JPS61126600A - Sound wave input processing - Google Patents

Abstract

An acoustic processor and a method of processing an acoustic wave input which includes a non-linear auditory model of the neural firing rate in the ear. The firing rate is determined by the replenishment of neurotransmitter and loss of neurotransmitter due to spontaneous decay, spontaneous firing, and acoustic wave inputs defined, preferably, in sones. The number of free parameters in the acoustic processor is reducible to one, namely the ratio R of two steady-state firing rates each resulting from a different loudness. Preferably, the free parameter is adjusted to minimize steady-state effects which are adversely impacted by speaker differences, background noise, distortion, and the like. The present invention addresses the general problem of processing an acoustic wave input in a speech recognition system and addresses the specific problem of adjusting acoustic processor performance to reduce adverse effects.

Description

[Detailed description of the invention] The present invention will be explained in the following order.

A. Industrial field of application B. Summary of the disclosure C1. Prior art (FIG. 2) D9. Problems to be solved by the invention E1. Means for solving the problems F. Example F 1. Design principle F 2. Hearing organ model (Figures 3 to 6) F 3
．． Operation of the acoustic processor (Figures 1 and 7) F4. General characteristics of the auditory organ model F5.
Alternative Embodiment G6 Effects of the Invention A, Industrial Field of Application The present invention relates primarily to the technology of speech recognition, and in particular to the technology of selecting features and parameters that affect the feature values at the front end of a speech recognition system.

B. SUMMARY OF THE DISCLOSURE An acoustic processor and acoustic wave input processing method that includes a nonlinear auditory organ model of the neural firing rate of the human ear is disclosed. The neural firing rate is determined by neurotransmitter replenishment and neurotransmitter loss due to spontaneous decay, spontaneous firing, and acoustic wave input, preferably converted into son units. The number of free parameters in the sound processor is 1
can be limited to This is the ratio R of two steady-state firing rates, each resulting from a different sound loudness. It is desirable to adjust the free parameters so as to minimize, if possible, the negative effects on the stable state due to speaker differences, background noise, distortion, etc. The present invention addresses the general problem of processing acoustic wave input in speech recognition systems, as well as the specific problem of adjusting the performance of an acoustic processor to reduce its negative effects.

C1 Prior Art Speech recognition systems or devices generally aim to automatically convert raw speech into another format, such as text format. In this regard, Pattern E Newsletter Volume 5 No. 2 (
1985) ``Maximum establishment method for continuous speech recognition'' (Bahl et al., 1985), pp. 179-190.
l,'A maximum Likelihood A
pproach to ContinuousSp
eech Recognition”, IEEE Tr.
anaaction'ns on Pattern
Analysis and Machine Int.
e 11 igences VolumePAMI-5
, No. 2, pp. 179-190 (1983)),
Several speech recognition methods have been described. In either method, hypotheses can be made about the text generator that determines what is said. The text generator is followed by a speaker that generates a live speech waveform. The audio waveform provides input to the audio processor. The output of the acoustic processor goes into a language decoder.

According to Barr et al., there are several ways to connect the components of this system. For example, a speaker and an acoustic processor may be combined to form an acoustic channel, where the speaker converts text to speech, and the acoustic processor acts as a data converter and compressor and feeds a string of symbols to a language decoder. can. A language decoder recovers source text from a string of symbols.

FIG. 2 shows a specific configuration example of a conventional audio processor.

Acoustic wave input (e.g. live audio) is sampled at a predetermined rate by an A/DC analog/digital converter 1.
Enter 02. A typical sampling rate is once every 50 microseconds.

A time window generator 104 is used to determine the edges of the digital signal.
is connected. The output of the time window generator 104 is FFT
(Fast 7-layer transform) device 106, FFT device 1
06 outputs a frequency spectrum for each time window.

The output of the FFT device 106 is represented by the symbol L 1L 2...
L.

processed in such a way as to result in Four devices--feature selection device 108, cluster device 110, prototype device 112, and symbolization device 114--cooperatively create symbols. When creating a symbol, a prototype is defined as a point (or vector) in space based on selected features and acoustic input, and then a corresponding point in space that can be contrasted with the prototype by the same selected features. (or a vector).

