Abstract: A decoder comprises a feature extraction circuit for calculating one or more feature vectors; an acoustic model circuit coupled to receive one or more feature vectors from said feature extraction circuit and assign one or more likelihood values to the one or more feature vectors; a memory for storing states of transition of the decoder; and a search circuit for receiving an input from said acoustic model circuit corresponding to the one or more likelihood values based upon the one or more feature vectors, and for choosing states of transition from the memory based on the input from said acoustic model.
Type:
Grant
Filed:
May 31, 2017
Date of Patent:
August 31, 2021
Assignee:
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
Inventors:
Michael R. Price, James R. Glass, Anantha P. Chandrakasan
Abstract: A method for real-time data-pattern analysis. The method includes receiving and queuing at least one data-pattern analysis request by a data-pattern analysis unit controller. At least one data stream portion is also received and stored by the data-pattern analysis unit controller, each data stream portion corresponding to a received data-pattern analysis request. Next, a received data-pattern analysis request is selected by the data-pattern analysis unit controller along with a corresponding data stream portion. A data-pattern analysis is performed based on the selected data-pattern analysis request and the corresponding data stream portion, wherein the data-pattern analysis is performed by one of a plurality of data-pattern analysis units.
Abstract: A hardware acoustic scoring unit for a speech recognition system and a method of operation thereof are provided. Rather than scoring all senones in an acoustic model used for the speech recognition system, acoustic scoring logic first scores a set of ciphones based on acoustic features for one frame of sampled speech. The acoustic scoring logic then scores senones associated with the N highest scored ciphones. In one embodiment, the number (N) is three. While the acoustic scoring logic scores the senones associated with the N highest scored ciphones, high score ciphone identification logic operates in parallel with the acoustic scoring unit to identify one or more additional ciphones that have scores greater than a threshold. Once the acoustic scoring unit finishes scoring the senones for the N highest scored ciphones, the acoustic scoring unit then scores senones associated with the one or more additional ciphones.
Abstract: A hardware implemented backend search stage, or engine, for a speech recognition system is provided. In one embodiment, the backend search engine includes a number of pipelined stages including a fetch stage, an updating stage which may be a Viterbi stage, a transition and prune stage, and a language model stage. Each active triphone of each active word is represented by a corresponding triphone model. By being pipelined, the stages of the backend search engine are enabled to simultaneously process different triphone models, thereby providing high-rate backend searching for the speech recognition system. In one embodiment, caches may be used to cache frequently and/or recently accessed triphone information utilized by the fetch stage, frequently and/or recently accessed triphone-to-senone mappings utilized by the updating stage, or both.
Abstract: Hidden Markov Model (HMM) parameters are updated using update equations based on growth transformation optimization of a minimum classification error objective function. Using the list of N-best competitor word sequences obtained by decoding the training data with the current-iteration HMM parameters, the current HMM parameters are updated iteratively. The updating procedure involves using weights for each competitor word sequence that can take any positive real value. The updating procedure is further extended to the case where a decoded lattice of competitors is used. In this case, updating the model parameters relies on determining the probability for a state at a time point based on the word that spans the time point instead of the entire word sequence. This word-bound span of time is shorter than the duration of the entire word sequence and thus reduces the computing time.
Abstract: The invention uses the ModelGrower program to generate possible candidates from an original or aggregated model. An isomorphic reduction program operates on the candidates to identify and exclude isomorphic models. A Markov model evaluation and optimization program operates on the remaining non-isomorphic candidates. The candidates are optimized and the ones that most closely conform to the data are kept. The best optimized candidate of one stage becomes the starting candidate for the next stage where ModelGrower and the other programs operate on the optimized candidate to generate a new optimized candidate. The invention repeats the steps of growing, excluding isomorphs, evaluating and optimizing until such repetitions yield no significantly better results.
Abstract: The method of recognizing speech in an acoustic signal comprises developing acoustic stochastic models of voice units in the form of a set of states of an acoustic signal and using the acoustic models for recognition by a comparison of the signal with predetermined acoustic models obtained via a prior learning process. While developing the acoustic models, the voice units are modeled by means of a first portion of the states independent of adjacent voice units and by means of a second portion of the states dependent on adjacent voice units. The second portion of states dependent on adjacent voice units shares common parameters with a plurality of units sharing same phonemes.
Abstract: A method and apparatus for indexing one or more audio signals using a speech to text engine and a phoneme detection engine, and generating a combined lattice comprising a text part and a phoneme part. A word to be searched is searched for in the text part, and if not found, or is found with low certainty is divided into phonemes and searched for in the phoneme parts of the lattice.
Abstract: Methods are given for improving discriminative training of hidden Markov models for continuous speech recognition. For a mixture component of a hidden Markov model state, a gradient adjustment is calculated of the standard deviation of the mixture component. If the calculated gradient adjustment is greater than a first threshold amount, an adjustment is performed of the standard deviation of the mixture component using the first threshold. If the calculated gradient adjustment is less than a second threshold amount, an adjustment is performed of the standard deviation of the mixture component using the second threshold. Otherwise, an adjustment is performed of the standard deviation of the mixture component using the calculated gradient adjustment.
Abstract: A speech recognition apparatus using a probability model that employs a mixed distribution, the apparatus formed by a standard pattern storage means for storing a standard pattern; a recognition means for outputting recognition results corresponding to an input speech by using the standard pattern; a standard pattern generating means for inputting learning speech and generating the standard pattern; and a standard pattern adjustment means, provided between the standard pattern generating means and the standard pattern storage means, for adjusting the number of element distributions of the mixed distribution of the standard pattern.