System for recognizing speech

Claim 3

The system of claim 1, wherein said output of the second filtering means indicates at least two of the three features, namely, the onset, rise and fall of any significant tones of the input speech signal over time.

Claim 4

(Amended) The system of claim 2, wherein said output of the second filtering means also indicates frequency of any significant tones of the input speech signal.

Claim 5

The system of claim 1, further comprising a neural network responsive to said second two dimensional representation of said input speech signal for identifying phonemes, groups of phonemes or suprasegmentals in said input signal.

Claim 6

The system of claim 4, wherein said network includes at least one pair of phoneme- or suprasegmental-related formation and deformation layers to provide formation and deformation maps, said deformation layer performing a local-averaging function on said formation map to provide the deformation map.

Claim 7

The system of claim 4, wherein said first filtering means is a filter bank having coefficients that are functions of selected dimensions of a representative biological cochlear.

Claim 8

The system of claim 6, said representative biological cochlear having a basilar membrane with a breadth, wherein said coefficients are functions of the breadth of the basilar membrane.

Claim 9

The system of claim 1, wherein said first filtering means includes a bank of M filters having different frequency pass bands, M being a positive integer greater than 1, each filter providing an output, said filters being indexed from 1 to M by pass band frequencies of the filters along a frequency axis and the outputs of the filters at N different times being indexed from 1 to N along a time axis, N being a positive integer greater than 1, forming an M by N array of filter outputs, said second filtering means including: means for superposing at least one p by q array of filter coefficients upon a portion of said M by N array of filtered outputs, p and q being positive integers, so that each coefficient corresponds to a filter output in said M by N array; and means for multiplying each coefficient by the corresponding filter output to obtain products and summing said products to obtain a first processed value for said superposition upon said one portion.

Claim 10

The system of claim 8, wherein said superposing means includes p(q-1) delay elements for delaying p outputs of the M filter outputs and pq gain or amplifying elements for amplifying the outputs.

Claim 11

The system of claim 8, wherein said superposition means superposes said p by q array upon each of mn different portions of the M by N array, m and n being positive integers, where the portions form an m by n array along the frequency and time axes, and wherein, for each superposition upon a portion, said multiplying means multiplies each coefficient by the corresponding filter output to obtain products and summing said products to obtain a first processed value, thereby obtaining from the mn superpositions and multiplications an m by n array of first processed values along said two axes.

Claim 12

The system of claim 10, wherein said portions are selected such that spacing between adjacent portions is the same for the m by n array of portions.

Claim 13

The system of claim 10, wherein said superposing means superposes each of 2 p by q arrays of coefficients upon one or more portions of said M by N array successively so that each coefficient corresponds to a filter output in said M by N array, and wherein, for each p by q array and for each superposition, said multiplying means multiplies each coefficient of such p by q array by the corresponding filter output to obtain products and sums said products for such p by q array to obtain one processed value for such superposition, and to obtain one or more such processed values for each p by q array.

Claim 14

The system of claim 12, wherein the coefficients of each p by q array are such that the processed values obtained for such array indicate at least two of the following four elementary tonotopic features: the onset, frequency, rise and fall of any significant tones in the input signal.

Claim 15

The system of claim 13, wherein said superposing means superposes each of 4 pairs of p by q arrays of coefficients upon one or more portions of said M by N array successively so that each coefficient corresponds to a filter output in said M by N array, and wherein the coefficients of each pair of p by q arrays are such that the processed values obtained for such pairs of arrays indicate respectively one of the following four elementary tonotopic features: the onset, frequency, rise and fall of any significant tones in the input signal.

Claim 16

The system of claim 8, wherein the coefficients of said p by q array are such that the processed value obtained indicates one of the following elementary tonotopic features: the onset, frequency, rise or fall of any significant tones in the input signal.

