Back to results
US 4,790,017 AGranted
Speech processing feature generation arrangement
AssigneeAmerican Telephone and Telegraph Company, AT&T Bell Laboratories
InventorHarvard S. Hinton
Life of this patent
Estimating the term…
Abstract
In a speech signal recognition system, signals are data-reduced by eliminating highly-transient intervals: each successive time frame is temporarily identified as an anchor frame and compared to m preceding and m succeeding frames, whereby defined uniform time intervals of 2m+1 frames having a prescribed similarity are retained, and highly-transient 2m+1 intervals are discarded.
Claims6
A speech pattern recognition arrangement comprising:
- means for partitioning an input speed pattern into successive time frame portions i=1, 2, . . . , i, . . . , I;
- means responsive to each speech pattern time frame portion i for generating a set of first signals U(i) representative of the acoustic features of the time frame portion;
- means responsive to the time frame sequence of first signals of the speech pattern for successively forming a plurality of speech pattern time intervals k, each time interval comprising a set of 2m+1 successive time frames of said speech pattern, the first signals of the time interval 2m+1 successive time frames having a prescribed degree of similarity;
- means responsive to the first signals of the 2m+1 time frames of each speech pattern interval k for generating a set of second signals I(k,1), I(k,2), . . . , I(k,p+3) representative of the acoustic features of said interval;
- means for storing a plurality of reference pattern templates, each template comprising an interval sequence of second signals corresponding to a reference pattern;
- and
- means jointly responsive to said speech pattern time interval sequence of second signals and said reference pattern time interval sequence of second signals for identifying said speech pattern as one of said reference patterns;
- said speech pattern time interval forming means comprising:
- means for sequentially designating each successive frame i is an anchor time frame for the set of time frames i-m to i+m;
- means responsive to the first signals of the anchor time frame i and the first signals of the m time frames i-1, i-2, . . . , i-m preceding said anchor time frame for comparing the first signals of said anchor time frame i with the first signals of each of the m time frames i-1, i-2, . . . , i-m preceding said anchor time frame i to generate m signals S(i-1), S(i-2), . . . , S(i-m) representative of the similarity between said anchor time frame i first signals and the first signals of each of the m preceding time frames i-1, i-2, . . . , i-m;
- means responsive to the similarity signal S(i) of the designated anchor time frame being less than or equal to each of the similarity signals S(i-m), S(i-m+1), . . . , S(i+m) for the preceding m time frames and the succeeding m time frames for generating a signal identifying the time interval of said 2m+1 successive time frames i-m, i-m+1, . . . , i+m as a second signal time interval;
- and
- means responsive to said identification signal for assigning said 2m+1 successive time frames from time frame i-m to i+m to a speech pattern time interval k.
10 more elements
A method for recognizing a speech pattern comprising the steps of:
- partitioning an input speech pattern into successive time frame portions i=1, 2, . . . , i, . . . , I;
- generating a set of first signals U(i) representative of the acoustic features of the time frame portion responsive to each speech pattern time frame portion i;
- successively forming a plurality of speech pattern time intervals k responsive to the time frame sequence of first signals of the speech pattern, each time interval comprising a set of 2m+1 successive time frames of said speech pattern and the first signals of the time interval 2m+1 successive time frames having a prescribed degree of similarity;
- generating a set of signal signals I(k,1), I(k,2), . . . , I(k,p+3) representative of the acoustic features of said interval responsive to the first signals of the 2m+1 time frames of each speech pattern interval k;
- storing a plurality of reference pattern templates, each template comprising an interval sequence of second signals corresponding to a reference pattern;
- and
- identifying said speech pattern as one of said reference patterns jointly responsive to said speech pattern time interval sequence of second signals and said reference pattern time interval sequence of second signals;
- said speech pattern time interval forming step including:
- sequentially designating each successive time frame i as an anchor time frame for the set of time frames i-m to i+m;
- comparing the first signals of said anchor time frame i with the first signals of each of the m time frames i-1, i-2, . . . , i-m preceeding said anchor time frame i to generate m signals S(i-1), S(i-2,), . . . , S(i-m) representative of the similarity between said anchor time frame i first signals and the first signals of each of the m preceding time frames i-1, i-2, , . . . , i-m response to the first signals of anchor time frame i and the first signals of the m time frames i-1, i-2, . . . , i-m preceding said anchor time frame;
- generating a signal identifying said time interval of said 2m+1 successive time frames i-m, i-m+1, . . . , i-m as a second signal time interval responsive to the similarity signal S(i) of the designated anchor time frame being less than or equal to each of the similarity signals S(i-m), S(i-m+1), . . . , S(i+m) for the preceding m time frames;
- and
- assigning said 2m+1 successive time frames from time to frame i-m to i+m to a speech pattern time interval responsive to said identification signal.
10 more elements
Dependencies are inferred from the claim text.