Google Inc.

Nuance Communications, Inc.

Mitchel Weintraub

Francoise Beaufays

Brian Strope

Ehsan Variani

Ananth Sankar

Ananda Theertha Suresh

Shaun Eric Williams

Vincent Olivier Vanhoucke

Patrick An Phu Nguyen

Georg Heigold

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.

Exponential modeling with deep learning features

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&#x3e;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&#x3e;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.

Parallel recognition

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.

Training an automatic speech recognition system using compressed word frequencies

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.

Speech recognition process

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.

Method for learning linguistically valid word pronunciations from acoustic data

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.

Method and system for learning linguistically valid word pronunciations from acoustic data

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. The scaled noise magnitude estimate is subtracted from the spectral magnitudes of the filtered speech data, to produce cleaned speech data, based on which a feature vector is generated.

Growing community of inventors

Cupertino, CA, United States of America

Mitchel Weintraub