Noise robust speech classifier ensemble
AssigneeMicrosoft Corporation
InventorsKunal Mukerjee, Kazuhito Koishida, Shankar Regunathan
Life of this patent
Estimating the term…
Abstract
Embodiments for implementing a speech recognition system that includes a speech classifier ensemble are disclosed. In accordance with one embodiment, the speech recognition system includes a classifier ensemble to convert feature vectors that represent a speech vector into log probability sets. The classifier ensemble includes a plurality of classifiers. The speech recognition system includes a decoder ensemble to transform the log probability sets into output symbol sequences. The speech recognition system further includes a query component to retrieve one or more speech utterances from a speech database using the output symbol sequences.
Claims20
A method, comprising:
- converting a plurality of feature vectors that represents a speech utterance into a plurality of log probability sets, the converting using a classifier ensemble including a plurality of classifiers;
- transforming the plurality of log probability sets into a plurality of output symbol sequences;
- combining the plurality of output symbol sequences, using an iterative a priori probability calculation algorithm, into a fusion output symbol sequence;
- and
- retrieving one or more speech utterances from a speech database using the plurality of output symbol sequences.
A computer-readable storage device storing computer-executable instructions that, when executed, cause one or more processors to perform acts comprising:
- converting a plurality of feature vectors into a plurality of log probability sets using a classifier ensemble that includes a plurality of classifiers, the plurality of feature vectors representing a speech utterance;
- transforming the plurality of log probability sets into a plurality of output symbol sequences using a decoder ensemble that includes a plurality of decoders;
- combining the plurality of output symbol sequences into a fusion output symbol sequence using a block fusion algorithm, the block fusion algorithm to use convergent iterative a-priori probability calculations to produce the fusion output sequence;
- and
- retrieving one or more stored speech utterances from a speech database based on at least one of the plurality of output symbol sequences or the fusion output symbol sequence.
A method comprising:
- segmenting an input audio signal into a plurality of frames using a frame capture component;
- applying a first Discrete Cosine Transform (DCT) (Type II) to the noise-suppressed Mel filter coefficients to each frame of the plaurity of frames along a frequency direction to derive corresponding output coefficients;
- populating a matrix with the output coefficients;
- applying a second DCT (Type II) to the output coefficients in the matrix along a time direction to generated a de-correlated matrix for the plurality of frames;
- applying a feature selection to the de-correlated matrix to extract feature vectors;
- converting the feature vectors into log probability sets using a classifier ensemble that includes a plurality of classifiers;
- transforming the log probability sets into an output symbol sequences using a decoder ensemble that includes a plurality of decoders;
- combining the output symbol sequences into a fusion output symbol sequence using a block fusion algorithm, the block fusion algorithm to use convergent iterative a-priori probability calculations to produce the fusion output sequence;
- and
- retrieving one or more stored speech utterances from a speech database based on at least one of the plurality of output symbol sequences or the fusion output symbol sequence.
Dependencies are inferred from the claim text.