In particular, when defining the prototype, a set of points is defined as cluster device 1
The data are classified into 10 clusters. A prototype of each cluster is created by a prototype device 112 (in relation to the center locus or other characteristics of the cluster). The created prototype and the acoustic input (both characterized by the same selection feature) are input to the encoder 114, which performs the matching.

PROBLEM SOLVED BY THE INVENTION The selection of appropriate features is an important factor in extracting symbols representing acoustic (speech) wave inputs. The real deal! A relates to a preferred feature selection device 108 and front-end processors and speech recognition systems that include such a feature selection device.

The feature selection device of the present invention is responsive to model the peripheral auditory system. That is, the feature selection device of the present invention considers the auditory nerve firing rate at the selected frequency as a feature that defines the acoustic input. Traditionally, the human ear has been modeled differently (Hall and Schrader's paper “A Model of Mechanical/Neural Transduction in the Receptors of the Auditory Organ” published in the Journal of the Acoustical Society of America, Volume 55, No. 5).
and Schroeder ”Model f
or Mechanical to Neural
Transduction In theAudi
However, in the speech recognition system of the present invention, a model different from that of Hall and Schrader is used. Furthermore, the model realized by the present invention is and takes into account factors that are not mentioned in Schrader's paper or are analyzed in a different way. For example, the speech amplitude input has a wide dynamic range, so the neural firing rates in Hall and Schrader's model are accurate. However, the model implemented in the present invention solves this problem.

According to the test results of a number of items, the respective word error rates are improved by replacing the conventional device with the parameter selection device of the present invention. The present invention aims to improve the performance of speech recognition systems by using a parameter selection device and an auditory organ model based on neural firing.

Furthermore, the present invention performs critical band filtering to match the human ear and reflects the behavior of the basilar membrane as a frequency analyzer.

That is, the present invention also provides a basilar membrane-like response to receiving an increased volume of two components of an audio input spread across different critical bands, and filtering acoustic waves in similar bands. desirable.

Furthermore, it is an object of the invention to define the characteristics of the feature selection device as a function of volume (loudness), preferably in compressed amplitude form, such as in son units. Furthermore, different frequencies cause non-uniformity in volume (per son),
And the volume in units of phons.

Since level fluctuations cause non-uniformity in volume (in units of 7), the present invention provides a volume equalization adjustment device and a volume compression device to perform normalization. The feature selection device achieves the above objectives in a speech recognition system and contributes to the realization of large word recognition in real-time systems where separation of speaker's pronunciation and different words is desirable.

The model of the present invention, which achieves the foregoing objectives, first digitizes the waveform and then determines the magnitude of the waveform as a function of frequency in successive discrete time periods. It is desirable to classify this thickness into critical bands (as in the case of the basement membrane). According to this model, modeled neural firing within the pinna (for each critical frequency) occurs at rate f;
And it is assumed that this neural firing is due in particular to the modeled amount n of neurotransmitters in the pinna. The change rate of neurotransmitters for each critical band is the neurotransmitter replenishment rate A
o and loss.

The loss of neurotransmitters over time appears to be due to several factors, including:

(a) Sh-n However, sh corresponds to the natural decay or extinction of neurotransmitters over time, regardless of acoustic wave input.

(b) So-n However, So corresponds to the rate of spontaneous neural firing that occurs regardless of acoustic wave input.

(c) DL-n This corresponds to neural firing as a function of the coefficient-compressed sound [L].

Based on these factors, the above model is expressed by the following equation.

dn/dt=Ao-(So+Sh+DL)n'(
1) f=(So+DL)・n (2
) Equations (1) and (2) are defined for each critical frequency band. However, t is time.

The present invention is also relevant to determining the "next state" of the amount of neurotransmitter that is then to be used in determining the firing rate f'f. In general, the following condition.

