a second processing entity, a second recognition process is performed on a second element, where the second recognition process includes: in a second state machine having N (N>1) states, determining a second best path cost in at least a subset of the N states for at least part of the second element. At least one of the following is done: (i) passing the first best path cost to the second state machine, or (ii) passing the second best path cost to the first state machine. The foregoing techniques may include one or more of the following features, either alone or in combination.

Abstract: Respective word frequencies may be determined from a corpus of utterance-to-text-string mappings that contain associations between audio utterances and a respective text string transcription of each audio utterance. Respective compressed word frequencies may be obtained based on the respective word frequencies such that the distribution of the respective compressed word frequencies has a lower variance than the distribution of the respective word frequencies. Sample utterance-to-text-string mappings may be selected from the corpus of utterance-to-text-string mappings based on the compressed word frequencies. An automatic speech recognition (ASR) system may be trained with the sample utterance-to-text-string mappings.

Abstract: A speech recognition process may perform the following operations: performing a preliminary recognition process on first audio to identify candidates for the first audio; generating first templates corresponding to the first audio, where each first template includes a number of elements; selecting second templates corresponding to the candidates, where the second templates represent second audio, and where each second template includes elements that correspond to the elements in the first templates; comparing the first templates to the second templates, where comparing comprises includes similarity metrics between the first templates and corresponding second templates; applying weights to the similarity metrics to produce weighted similarity metrics, where the weights are associated with corresponding second templates; and using the weighted similarity metrics to determine whether the first audio corresponds to the second audio.

Abstract: Respective word frequencies may be determined from a corpus of utterance-to-text-string mappings that contain associations between audio utterances and a respective text string transcription of each audio utterance. Respective compressed word frequencies may be obtained based on the respective word frequencies such that the distribution of the respective compressed word frequencies has a lower variance than the distribution of the respective word frequencies. Sample utterance-to-text-string mappings may be selected from the corpus of utterance-to-text-string mappings based on the compressed word frequencies. An automatic speech recognition (ASR) system may be trained with the sample utterance-to-text-string mappings.

Abstract: A computerized method is provided for generating pronunciations for words and storing the pronunciations in a pronunciation dictionary. The method includes graphing sets of initial pronunciations; thereafter in an ASR subsystem determining a highest-scoring set of initial pronunciations; generating sets of alternate pronunciations, wherein each set of alternate pronunciations includes the highest-scoring set of initial pronunciations with a lowest-probability phone of the highest-scoring initial pronunciation substituted with a unique-substitute phone; graphing the sets of alternate pronunciations; determining in the ASR subsystem a highest-scoring set of alternate pronunciations; and adding to a pronunciation dictionary the highest-scoring set of alternate pronunciations.

Abstract: A computerized pronunciation system is provided for generating pronunciations for words and storing the pronunciations in a pronunciation dictionary. The system includes a word list including at least one word; transcribed acoustic data including at least one waveform for the word and transcribed text associated with the waveform; a pronunciation-learning module configured to accept as input the word list and the transcribed acoustic data, the pronunciation-learning module including: sets of initial pronunciations of the word, a scoring module configured score pronunciations and to generate phone probabilities, and a set of alternate pronunciations of the word, wherein the set of alternate pronunciations include a highest-scoring set of initial pronunciations with a highest-scoring substitute phone substituted for a lowest-probability phone; and a pronunciation dictionary configured to receive the highest-scoring set of initial pronunciations and the set of alternate pronunciations.

Abstract: A method and apparatus for generating a noise-reduced feature vector representing human speech are provided. Speech data representing an input speech waveform are first input and filtered. Spectral energies of the filtered speech data are determined, and a noise reduction process is then performed. In the noise reduction process, a spectral magnitude is computed for a frequency index of multiple frequency indexes. A noise magnitude estimate is then determined for the frequency index by updating a histogram of spectral magnitude, and then determining the noise magnitude estimate as a predetermined percentile of the histogram. A signal-to-noise ratio is then determined for the frequency index. A scale factor is computed for the frequency index, as a function of the signal-to-noise ratio and the noise magnitude estimate. The noise magnitude estimate is then scaled by the scale factor.

Abstract: Pronunciation quality is automatically evaluated for an utterance of speech based on one or more pronunciation scores. One type of pronunciation score is based on duration of acoustic units. Examples of acoustic units include phones and syllables. Another type of pronunciation score is based on a posterior probability that a piece of input speech corresponds to a certain model such as an HMM, given the piece of input speech. Speech may be segmented into phones and syllables for evaluation with respect to the models. The utterance of speech may be an arbitrary utterance made up of a sequence of words which had not been encountered before. Pronunciation scores are converted into grades as would be assigned by human graders. Pronunciation quality may be evaluated in a client-server language instruction environment.

