Distributed speech recognition with back-end voice activity detection apparatus and method
AssigneeMotorola, Inc.
InventorTenkasi Ramabadran
Life of this patent
Estimating the term…
Abstract
In a distributed voice recognition system, a back-end pattern matching unit 27 can be informed of voice activity detection information as developed through use of a back-end voice activity detector 25. Although no specific voice activity detection information is developed or forwarded by the front-end of the system, precursor information as developed at the back-end can be used by the voice activity detector to nevertheless ascertain with relative accuracy the presence or absence of voice in a given set of corresponding voice recognition features as developed by the front-end of the system.
Claims20
A method for facilitating recognition of speech, comprising:
- at a back-end platform:
- receiving a plurality of speech recognition features;
- receiving a plurality of cepstral features as part of receiving the plurality of speech recognition features;
- processing at least some of the plurality of cepstral features to provide at least an approximation of precursor information that yielded the plurality of cepstral features;
- processing the at least an approximation of precursor information to detect portions that likely correspond to speech activity and provide a corresponding identification of speech portions;
- processing the plurality of speech recognition features as a function, at least in part, of the identification of speech portions to facilitate recognition of speech content represented by at least some of the speech recognition features.
An apparatus for use in facilitating distributed speech recognition, the apparatus comprising:
- at a back-end platform:
- a voice activity detector having an input operably coupled to receive cepstral features as derived from at least an approximation of precursor information and having at least a first output that provides a speech activity detection signal identifying the at least an approximation of the precursor information that likely corresponds to speech;
- a segmentation unit having an input operably coupled to the first output of the voice activity detector and having an output that provides a speech detection signal identifying portions of the at least an approximation of the precursor information that likely correspond to speech;
- and
- a pattern matching unit having inputs operably coupled to receive the cepstral features and the speech detection signal and having an output that provides recognized speech as corresponds to the cepstral features.
A method comprising:
- at a back-end platform:
- receiving over a wireless channel a signal including a bit stream containing information derived from a plurality of input samples;
- decoding the bit stream to recover thirteen Mel frequency cepstral coefficients and a logarithmic value of energy as corresponds to each frame of input samples;
- using an inverse discrete cosine transformation and exponentiation to convert the Mel frequency cepstral coefficients into a spectral representation thereof;
- using the spectral representation to detect portions thereof that likely include speech activity;
- using the Mel frequency cepstral coefficients, the logarithmic value for each frame of the input samples, and information identifying portions that likely include speech to facilitate a pattern matching process to recognize speech content in the input.
Dependencies are inferred from the claim text.