Claim 17

The system of claim 8, wherein said superposition means superposes said p by q array upon different portions of the M by N array so that each of the M by N filter outputs is superposed upon at least one coefficient, and wherein said multiplying means multiplies each of the M by N filter outputs by at least one coefficient.

Claim 18

The system of claim 8, wherein said coefficients are symmetric or anti-symmetric functions of frequency and time.

Claim 19

The system of claim 8, wherein said coefficients are symmetric or anti-symmetric functions of frequency and time modulated by a Gaussian function.

Claim 20

The system of claim 18, wherein said coefficients are Gabor functions of frequency and time.

Claim 21

The system of claim 1, said second filtering means including at least two filters at different resolutions for filtering the output of the first filtering means to provide speech context information for the recognition of phonemes.

Claim 22

The system of claim 20, further comprising a neural network responsive to said two dimensional representation of said input speech signal for identifying phonemes, groups of phonemes or suprasegmentals in said input signal.

Claim 23

The system of claim 21, wherein said network includes at least one pair of phoneme- or suprasegmental-related formation and deformation layers to provide formation and deformation maps, said deformation layer performing a local-averaging function on said formation map to provide the deformation map.

Claim 24

A filter bank for improving signal to noise ratio in processing an input signal comprising: a plurality of M filters arranged in parallel, M being a positive integer, each filter having a first stage and a second stage; wherein the first stage of each filter includes: (a) first delay means for delaying the input signal, and (b) means for subtracting from the input signal or a signal derived therefrom the delayed input signal or a signal derived therefrom and adding thereto feedback signals from at least some of the second stages of the M filters or signals derived therefrom to derive an output signal; wherein the second stage of each filter provides an output signal and includes: (c) first means for adding the output signals of the first stages of at least some of the filters or signals derived therefrom to obtain a first sum signal; (d) second means for delaying said first sum signal and supplying said delayed first sum signal or a signal derived therefrom to the first stages of at least some of the filters; (e) second means for adding the sum signal and the delayed sum signal or signals derived therefrom to obtain a second sum signal; (f) third means for delaying the output signal of the second stage and supplying said delayed output signal of the second stage or a signal derived therefrom to the first stages of at least some of the filters; and (g) means for adding to the second sum signal or a signal derived therefrom the delayed output signal of the second stage or a signal derived therefrom to derive the output signal of the second stage.

Claim 25

A filter bank for improving signal to noise ratio in processing an input signal comprising: a plurality of M filters arranged in parallel, M being a positive integer, each filter having a first stage and a second stage; wherein the first stage of each filter includes: (a) first delay means for delaying the input signal, and (b) means for subtracting from the input signal or a signal derived therefrom the delayed input signal or a signal derived therefrom and adding thereto feedback signals from at least some of the second stages of the M filters or signals derived therefrom to derive an output signal; wherein the second stage of each filter provides an output signal and includes: (c) first means for adding the output signals of the first stages of at least some of the filters or signals derived therefrom to obtain a first sum signal; (d) second means for delaying said first sum signal and supplying said delayed first sum signal or a signal derived therefrom to the first stages of at least some of the filters; (e) second means for adding the sum signal and the delayed sum signal or signals derived therefrom to obtain a second sum signal; (f) third means for delaying the output signal of the second stage and supplying said delayed output signal of the second stage or a signal derived therefrom to the first stages of at least some of the filters; and (g) means for adding to the second sum signal or a signal derived therefrom the delayed output signal of the second stage or a signal derived therefrom to derive the output signal of the second stage.

Claim 26

The filter bank of claim 23, wherein M is greater than or equal to 10, and wherein the subtracting and adding means (b) adds feedback signals from at least half of the second stages of the M filters or signals derived therefrom in deriving the output signal, and wherein the first adding means adds the output signals of the first stages of at least half of the filters or signals derived therefrom to obtain said first sum signal.

Claim 27

The filter bank of claim 23, wherein said first, second and third delaying means introduce substantially the same delay.