Abstract: Pronunciation quality is automatically evaluated for an utterance of speech based on one or more pronunciation scores. One type of pronunciation score is based on duration of acoustic units. Examples of acoustic units include phones and syllables. Another type of pronunciation score is based on a posterior probability that a piece of input speech corresponds to a certain model, such as a hidden Markov model, given the piece of input speech. Speech may be segmented into phones and syllable for evaluation with respect to the models. The utterance of speech may be an arbitrary utterance made up of a sequence of words which had not been encountered before. Pronunciation scores are converted into grades as would be assigned by human graders. Pronunciation quality may be evaluated in a client-server language instruction environment.

Abstract: Adverse effects of type mismatch between acoustic input devices used during testing and during training in machine-based recognition of the source of acoustic phenomena are minimized. A normalizing model is matched to a source model based, or dependent, upon an acoustic input device whose transfer characteristics color acoustic characteristics of a source as represented in the source model. An application of the present invention is to speaker recognition, i.e., recognition of the identity of a speaker by the speaker's voice.

Abstract: In a method for determining likelihood of appearance of keywords in a spoken utterance as part of a keyword spotting system of a speech recognizer, a new scoring technique is provided wherein a confidence score is computed as a probability of observing the keyword in a sequence of words given the observations. The corresponding confidence scores are the probability of the keyword appearing in any word sequence given the observations. In a specific embodiment, the technique involves hypothesizing a keyword whenever it appears in any of the "N-Best" word lists with a confidence score that is computed by summing the likelihoods for all hypotheses that contain the keyword, normalized by dividing by the sum of all hypothesis likelihoods in the "N-best" list.

Abstract: A dual operational exercise resistance device that is usable in the home or fitness center by attachment to a supporting structure. The supporting structure can be a freestanding platform base, a wall, or a door. The device can be utilized to replace the weight stack in a universal gym or any exercise machine. The device can be utilized to provide resistance from more than one point of attachment for standard exercise grips. In addition the resistance provided by the device is easily adjustable, providing the appropriate resistance for the particular muscle group being exercised. The device comprises a base that is configured for attachment to a support or exercise machine, a lever arm member having a peripheral edge being pivotally attached to the base having a pulley on end, two pulleys on opposite ends of the base being pivotally mounted on a horizontal axis within vertically orientated mounted brackets and a pulley being pivotally attached to an intermediate point on the base.

Abstract: A portable exercise device that is usable in the home by attachment to supporting structure. The device comprises a pair of bases that are configured for attachment to a support, a member having a peripheral edge being pivotally attached to one of the bases and a bar being pivotally attached to the other base. One end of the bar is attached to the base to which the member is attached by a resistance mechanism. To the other end of the bar is attached a strap that engages a portion of the peripheral edge of the member and extends therefrom for attachment to a grip. This structure reduces the resistance force produced by the resistance mechanism as the exerciser approaches full contraction of the muscle group being exercised to ensure a completely full contraction is made. The device also compensates for the use of a non-linear resistance mechanism. In addition the resistance provided by the device is easily adjustable, providing the appropriate resistance for the particular muscle group being exercised.

Abstract: An automatic speech recognition methodology, wherein words are modeled as probabilistic networks of allophones, collects nodes in the probabilistic network into equivalence classes when those nodes have the same allophonic choices governed by the same phonological rules. The allophonic choices allow for representation of dialectic pronunciation variations between different speakers. Training data is shared among nodes in an equivalence class so that accurate pronunciation probabilities may be determined even for words for which there is only a limited amount of training data. A method is used to determine probabilities for each of a multitude of pronunciation models for each word in the vocabulary, based on automatic extraction of linguistic knowledge from sets of phonological rules, in order to robustly and accurately model dialectal variation.

Abstract: An automatic speech recognition methodology takes advantage of linguistic constraints wherein words are modeled as probabilistic networks of phonetic segments (herein phones), and each phone is represented as a context-independent hidden Markov phone model mixed with a number of context-dependent phone models. Recognition is based on use of methods to design phonological rule sets based on measures of coverage and overgeneration of pronunciations which achieves high coverage of pronunciations with compact representations. Further, a method estimates probabilities of the different possible pronunciations of words. A further method models cross-word coarticulatory effects. In a specific embodiment of the system, a specific method determines the single most-likely pronunciation of words. In further specific embodiments of the system, methods generate speaker-dependent pronunciation networks.

Abstract: A method is disclosed for use in preprocessing noisy speech to minimize likelihood of error in estimation for use in a recognizer. The computationally-feasible technique, herein called Minimum-Mean-Log-Spectral-Distance (MMLSD) estimation using mixture models and Marlov models, comprises the steps of calculating for each vector of speech in the presence of noise corresponding to a single time frame, an estimate of clean speech, where the basic assumptions of the method of the estimator are that the probability distribution of clean speech can be modeled by a mixture of components each representing a different speech class assuming different frequency channels are uncorrelated within each class and that noise at different frequency channels is uncorrelated.