"state" is defined by the following equation.

n(t+Δt) =n(t)+(dn/dt
)Δt(3) Equation (3) for “next state” and equation (1) for neurotransmitter changes serve to define the next value of firing rate f. Incidentally, the firing rate f (per frequency band) is nonlinear in that it depends multiplicatively on the preceding state. This closely follows the temporal adaptability of the auditory system (as discussed above).

The firing rate of each frequency band simultaneously gives characteristics of speech recognition encoding. For example, for 20 bands, 20 firing rates (one for each band) simultaneously give a vector in 20 dimensions of space. This vector is input to the encoder 114, and the vector corresponding to the acoustic wave input can be compared to the stored data and the created symbol.

Both f and n in equations (1) and (2) tend to produce large DC pedestals. If we want to increase the dynamic range of each term in the equation, a set of equations is given that reduces the pedestal height. Incidentally, in the present invention, n
is divided into a steady state component n and a fluctuation component n (t), and equation (2) becomes as follows.

Similarly, if we define n = n + n(t) and ignore the constant term, equation (3) becomes as follows.

n(t+Δt)=n(1-8oΔt)−f(1(5
) Equations (4) and (5) constitute the special case output and update equations, respectively, that are applied to the signal of each critical frequency band during successive time frames.

For each frequency band, equation (4) defines the vector dimension of each time frame. This is better than the basic output from equations (1)-(3).

Although the performance of a speech recognition system can be improved by adjusting or changing the values of parameters that affect the feature values, it is especially important to check the system for improvement after each adjustment or change. It is a time consuming process when there are many parameters that can be adjusted or changed.

Therefore, another object of the present invention is to provide a functional auditory organ model with as few free parameters as possible as a feature selection device. By using empirical data specifying some of the terms in the above equation, the number of free parameters is reduced to one.

Thus, the present invention allows one to tune the system model by changing one parameter and determine how much the performance of the system can be changed or improved. In particular, 1
The two parameters are ratios defined as:

R represents the ratio of (a) the stable excitation rate when the volume is maximum (eg, the limit of sensation) and (b) the stable excitation rate when the volume is minimum (eg, 0).

According to the present invention, R is preferably the only system variable that adjusts or changes the parameter.

By equating the volume included in the aforementioned model with respect to frequency and compressing it with the volume, the present invention can reduce the occurrence of inconsistent output patterns in the case of similar acoustic wave inputs. This is done by emphasizing transient parts of the acoustic (speech) input that are unaffected by factors such as differences in the frequency response of the acoustic channels, speaker differences, background noise and distortion.

Finally, further improvements are made regarding the definition of equivalent loudness in extracting the relationship between loudness and intensity from the acoustic input. In particular, a histogram is maintained for each of the critical frequency bands;
Speech is assumed to be present if a predetermined number of filters (in a critical frequency band) produce an output above a certain value for a predetermined period of time. Sensory limits and audible limits are then determined and used to normalize the voice based on the histogram of the hypothesized voice over time.

Means for Solving the E9 Problem The improved auditory organ model of the present invention is used in a speech recognition system. In an embodiment of the invention, a method of processing acoustic wave input in a speech recognition system includes the following steps.

(a) measuring the acoustic wave input sound in each of at least one frequency band;

(bl Determine the neural firing rate as a function of the measured sound level in each frequency band in the auditory organ model.

(c) Expressing acoustic wave input as neural firing rates determined in each frequency band.

(d) For each frequency band, determine the current tk of neurotransmitters available for neural firing.

(, for each frequency band, the rate of change of the neurotransmitter is determined based on the recruitment constant representing the rate at which the neurotransmitter is produced and the neural firing rate determined for each frequency band.

The neural firing rate is determined by the amount of neurotransmitters that can be used for neural firing, and the amount of neurotransmitters that can be used for neural firing in the "next state" depends on the amount of neurotransmitters that can be used in the "current state" and Determined by the rate of change of the transmitter.