Claim 28

The filter bank of claim 23, further comprising means for sampling the input signal, the first and second sum signals, and the output signals of the two stages, at a substantially constant sampling rate and at sampling interval T to obtain samples of such signals, wherein at (n-1)th and at nth time instants when such signals are sampled, n being a positive integer, an (n-1)th sample and an nth sample of each of such signals are obtained.

Claim 29

The filter bank of claim 26, i and j being integers in the range 1 to M, wherein the nth sample z i (n) of said output signal of said first stage of the ith filter is given by: z.sup.i (n)=a.sub.1.sup.i * z.sub.1.sup.i (n-1)+c.sup.i * Σ.sub.j d.sup.j Θ.sup.i.sub.j * z.sub.1.sup.j (n-1)+a.sub.2.sup.i * z.sub.2.sup.i (n-1)+b.sup.i * (u(n)-u(n-1)) where the (n-1) th sample z 1 i (n-1) and the nth sample z 1 i (n) of the first sum signal for the ith filter are respectively given by: z.sub.1.sup.i (n-1)=Σ.sub.j e.sup.i.sub.j * Z.sup.j (n-1); z.sub.1.sup.i (n)=Σ.sub.j e.sup.i.sub.j * Z.sup.j (n); where the nth sample z 2 i (n) of the output signal of the second stage of the ith filter is given by: z.sub.2.sup.i (n)=z.sub.2.sup.i (n-1)+(T/2) * (z.sub.1.sup.i (n)+z.sub.1.sup.i (n-1)); where z 2 i (n-1) is the (n-1)th sample of the output signal of the second stage of the ith filter; where the (n-1) th sample and nth sample of the input signal are respectively u(n-1) and u(n); where ##EQU7## Σ.sub.k [δ.sup.i.sub.k +c.sup.i Θ.sup.i.sub.k d.sup.k ]e.sup.k.sub.j =δ.sup.i.sub.j ; and where a 1 i , a 2 i , b i , c i , d i are constants.

Claim 30

The filter bank of claim 27, said bank suitable for use in recognizing speech features in speech input signals, wherein when a i 1 is in the range between 0 and about 1, the coefficients a 2 i , b i , c i , d i are respectively in the ranges between about -1,000,000,000 and 0, about -100,000 and 0, 0 and about 10, and 0 and about 10,000, for all values of i between 1 and M.

Claim 31

The filter bank of claim 27, said bank suitable for use in recognizing speech features in speech input signals, wherein a 1 i , a 2 i , b i , c i , d i are related to characteristics of a representative human cochlear according to the following relations, where the height and width of the scalae vestibuli and tympani of the cochlear are H, where the density of a fluid in the scalae is ρ, where the basilar membrane of length L of the cochlear comprises M sections, and where the ith section has mass m i , breadth B i , damping coefficient q i , stiffness coefficient K i : μ.sup.i =m.sup.i +(1/3)ρHB.sup.i +2ρ(D/H).sup.2 (B.sup.i).sup.2 * i+q.sup.i (T/2)+K.sup.i (T/2).sup.2 ; a.sup.i.sub.1 =(m.sup.i +(1/3)ρHB.sup.i +2ρ(D/H).sup.2 (B.sup.i).sup.2 *i-q.sup.i (T/2)-K.sup.i (T/2).sup.2)/μ.sup.i ; a.sup.i.sub.2 =-2K.sup.i (T/2)/μ.sup.i ; b.sub.i =-2ρDB.sup.i * i/μ.sup.i ; c.sup.i =2ρD.sup.2 (B.sup.i /H)/μ.sup.i ; d.sup.i =B.sup.i /H; and D=L/M.

Claim 32

The filter bank of claim 29, wherein the quantity B i varies linearly with the distance of the ith section from one end of the basilar membrane.

Claim 33

The filter bank of claim 23, wherein said M filters have substantially non-overlapping frequency pass bands.

Claim 36

The system of claim 32, said filtering means including at least two contrast filters at different resolutions.

