Method and apparatus for the recognition of individually spoken words

Claim 1

A system for the recognition of separately spoken words, comprising: an acousto-electric transducer for receiving the spoken words to be recognized and producing an electrical signal therefrom; a controllable amplifier connected to an output of said acousto-electric transducer to amplify the electrical signal from said transducer; a high-pass filter connected to an output of said controllable amplifier to filter the electrical signal; a signal analysis path connected at an output of said high-pass filter to receive the electrical signal, said signal analysis path including: a low pass filter for filtering the electrical signal; a first one-bit analog - to - digital converter connected at an output of said low pass filter for converting the electrical signal into a digital signal; a digital processing means for sign correlation of the signal in sections from said first analog-to-digital converter to produce digital data words; a microprocessor having a digital input connected to receive said digital data words from said digital processing means; a level calculation path connected to the output of said high pass filter, said level calculation path including: an envelope detection circuit operable to calculate the level of the output signal of said transducer and produce an envelope signal; a second one-bit analog-to-digital converter having an input connected to an output of said envelope detection circuit and having and output connected to a digital input of said microprocessor for producing a digital envelope signal from said envelope signal; and means for supplying a master clock signal to synchronize said microprocessor and said first and second analog-to-digital converters; said microprocessor processing signals from said first and second analog to-digital converters to recognize the spoken words, said microprocessor: receiving the beginning and end of the digital envelope signal corresponding to a separately spoken word t determine the word length, dividing the signals corresponding to a separately spoken word into regions having feature vectors, averaging the feature vectors to produce resulting vectors, and comparing the resulting vectors to lexicon entries for election of an appropriate lexicon word.

Claim 2

A system for the recognition of separately spoken words, comprising: an acousto-electric transducer for receiving the spoken words to be recognized and producing an electrical signal therefrom; a controllable amplifier connected to an output of said acousto-electric transducer to amplify the electrical signal from said transducer; a high-pass filter connected to an output of said controllable amplifier to filter the electrical signal; a signal analysis path connected at an output of said high-pass filter to receive the electrical signal, said signal analysis path including: a low pass filter for filtering the electrical signal; a first one-bit analog - to - digital converter connected at an output of said low pass filter for converting the electrical signal into a digital signal; a digital processing means for sign correlation of the signal in sections from said first analog-to-digital converter to produce digital data words; a microprocessor having a digital input connected to receive said digital data words from said digital processing means; a level calculation path connected to the output of said high pass filter, said level calculation path including: an envelope detection circuit operable to calculate the level of the output signal of said transducer and produce an envelope signal; a second one-bit analog-to-digital converter having an input connected to an output of said envelope detection circuit and having and output connected to a digital input of said microprocessor for producing a digital envelope signal from said envelope signal; and means for supplying a master clock signal to synchronize said microprocessor and said first and second analog-to-digital converters; said microprocessor processing signals from said first and second analog to-digital converters to recognize the spoken words, said microprocessor: receiving the beginning and end of the digital envelope signal corresponding to a separately spoken word t determine the word length, dividing the signals corresponding to a separately spoken word into regions having feature vectors, averaging the feature vectors to produce resulting vectors, and comparing the resulting vectors to lexicon entries for election of an appropriate lexicon word.

Claim 15

A method for operating a separately spoken word recognition system having first and second converters, a scanner and a lexicon, comprising the steps of: executing a sign correlation with a vector composed of vector components according to the function: ##EQU3## for feature calculation wherein AKF (n) are vector components, s(t) is the output signal of the first converter means, t is the scan time of the scannermeans, T is the length of the time interval to be observed and N+1, 2, 3 . . . N whereby N≦T represents the shift parameter of a correlation; representing a full separately spoken word by a sequence of sign correlation vectors, said vectors being features of successive time intervals of the length of the time interval to be observed within the word; considering at time tp=m×T for m=1, 2, 3 . . . a level signal output of the second converter means; perceiving the beginning of separately spoken words at the time t PA when said level signal has a logic level of "1" and subsequently assumes a logic level of "1" more frequently in a predetermined time span than a logic level of "0"; perceiving the end of separately spoken words at a time t PE for a last level signal having a logic level of "1" and followed by a predetermined number of level signals of logic level of "0"; determining a length of separately spoken words as a difference between the end time t PE and the beginning time t PA ; dividing the separately spoken words into regions having an equal number of feature vectors so that each of said regions has ##EQU4## by the total number of vectors in each region, where N B is the number of regions and so that the beginning of each respective region is at a distance of: ##EQU5## from the beginning of said words; averaging said feature vectors in each region to produce resulting vectors; depositing said resulting vectors and said length and an appropriate code as an identifier in the lexicon during a learning phase for each word to be recognized; eliminating lexicon words having lengths smaller by a factor of F1 and larger by a factor of F2 than said length of the word to be recognized in a recognition phase; identifying the distance of the word to be recognized from remaining ones of said lexicon words by summation of the differences of said respective resulting vectors of the word to be recognized and the remaining lexicon words; selecting as fitting one of said lexicon words having the least distance from the word to be recognized; rejecting said selected lexicon word if said least distance is greater than a prescribed value; outputting said appropriate code selected as an identifier for said fitting lexicon word; and triggering a predetermined action by said output coding.

Claim 16

A method for operating a separately spoken word recognition system having first and second converters, a scanner and a lexicon, comprising the steps of: executing a sign correlation with a vector composed of vector components according to the function: ##EQU3## for feature calculation wherein AKF (n) are vector components, s(t) is the output signal of the first converter means, t is the scan time of the scannermeans, T is the length of the time interval to be observed and N+1, 2, 3 . . . N whereby N≦T represents the shift parameter of a correlation; representing a full separately spoken word by a sequence of sign correlation vectors, said vectors being features of successive time intervals of the length of the time interval to be observed within the word; considering at time tp=m×T for m=1, 2, 3 . . . a level signal output of the second converter means; perceiving the beginning of separately spoken words at the time t PA when said level signal has a logic level of "1" and subsequently assumes a logic level of "1" more frequently in a predetermined time span than a logic level of "0"; perceiving the end of separately spoken words at a time t PE for a last level signal having a logic level of "1" and followed by a predetermined number of level signals of logic level of "0"; determining a length of separately spoken words as a difference between the end time t PE and the beginning time t PA ; dividing the separately spoken words into regions having an equal number of feature vectors so that each of said regions has ##EQU4## by the total number of vectors in each region, where N B is the number of regions and so that the beginning of each respective region is at a distance of: ##EQU5## from the beginning of said words; averaging said feature vectors in each region to produce resulting vectors; depositing said resulting vectors and said length and an appropriate code as an identifier in the lexicon during a learning phase for each word to be recognized; eliminating lexicon words having lengths smaller by a factor of F1 and larger by a factor of F2 than said length of the word to be recognized in a recognition phase; identifying the distance of the word to be recognized from remaining ones of said lexicon words by summation of the differences of said respective resulting vectors of the word to be recognized and the remaining lexicon words; selecting as fitting one of said lexicon words having the least distance from the word to be recognized; rejecting said selected lexicon word if said least distance is greater than a prescribed value; outputting said appropriate code selected as an identifier for said fitting lexicon word; and triggering a predetermined action by said output coding.