Continuous speech recognition apparatus

• US 4,087,630 A
• Filed: 05/12/1977
• Issued: 05/02/1978
• Est. Priority Date: 05/12/1977
• Status: Expired due to Term

• Current Assignee: Centigram Communications Corporation (TE Connectivity Limited)
• Original Assignee: Centigram Corporation
• Inventors: Madden, John D., Postas, L. John, Browning, Iben, Chapman, Robert G. Jr., Berney, Carl L., Glaser, George
• Primary Examiner(s): Claffy, Kathleen H.
• Assistant Examiner(s): Kemeny, E. S.
• Application Number:
  US05/796,191
• Time in Patent Office: 355 Days
• Field of Search: 179/1 SD, 340/146.3 WD
• US Class Current: 704/236
• CPC Class Codes: B60R 16/0373 Voice control in general G10L
• CPC Class Codes: G10L 15/00 Speech recognition G10L17/00 takes precedence

Abstract

An apparatus and method wherein speech or other signals are sampled during a time slice of approximately 1/30 second and spectrum analysis is performed on the samples, producing measures of amplitude in several frequency bands with each frequency band being characterized by a binary digit indicating the presence or absence of significant amplitude. The binary digits are collectively referred to as a sonogram. Sonograms for several time slices are then concatenated, randomized and decoded using an n-tuple technique to produce a pattern corresponding to the current speech signal. This pattern is learned by superimposing it on an existing vocabulary entry and is subsequently recognized if it is sufficiently similar to one vocabulary entry and different from all others.

Claims

1. Speech recognition apparatus capable of operating in either a learn mode or a recognize mode comprising:

means for converting audible speech into an electrical signal;

3. Speech recognition apparatus as recited in claim 1 wherein said electrical signal is in the form of an analog signal and said signal processing means includes an analog-to-digital converter which periodically samples said analog signal and develops digital signals corresponding to the amplitude of each analog signal sample, and buffer means for storing each said digital signal in a time slice.

5. Speech recognition apparatus as recited in claim 3 wherein said means for receiving and repetitively outputting said digital signals at predetermined rates includes a variable output rate shift register having recirculating circuit means coupling its output to its input.

6. Speech recognition apparatus as recited in claim 3 wherein said spectrum analyzing means further includes averaging circuit means for normalizing said component frequency signals.

7. Speech recognition apparatus as recited in claim 5 wherein said averaging circuit means includes a rectifier means for rectifying the output of said filter means, an integrating circuit for integrating the rectified signals, means for generating time base signals, and means for dividing the integrated signals by the time base signals to develop said frequency component signals.

8. Speech recognition apparatus as recited in claim 3 wherein said means for converting said series of frequency component signals includes an analog-to-binary digital converter which develops a series of binary numbers, each of which is representative of the amplitude of one of said frequency component signals, and means for ranking said binary numbers, discarding those binary numbers which fall below a predetermined threshold and generating a binary word including one bit for each said frequency component, the bits of said binary word corresponding to the X largest numbers, where X is a predetermined integer, being set.

9. Speech recognition apparatus as recited in claim 7 wherein said means for decoding, said means for storing, and said means for comparing comprise a preprogrammed microprocessor.

11. Signal recognition apparatus as recited in claim 9 wherein said preprocessing means includes an analog-to-digital converter which periodically samples said analog signal to develop said digital signals, and buffer means for storing each said digit signal in said segment.

13. Signal recognition apparatus as recited in claim 11 wherein said spectrum analyzing means further includes averaging circuit means for normalizing said frequency component signals.

15. Signal recognition apparatus as recited in claim 13 wherein said means for concatenating, and said means for comparing comprise a preprogrammed microprocessor.

17. A speech recognition method as recited in claim 15 wherein said analyzing step includes sampling each said time slice of signal a plurality of times to develop a series of digital signals representative of the amplitude of each sample;

outputing said series of digital signals a predetermined number of times at different rates;

converting each read out series of digital signals to a time-shifted analog signal; and

filtering each said time-shifted signal through a signal band pass filter to develop a series of frequency component signals from which said sonograms are developed.

19. A speech recognition method as recited in claim 15 wherein said predetermined criteria is the requirement that the count of matched bits of the highest count exceed a predetermined minimum value and differ from the next highest count by a predetermined number of counts.