Claim 37

The system of claim 33, wherein the input speech signal contains significant local variations in the two dimensional representation of the input speech signal, said two dimensional representation having a tonotopic direction and a time direction, said filtering means further comprising an orientation filter providing an output indicating orientations of said local variations with respect to the two directions.

Claim 38

The system of claim 34, wherein said orientation filter has a resolution different from those of the contrast filters.

Claim 39

The system of claim 34, wherein said local variations define boundaries with said orientations, said contrast filters providing outputs indicating the presence of significant speech features in the input speech signal by indicating said boundaries.

Claim 40

The system of claim 34, said two dimensional representation representing tonotopic characteristics of said input signal in tonotopy and time, said system further comprising: a neural network including: at least one phoneme- or suprasegmental-related formation layer for processing the contrast information and the output of the orientation filter to provide a plurality of phoneme- or suprasegmental-related feature formation maps.

Claim 41

The system of claim 32, wherein the input speech signal contains local variations in the two dimensional representation of the input speech signal, said variations having orientations and defining boundaries in said representation, said two dimensional representation having a tonotopic direction and a time direction, said filtering means comprising: one or more orientation filters each providing an output indicating orientations of said local variations with respect to the two directions; and one or more contrast filters, each filtering the output of one of the orientation filters or a signal derived therefrom to provide an output indicating the locations of the boundaries.

Claim 42

The system of claim 38, said system comprising two or more orientation filters providing outputs at different resolutions and two or more contrast filters, wherein each contrast filter filters the output of one of the orientation filters to provide an output indicating the locations of the boundaries.

Claim 43

The system of claim 38, said system further comprising one or more smoothing filters connected in series in at least one sequence, each sequence of smoothing filters filtering the output of one of said one or more orientation filters or a signal derived therefrom, defining a corresponding orientation filter, to each provide an output, so that the output of each smoothing filter in a sequence contains average value information indicating orientations of said local variations at a coarser resolution than those indicated by the output of the corresponding orientation filter and by the output of a different smoothing filter upstream in the same sequence, and wherein each contrast filter filters either the output of an orientation filter or the output of a smoothing filter, or a signal derived therefrom to provide an output, said contrast filters each providing an output that is coarser in resolution than the output of the orientation filter or smoothing filter filtered by said each contrast filter, said output of each contrast filter containing contrast value information indicating the locations of the boundaries at different resolutions.

Claim 44

The system of claim 40, wherein the output of each smoothing filter in the sequence indicates orientations of said local variations at a resolution coarser than those indicated by the output of the immediately preceding orientation filter or smoothing filter by a factor of 2, and said contrast filters each provides an output that is coarser in resolution than the output of the orientation filter or smoothing filter filtered by said each contrast filter by a factor of 2.

Claim 45

The system of claim 40, wherein said sequence includes at least three smoothing filters, said system including at least three contrast filters, one of said contrast filters filtering the output of the orientation filters and each of the remaining contrast filters filtering the output of a corresponding smoothing filter, except for the last smoothing filter in the sequence.

Claim 46

The system of claim 40, said system comprising at least a first and a second orientation filter for detecting local variations with a first and a second orientation respectively, wherein the output of the contrast filter filtering the output of the first orientation filter indicates the presence of a boundary with said first orientation, defining a first type of boundary, different from a second type of boundary, the presence of which is indicated by the output of the contrast filter filtering the output of the second orientation filter.

Claim 47

The system of claim 43, wherein the orientation of each of the two types of orientation filters is vertical, horizontal, or inclined with respect to the vertical and horizontal directions of the two dimensional representation.

Claim 48

The system of claim 43, wherein said smoothing filters are connected in series in at least two sequences, each sequence filtering the output of its corresponding orientation filter to provide outputs indicating orientation of the first or second type of boundaries at coarser resolutions, said contrast filters filtering the outputs of said orientation or smoothing filters in each sequence providing outputs indicating the locations of said first or second type of boundaries.

