Speech recognition using phoneme waveform parameters

A system for speech recognition based on direct measurement of phoneme waveform structural features, comprising:

• (a) means for providing an input analog speech signal;
• (b) means for pre-processing said input analog signal to convert it to a normalized-amplitude sampled digital signal;
• (c) means for buffering said pre-processed signal;
• (d) means for fetching and analyzing a short interval, called a frame, of said buffered signal, said analyzing by Mainwave-Ripple Analysis comprising
• (d1) locating any signal waveform present in the frame, and
• (d2) measuring the waveform fine-structure parameters within very short time intervals, including amplitude valley and zero-crossing counts, called local values, and thereby deriving over the frame a distribution array of local values, whereby the waveform is identified as one of four or more basic phoneme waveforms, called protowaves;
• (e) means for further analyzing said input waveform in the frame to measure main-structure parameters, including
• (e1) mainwave parameters including flyback-pulse amplitude valley and pulse width, and
• (e2) ripplewave parameters including normalized ripple amplitude and normalized ripple rate;
• (f) means for combining said measures to form a set of waveform parameters called a single frame analysis array which uniquely defines the waveform structure;
• (g) means for repeatedly fetching and analyzing frames and outputting single-frame analysis arrays per steps (d) through (f), thus outputting a multi-frame sequence of analysis arrays;
• (h) mode switch means for selectively connecting the output sequence of analysis arrays to either a training reference generator or to a recognizer;
• (i) training reference generator means for generating and storing known-phoneme single frame reference arrays by combining a most-probable single-frame analysis array with its corresponding typed-in phoneme symbol;
• (j) single frame recognizer means for comparing an unknown single frame analysis array with a pre-stored set of known-phoneme single frame reference arrays, the best match yielding a single frame decision output phoneme;
• (k) multi-frame recognition means for storing and examining the sequence of single-frame decision phoneme outputs to find a significant pattern which provides a multi-frame decision output;
• (l) context-correction recognition means for examining the multi-frame decision output phoneme sequence for a word recognition;
• (m) function recognition means for examining the word recognized and outputting any functional meaning;
• and
• (n) master control means for controlling and accessing all system means and providing selectable throughput with peripheral devices and human operator.

A method of speech recognition based on direct measurement of phoneme waveform structural features, comprising the steps of:

• (a) providing an input analog speech signal;
• (b) pre-processing said input analog signal to convert it to a normalized-amplitude sampled digital signal;
• (c) buffering said pre-processed signal;
• (d) fetching and analyzing a short interval, called a frame, of said buffered signal, said analyzing by Mainwave-Ripple Analysis comprising
• (d1) locating any signal waveform present in the frame, and
• (d2) measuring the waveform fine-structure parameters within very short time intervals, including amplitude valley and zero-crossing counts, called local values, and thereby deriving over the frame a distribution array of local values, whereby the waveform is identified as one of four or more basic phoneme waveforms, called protowaves;
• (e) further analyzing said waveform in the frame to measure main-structure parameters, including
• (e1) mainwave parameters including flyback-pulse amplitude valley and pulse width, and
• (e2) ripplewave parameters including normalized ripple amplitude and normalized ripple rate;
• (f) combining said measures to form a set of waveform parameters called a single frame analysis array which uniquely defines the waveform structure;
• (g) repeatedly fetching and analyzing frames and outputting single-frame analysis arrays per steps (d) through (f), thus outputting a multi-frame sequence of analysis arrays;
• (h) mode-switch selecting the output sequence of analysis arrays to either training or recognition procedures;
• (i) training the system by generating and storing
• (i1) singular-reference arrays which are known-phoneme single-frame reference arrays formed by combining a most-probable single-frame analysis array with its corresponding typed-in phoneme symbol,
• (i2) universal-reference arrays which are a plurality of sets of phoneme reference arrays formed by interpolation of parameter values between singular arrays of a basso and soprano to cover the male-female analysis array parameter range,
• (i3) personal-reference arrays which are generated by prompted calibrating commands, and
• (i4) phonode-reference arrays which are generated by broadening parameter value ranges to include confusion group values;
• (j) recognizing phonemes using
• (j1) single-frame recognition which compares an input unknown single-frame analysis array with a pre-stored set of known-phoneme single-frame reference arrays, the best match yielding a single-frame decision output phoneme, and
• (j2) multi-frame recognition by storing and examining the sequence of single-frame decision phoneme outputs to find a significant pattern which provides a multi-frame decision phoneme output;
• (k) context-correction recognizing by examining the multi-frame decision output phoneme sequence for a word recognition;
• (l) function recognizing by examining the word recognized and outputting any functional meaning;
• and
• (m) master controlling all processing including accessing all processing procedures and providing selectable throughput with peripheral devices and human operator.