Measuring the sound includes measuring the volume of the input acoustic wave in each of a plurality of frequency bands, each frequency band corresponding to a critical frequency band associated with the human ear and defined in compressed amplitude form. It is desirable to include the sound volume.

F, Example F 1. Design Principles In FIG. 2, a conventional audio channel typically comprises multiple parameters whose values can be adjusted to change its performance. Inspecting performance that changes in response to parameter variations requires operating the entire acoustic processor 100, and generally takes about one day. Therefore, the more parameters that change, the more difficult and time-consuming the task of testing for performance changes.

The design principle of the present invention is that the acoustic processor 1001C facilitates performance improvement with a minimum number of adjustable parameters.
It is to provide.

F2. Hearing organ model (FIGS. 5-6) A hearing organ model is created in accordance with the present invention and used in an acoustic processor of a speech recognition system.

The auditory organ model will be explained with reference to FIG.

Figure 3 shows parts of the human inner ear. In particular, white hair cells 20
0 and the distal end 202 extending into a liquid-containing groove 204 is shown in detail. Further, upstream from the white hair cell 200, there are an exogenous cell 206 and an end portion 208 extending into the groove 204.
It is shown. Nerves that transmit information to the brain are connected to the white hair cells 200 and the exogenous cells 206. In particular, neurons undergo chemical changes that cause electrical pulses to be carried along to the brain and processed. Electrochemical changes are stimulated by mechanical movement of basement membrane 210.

It is known in the art that basilar membrane 210 acts as a frequency analyzer of acoustic waveform input, with sections along basilar membrane 210 responding to respective critical frequency bands. Each portion of basilar membrane 210 that responds to a corresponding frequency band influences the perceived loudness of the acoustic waveform input. That is, the loudness of the tones is perceived to be greater when the two tones are in distinct critical frequency bands than when the two tones of similar power intensity occupy the same frequency band. It has been found that there are 22 orders of magnitude critical frequency bands defined by the basilar membrane 210.

Tailored to the frequency response of the basilar membrane 21°0, the present invention physically defines in good form the input acoustic waveform in some or all of the critical frequency bands, and then for each defined critical frequency band. Examine signal components separately.

This function is performed by suitably filtering the signal from the FF,,T device 106 (FIG. 2) and providing a separate signal to the feature selection device 108 for each critical frequency band examined.

A separate input is also provided by the time window generator 104 (preferably two
5.6 ms) time frame. Therefore, feature selection device 108 preferably includes 22 signals. Each of these signals represents the sound intensity 'jt of a particular frequency for each time frame.

The signal is preferably filtered by a conventional critical band filter 300 of FIG. Then separately a volume equalization converter 302 that perceives changes in volume as a function of frequency.
Processed by By the way, given dB at one frequency
The perceived loudness of a first tone at a level may differ from the loudness of a second tone at the same dB level at another frequency. Volume equalization converter 302 transforms the signals of each frequency band so that each is measured on the same loudness scale based on empirical data. For example, the volume equalization converter 602 may be used as described by Fletcher and Munson in 1963.
With some modifications to the study, acoustic energy can be mapped to equivalent loudness. Figure 5 shows the results of modifications made to the previous study. According to Figure 5, 40dB
It can be seen that a 1 KHz tone corresponds to the volume level of a 100 Hz tone at 60 dB.

The volume equalization converter 602 adjusts the volume according to the curve shown in FIG. 5, producing the same volume regardless of frequency.

In addition to the dependence on frequency, changes in power do not correspond to changes in volume, as is clear from looking at specific frequencies in Figure 5. That is, variations in sound intensity, or amplitude, are not reflected in similar changes in perceived loudness at all points. For example, at a frequency of 100Hz, approximately 110
A 1 0 dB perceived volume change around dB is much larger than a 1 0 dB perceived volume change around 20 dB. This difference is illustrated by a volume compressor 504 (FIG. 4) that compresses the volume in a predetermined manner.

The volume compression device 604 compresses the power P to its cube root P1/3 by replacing the volume amplitude measurement value in units of phone to units of son.