Claim 49

The system of claim 40, said two variables being frequency and time, said two dimensional representation representing tonotopic characteristics of said input signal in tonotopy and time, said system further comprising: a neural network including: at least one phoneme- or suprasegmental-related formation layer for processing the contrast value information provided by at least some of the contrast filters and the average value information provided by at least one smoothing filter in the sequence to provide a plurality of phoneme- or suprasegmental-related feature formation maps.

Claim 50

The system of claim 46, wherein said neural network processes the average value information provided by the last smoothing filter in the sequence to provide said plurality of phoneme- or suprasegmental-related feature formation maps.

Claim 51

The system of claim 47, wherein said layer provides said phoneme- and suprasegmental-related feature formation maps by processing contrast value information derived by the contrast filters of all the resolutions except the contrast filter of the highest resolution and the average value information derived by the last smoothing filter of the sequence, and provides consonant-related feature formation maps by processing contrast value information provided by all of the contrast filters and the average value information provided by the last smoothing filter in the sequence.

Claim 52

The system of claim 48, said neural network further comprising at least a first phoneme- or suprasegmental-related deformation layer which performs a local-averaging operation on each phoneme- or suprasegmental-related feature formation map to provide a phoneme- or suprasegmental-related deformation map, and on each consonant-related formation feature map to provide a consonant-related deformation map, said formation and deformation layers defining a first pair of formation and deformation layers.

Claim 53

The system of claim 49, said neural network further comprising additional pairs of formation and deformation layers, said pairs connected in a sequence of pairs with the first pair as the first in the sequence, wherein at least one additional pair including: an additional phoneme- or suprasegmental-related formation layer for processing said phoneme- or suprasegmental-related deformation map provided by the deformation layer of a preceding pair to provide an additional set of phoneme- or suprasegmental-related feature formation maps and for processing said consonant-related and phoneme- or suprasegmental-related deformation maps provided by the deformation layer of a preceding pair to provide an additional group of consonant-related feature formation maps.

Claim 54

The system of claim 50, said at least one additional pair further comprising: an additional phoneme- or suprasegmental-related deformation layer for processing said additional set of phoneme-related or suprasegmental-related feature formation maps to provide an additional phoneme- or suprasegmental-related deformation map.

Claim 55

The system of claim 51, wherein the last pair of formation and deformation layers in the sequence of pairs of formation and deformation layers provide a vowel-related deformation map, a suprasegmental-related deformation map, and a consonant-related deformation map that together provide information concerning a phoneme, a group of phonemes, or a suprasegmental.

Claim 56

The system of claim 32, said filtering means providing an output indicating the presence of one of three types of boundaries: vertical, horizontal or inclined boundaries.

Claim 57

The system of claim 32, wherein the filtering means output indicates at least one of the following: onset, frequency, rise or fall of significant tones in the input signal.

Claim 60

The system of claim 55, said phoneme- or suprasegmental-related formation layer processing the output of the filtering means to provide a plurality of phoneme- or suprasegmental-related feature formation maps; and said phoneme- or suprasegmental-related deformation layer performing a local averaging operation on said plurality of phoneme- or suprasegmental-related feature formation maps to provide invariant phoneme- or suprasegmental-related feature maps.

Claim 61

The system of claim 56, wherein said filtering means includes at least one bank of R filters, R being a positive integer greater than 1, each filter providing an output, said filters being indexed from 1 to R along a tone axis and the outputs of the filters at S different times being indexed from 1 to S along a time axis, forming an R by S array of filter outputs; wherein said at least one formation layer includes a plurality of tracts, each tract including: means for superposing at least one t by u array of weights upon a portion of said R by S array, t and u being positive integers, so that each weight corresponds to a filter output in said R by S array; and means for multiplying each weight by the corresponding filter output to obtain products and summing said products to obtain a first processed value for said superposition upon said one portion.