In a method of speech recognition, a method of waveform analysis comprising the steps of:

• (a) providing an input analog speech signal;
• (b) pre-processing said input analog signal to convert it to a normalized-amplitude sampled digital signal;
• (c) buffering said pre-processed signal;
• (d) fetching and analyzing a short interval, called a frame, of said buffered signal, said analyzing by Mainwave-Ripple Analysis comprising
• (d1) locating any signal waveform present in the frame, and
• (d2) measuring the waveform fine-structure parameters within very short time intervals, including amplitude valley and zero-crossing counts, called local values, and thereby deriving over the frame a waveform analysis array comprising a distribution of local values, including frame high and low local values, whereby the waveform is identified as one of four or more basic phoneme waveforms, called protowaves.

In a method of speech recognition using Mainwave-Ripple Analysis, a method of training for recognition comprising the steps:

• (a) generating and storing
• (a1) singular-reference arrays which are known-phoneme single frame reference arrays formed by combining a most-probable single-frame waveform analysis array with its corresponding phoneme sysmbol,
• (a2) universal-reference arrays which are a plurality of sets of phoneme reference arrays formed by interpolation of parameter values between singular arrays of a basso and soprano to cover the male-female analysis array parameter range, and
• (a3) personal-reference arrays which are generated from prompted calibrating commands;
• and
• (b) generating an extended set of personal reference arrays by adding related-value universal reference arrays.

In a method of speech recognition using Mainwave-Ripple Analysis, a method of coarse recognition using a phoneme/phonode approximation comprising the steps of:

• (a) grouping phonemes having similar sounds and similar waveform analysis arrays into confusion groups;
• (b) selecting from each confusion group one phoneme, called a phonode, which is an optimal representative for the group, for print or phonemic talkback;
• and
• (c) generating and storing a set of phonode-reference arrays which have broadened ranges of phoneme waveform parameters to cover the confusion group, whereby recognition processing can be limited to recognizing the phoneme's confusion group and outputting the corresponding phonode.

A computer program storage device readable by a computer system, tangibly embodying a program of instructions executable by the computer system to perform a speech recognition method, said method comprising the steps of:

• (a) pre-processing an input speech signal to output an amplitude-normalized digital speech signal;
• (b) sequentially locating each signal waveform in a short time interval called a frame;
• (c) analyzing each said waveform by Mainwave-Ripple Analysis comprising locating and measuring structural features to derive a waveform anaysis array comprising measures of the distribution of amplitude and zero-crossings in very short intervals called local intervals, and high and low values of the local interval values for the frame called minimax values;
• (d) training for speech recognition by combining a most-probable waveform analysis array with its known phoneme symbol to generate reference arrays, and
• (e) recognizing unknown phoneme waveforms by comparing each unknown input speech signal waveform analysis array with prestored reference arrays, whereby a best match effects phoneme recognition.

A system for speech recognition which converts an input analog speech signal to corresponding phonemic text, said system comprising:

• (a) means for pre-processing an input speech signal to output an amplitude-normalized digital speech signal;
• (b) means for sequentially locating each signal waveform in a short time interval called a frame;
• (c) means for analyzing said waveform by Mainwave-Ripple Analysis comprising locating and measuring structural features to derive a waveform analysis array comprising measures of the distribution of amplitude and zero-crossings in very short intervals called local intervals, and high and low values of the local interval values for the frame called minimax values;
• (d) means for selecting one of several operating mode means, including training and recognition;
• (e) training means for combining a most-probable single-frame waveform analysis array with a known phoneme symbol to generate single-frame reference arrays, and
• (f) means for recognizing unknown phoneme waveforms by comparing each unknown input speech signal waveform analysis array with prestored single-frame reference arrays, whereby a best-match effects phoneme recognition.

A method for speech recognition which converts an input analog speech signal to corresponding phonemic text, said method comprising the steps of:

• (a) pre-processing an input speech signal to output an amplitude-normalized digital speech signal;
• (b) sequentially locating each signal waveform in a short time interval called a frame;
• (c) analyzing said waveform by Mainwave-Ripple Analysis comprising locating and measuring structural features to derive a waveform analysis array comprising measures of the distribution of amplitude and zero-crossings in very short intervals called local intervals, and high and low values of the local interval values for the frame called minimax values;
• (d) selecting one of several operating modes including training and recognition;
• (e) training by combining a most-probable single-frame waveform analysis array with its known phoneme symbol to generate single-frame reference arrays, and
• (f) recognizing unknown phoneme waveforms by comparing each unknown input speech signal waveform analysis array with prestored single-frame reference arrays, whereby a best-match effects phoneme recognition.