FIG. 6 shows the known empirically determined Hong-to-Thorn relationship. Through the use of sone units, our model remains nearly accurate at large audio signal amplitudes. One sone is defined as a volume of 40 dB at IKHz.

In FIG. 4, a novel time-varying response device 306 is shown. The device operates with volume equalization and volume compression signals associated with each critical frequency band. In particular, for each frequency band examined, the neural firing rate f is determined for each time frame. The firing rate f is defined according to the invention as follows.

f=(So+DL)n
f force, where n: amount of neurotransmitter, So: spontaneous firing constant related to nerve firing independent of acoustic waveform input;
L: Volume measurement value, D: Displacement constant.

5o-n corresponds to the spontaneous neural firing rate that occurs regardless of the presence or absence of acoustic wave input, and DLn corresponds to the firing rate due to acoustic wave input.

An important point is that in the present invention, the value of n changes with time according to the following equation.

dn/dt=Ao-(So+Sh+DL)n (
8) However, Ao: recruitment constant, Sh: spontaneous neurotransmitter decay constant. The new relationship shown in equation (8) is that while neurotransmitters are generated at a constant rate Ao,
Decay (sh −n ), (bin spontaneous firing (So・
n), (e) Nerve firing due to acoustic wave input (DL-
n). It is assumed that these modeled phenomena occur at the locations shown in FIG.

Equation (8) reflects the fact that the order quantity and the order firing wheel of the neurotransmitter are non-linear in that they are at least multiplicatively related to the current quantity of the neurotransmitter. In other words, the amount of neurotransmitter in state (t+Δt) is
+dn/dt). That is, n(t+Δt)=n(tl+(dn/dt)
) ・Δt(9) holds true.

Equations (7), t8) and (9) represent the operation of the time-varying signal analyzer. Time-varying signal analyzers point to the fact that the auditory system is adaptive and the signals of the auditory nerve are non-linearly related to the acoustic wave input. By the way, the present invention
By closely following the apparent temporal changes in the neural system, it provides the first model for implementing non-articulated signal processing in speech recognition systems.

In order to reduce the number of unknown terms in equation (8), the present invention uses the following equation, which is applied to a constant volume level.

So + Sh + D L = 1/τ (10) where τ is the measurement of the time after the audio wave input is generated until the auditory response drops to 37% of its maximum 1 position. τ is a function of volume, and according to the present invention is taken from a known graph displaying the attenuation of the response for various volume levels. That is, when a tone of constant volume is generated, an initial high level response occurs, after which the response decays to a steady state level with a time constant r. When there is no acoustic wave input, τ=τ
It is 0. This is about 50 milliseconds.

When the volume is L, τ=r. This is max
The maximum length is about 60 mm. By setting Ao=1, 1/(
So+Sh) is determined to be 5 centiseconds when L=00. L is L, L = 20max
In the case of max Thorn, the following formula (1 holds true).

So+Sh+D(20)=1/! 10
(11) Based on the above data and formulas, So and sh are determined according to formulas (12) and (15).

So=DLmax/[R+(DLm, Lx・τ.・R)
-1) (12) Sh = 1/τo So
(15) where f steady state 1 represents the firing rate at a given volume when d n /d t is 0.

R is the only variable left in the sound processor.

Therefore, the performance of this processor is changed simply by changing Rfe. That is, R is one parameter that can be adjusted to change performance, usually meant to minimize steady-state effects with respect to transient-state effects. Inconsistent output patterns for similar speech inputs are generally caused by differences in frequency response, speaker differences, background noise, and other factors (affecting steady-state but not transient portions of the speech signal). It is desirable to minimize steady-state effects since they are caused by distortions (no). The value of R is preferably set by optimizing the error rate of the complete speech recognition system. The optimum value thus found is R=1.5. In that case, the values of So and sh are 0.0888 and 0.11111, respectively, and the value of D is 0.00666.
is obtained.