2. signal processing means for sampling a time slice of said electrical signal and for developing a plurality of digital signals representative thereof;

spectrum analyzing means for receiving said plurality of digital signals and for developing a series of frequency component signals each of which is indicative of the amplitude of a particular frequency component in said time slice;

means for converting said series of frequency component signals into a series of binary digits respectively indicating the presence or absence of significant amplitude in each said frequency component signal;

means for pseudo-randomly selecting various ones of said binary digits from a plurality of concatenated series of said digits and for combining the selected binary digits into groups of n data bits, where n is an integer;

means for decoding each of said groups of n bits to develop a corresponding binary word of length 2ⁿ ;

means for storing said binary word when said apparatus is operated in the learn mode; and

means for comparing words developed from subsequently input speech to each of said stored words, and for developing an output signal when a predetermined correlation is found to exist between the compared input word and a particular stored word, such output signal indicating that said subsequently input speech has been recognized.

4. Speech recognition apparatus as recited in claim 2 wherein said spectrum analyzing means includes means for receiving the digital signals stored in said buffer means and corresponding to a first time slice, and for repetitively outputting such signals at predetermined different rates during the time that digital signals from a second time slice are being loaded into said buffer means;

digital-to-analog means for converting each of said time-shifted digital signals to corresponding analog signals; and

bandpass filter means for filtering out a particular frequency component from each of said analog signals to develop said series of frequency component signals.

10. Signal recognition apparatus for recognizing data contained in an analog electrical signal comprising:

preprocessing means for converting a segment of the analog signal into a predetermined number of digital signals, each representing the amplitude of a portion of said segment;

spectrum analyzing means for receiving said digital signals and for determining the frequency content of the signal segment represented by said digital signals and for developing a series of frequency component signals each corresponding to the relative magnitude of a particular frequency component of said segment;

means for converting said series of frequency component signals into an input sonogram comprised of a number of data bits corresponding to said frequency component signals and indicating the presence or absence of component frequencies of significant amplitude;

means for concatenating a plurality of said sonograms and for using an n-tuple pattern generating technique to convert the concatenated sonograms into an n-tuple input binary word;

means for comparing said input binary word with each of a predetermined number of previously stored binary words and for developing an output signal when a predetermined number of bits of said input binary word correspond with a like number of bits of one of said stored binary words; and

means responsive to said output signal for indicating recognition of the data.

12. Signal recognition apparatus as recited in claim 10 wherein said spectrum analyzing means includes means for receiving the digital signals stored in said buffer means and corresponding to said first segment, and for repetitively outputting said signals at predetermined different rates during the time that digital signals from a second segment are being loaded into said buffer means;

digital-to-analog means for converting the time shifted digital signals to corresponding analog signals; and

band pass filter means for filtering out a particular frequency component from each of the time shifted analog signals to develop said series of frequency component signals.

14. Signal recognition apparatus as recited in claim 12 wherein said means for converting said series of frequency component signals includes an analog-to-binary digital converter which is representative of the amplitude of one of said frequency component signals, and means for ranking said binary numbers, discarding those binary numbers which fall below a predetermined threshold and generating a binary word including one bit for each frequency component, the bits of said binary word corresponding to the X largest numbers, where X is a predetermined integer, being set.

16. A speech recognition method comprising;

converting a voice signal into a corresponding analog electrical signal;

separating said analog signal into time slices of signal;

analyzing each said time slice for frequency content to develop a sonogram comprised of a series of digital characters each of which corresponds to the relative magnitude of a particular frequency component of the time slice;

accumulating a plurality of said sonograms;

using an n-tuple technique to develop a word pattern from said accumulated sonograms;

comparing said word pattern to each of a plurality of previously stored word patterns, each time counting the number of bits in the compared patterns; and

generating a recognition indication when the number of matched bits satisfies a predetermined criteria.

18. A speech recognition method as recited in claim 16 wherein the step of utilizing said n-tuple technique includes pseudo-randomly selecting various ones of said digital characters from said accumulated sonograms, combining the selected digital characters into groups of n data bits, where n is an integer, and decoding each of said groups of said n bits to develop a corresponding binary word of length 2ⁿ.