Claim 62

The system of claim 57, wherein said superposition means superposes said t by u array upon each of rs different portions of the R by S array, r and s being positive integers smaller than R and S respectively, where the portions form an r by s array along the frequency and time axes, and wherein, for each superposition upon a portion, said multiplying means multiplies each weight by the corresponding filter output to obtain products and summing said products to obtain a first processed value, thereby obtaining from the rs superpositions an r by s array of first processed values along said two axes.

Claim 63

The system of claim 58, wherein said at least one deformation layer includes a plurality of tracts each comprising: second means for superposing at least one v by w second array of coefficients upon a portion of said r by s array of first processed values, v and w being positive integers smaller than r and s respectively, so that each weight in the second array corresponds to a processed value in said r by s array of first processed values; and second means for multiplying each coefficient in the second array by the corresponding first processed value to obtain products and summing said products to obtain a second processed value for said superposition upon said one portion, said coefficients of the second array being such that said second processed value is an average value of the corresponding first processed values.

Claim 64

The system of claim 59, wherein said v by w second array of coefficients is a matrix with most of its coefficients substantially equal to each other.

Claim 65

The system of claim 57, wherein said superposing means includes t(u-1) delay elements for delaying t outputs of the R filter outputs and tu amplifying elements for amplifying the tu outputs of the R filter outputs.

Claim 66

The system of claim 55, said output of the filtering means indicating one or more elementary tonotopic features of the input speech signal in a two dimensional representation of the signal in tonotopy and time, said features including onset, rise and fall of any significant tones in the input speech signal over time.

Claim 69

The method of claim 63, wherein said filtering step in step (b) provides an output that indicates at least two of the three features, namely, the onset, rise and fall of any significant tones of the input signal over time.

Claim 70

The method of claim 63, wherein said filtering step in step (b) provides an output that indicates also frequency of any significant tones of the input signal.

Claim 71

The method of claim 63, wherein said filtering step in step (a) employs a bank of M filters having different frequency pass bands, M being a positive integer greater than 1, each filter providing an output, said filters being indexed from 1 to M by pass band frequencies of the filters along a frequency axis and the outputs of the filters at N different times being indexed from 1 to N along a time axis, N being a positive integer greater than 1, forming an M by N array of filter outputs, said filtering step in step (b) including: superposing at least one p by q array of filter coefficients upon a portion of said M by N array of filtered outputs, p and q being positive integers, so that each coefficient corresponds to a filter output in said M by N array; and multiplying each coefficient by the corresponding filter output to obtain products and summing said products to obtain a first processed value for said superposition upon said one portion.

Claim 72

The method of claim 66, wherein the coefficients of said p by q array in the superposing step indicates one of the following elementary tonotopic features: the onset, frequency, rise or fall of any significant tones in the input signal.

Claim 73

The method of claim 63, said filtering step in step (b) including filtering the output of the filtering step in step (a) at different resolutions to provide speech context information for the recognition of phonemes.

Claim 76

The method of claim 69, said filtering including filtering at different resolutions to derive contrast information.

Claim 77

The method of claim 70, wherein the input speech signal contains significant local variations in the two dimensional representation of the input speech signal, said two dimensional representation having a vertical tonotopic direction and a horizontal time direction, said filtering further comprising providing an output indicating orientations of said local variations with respect to the two directions.

Claim 78

The method of claim 71, wherein said providing step provides an output at a resolution different from those of the contrast information derived by the filtering.

Claim 79

The method of claim 71, wherein said local variations define boundaries with said orientations, said filtering at different resolutions generating outputs that indicate the presence of significant speech features in the input speech signal by indicating said boundaries.

Claim 82

The method of claim 74, said processing and performing step processing the output of the filtering step to provide a plurality of phoneme- or suprasegmental-related feature formation maps and performing a local averaging operation on said plurality of phoneme- or suprasegmental-related feature formation maps to provide invariant phoneme- or suprasegmental-related feature maps.