Technique for modifying reference vector quantized speech feature signals
AssigneeAT&T Corp.
InventorsAaron E. Rosenberg, Frank K.-P. Soong
Life of this patent
Estimating the term…
Abstract
A signal processing arrangement uses a codebook of first vector quantized speech feature signals formed responsive to a large collection of speech feature signals. The codebook is altered by combining the first speech feature signals of the codebook with second speech feature signals generated responsive to later input speech patterns during normal speech processing. A speaker recognition template can be updated in this fashion to take account of change which may occur in the voice and speaking characteristics of a known speaker.
Claims21
An arrangement for modifying a codebook of vector quantized speech feature signals for speaker verification or adaptation to the voice of a particular speaker comprising:
- means for storing a set of vector quantized feature signals q(i), i=1, 2, . . . , K, where K is the number of vector quantized feature signals in said set, and a set of occupancy signals n(i) each corresponding to one of said vector quantized feature signal q(i);
- means for receiving an input pattern;
- means for analyzing said input pattern to generate a set of input feature vector signals v(t), t=1,2, . . . , M, where M is the number of vector-quantized feature signals in an input pattern;
- means responsive to said set of input feature signals and said set of vector quantized feature signals for classifying each input feature signal v(t) as one of said set if vector quantized feature signals q(i);
- means responsive to the classification of each input feature signal as one of said vector quantized feature signals for generating an occupancy signal m(i) corresponding to the count of input feature signals classified as each one of said vector quantized feature signals q(i);
- and
- means for modifying said set if vector quantized feature vector signals, q(i), to produce a modified set of vector quantized feature vector signals, q'(i), based on said classified input feature vector signals, said input feature occupancy signals, said vector quantized feature signals, and said occupancy signals of said vector quantized feature signals
In a signal processing arrangement having a stored codebook of a set of vector quantized speech feature signals q(i) i=1, 2, . . . , K, where K is the number of vector quantized feature signals in said set, and occupancy signals n(i) each corresponding to one of said vector quantized feature signals, a method for modifying the codebook of vector quantized feature signals for speaker verification or adaptation to the voice of a particular speaker comprising the steps of:
- receiving an input pattern;
- analyzing said input pattern to generate a set of input feature vector signals v(t), t=1, 2, . . . , M, where M is the number of vector-quantized feature signals in an input pattern;
- classifying each input feature vector signal v(t) as one of said set of vector quantized feature signals q(i) responsive to said input feature signals and said set of vector quantized feature signals;
- generating an occupancy signal m(i) corresponding to the count input feature vector signals classified as respective ones of said vector quantized feature signals q(i) responsive to the classifying step;
- and
- modifying said set of vector quantized feature vector signals, q(i), based on said classified input feature vector signals, said input feature occupancy signals, said vector quantized feature signals and said occupancy signals of said vector quantized feature signals.
In a signal processing arrangement having a stored codebook of a set of vector quantized feature signals q(i) i=1, 2, . . . K, where K is the number of vector quantized feature signals in said set, and occupancy signals n(i) each corresponding to one of said vector quantized feature signals, a method for modifying the codebook of vector quantized feature signals, the method comprising the steps of:
- receiving an input pattern;
- analyzing said input pattern to generate a set of input feature vector signals v(t), t=1, 2, . . . , M, where M is the number of vector-quantized feature signals in an input pattern;
- classifying each input feature vector signal v(t) as one of said set of vector quantized feature signals q(i) responsive to said input feature signals and said set of vector quantized feature signals;
- generating an occupancy signal m(i) corresponding to the count input feature vector signals classified as respective ones of said vector quantized feature signals q(i) responsive to the classifying step;
- and
- modifying said set of vector quantized feature vector signals, q(i), based on said classified input feature vector signals, said input feature occupancy signals, said vector quantized feature signals and said occupancy signals of said vector quantized feature signals.
Dependencies are inferred from the claim text.