F3. Flowchart of the operation of the acoustic processor (FIGS. 1 and 7) FIG. 1 is a flowchart of the operation of the acoustic processor according to the present invention. Preferably sampled at 20KHz, 25
．． The digitized audio during a time frame of 6 milliJ seconds is preferably passed through a Hanning window, the output of which is subjected to a Fourier transform at 10 millisecond intervals. The conversion output is filtered to provide a full power density output in each of at least one frequency band (preferably all critical frequency bands, or at least 20 bands). The power density is then converted from log magnitude to volume level. This can be done by the modified graph of FIG. 5 or by the process shown in FIG.

In FIG. 7, first, the perceptual limit Tf and audible limit Th of each filtered frequency band m are assigned 120 dB and OdB, respectively (step 410). Thereafter, the audio counter, total frame register, and histogram register are reset (step 420).

Each of the histograms includes bins, with each bin representing the number of samples or counts during which the power or similar measurement (in a given frequency band) is within a respective range. In the present invention, the histogram preferably represents the number of centiseconds during which the volume is in each of a given number of volume ranges (for a given frequency band). For example, the sixth
In the frequency band, there may be 20 centiseconds between 10 dB and 20 dB power. Similarly, in the 20th frequency band, a total of 1000 between 50dB and 60dB
There may be 150 centiseconds of a centisecond. Percentiles are taken from the total number of samples (ie, centiseconds) and the binned counts.

A frame of filter output for each frequency band is examined (step 430). The bins in the appropriate histogram (91 per filter) are incremented (step 440).
. The number of bins with amplitudes greater than 55 dB are aggregated into filters (i.e., frequency bands) (step 450);
The number of filters that indicate the presence of audio is determined (step 460).

Minimal hint of voice presence (e.g. 6 out of 20)
If there is no filter, the next frame is examined (
Step 430. If there are enough filters to indicate the presence of audio, the audio counter is incremented (step 4).
70). The audio counter is incremented until audio occurs for 10 seconds (step 480) and new T and T5 values are determined for each filter.

The new values of and Th for a given filter are determined as follows. For T, the dB value of the bin holding the 65th sample from the top of 1000 bins (i.e., the 96.5th percentile of audio) is defined as BINH, and T, =BINH+40dB is set. For T, the dB of the bin holding the (0,01)th (total number of bins - voice count) value from the lowest bin
The value is defined as BINL. That is, BINL is one bin of the number of samples in the histogram excluding those classified as audio. T is T = BINL
-30dBh.

In FIG. 1, the sound amplitude is converted to tones and compressed (steps 670 and 375) based on the updated limits, as described above. An alternative method of introducing and compressing sonic units is to reduce the filter amplitude “a” (after the bins have been incremented) to
'' and convert it to 9 dB using the following formula.

B a = 201og1o(a) -10(15) Each of the filter amplitudes is then compressed to a range between 0 and 120, giving equivalent loudness by:

aeq=120(adB-T)/(T-T) (
16) h fh Next, aeql is converted from the volume level (in units of phons) to an approximate value of the volume in units of 40 dB (I at 40 dB).
It is desirable to map the KHz signal to 1).

B L = (a8q'-50)/4 (17
) Then, the approximate value L of the sound volume in units of sones is obtained from the following equation.

L = 10 (LdB)/20 (18)
L is used as input to equation (force and (8)) to determine the output firing rate f for each frequency band (step 38
0). In the case of 22 frequency bands, a 22-dimensional vector characterizes the acoustic wave input over successive time frames, but typically the 20 frequency bands are examined using a scaled filter bank in Mel.

n's "next state" before processing the next time frame.

” is determined according to equation (9) (step 390).
.

The acoustic processors described above require improvement when used with DC pedestals where the firing rate f and neurotransmitter content n are large. That is, when the dynamic range of the f and n equation terms is important, the height of the sibe destal is reduced according to the following equation. ``In the stable state, when there is no acoustic wave input signal (L=0), equation (8) can be solved for the internal state n in the stable state by the following equation.

n=A/(So+Sh) (19)
The internal state of the amount of neurotransmitter n (tJ is expressed as a steady-state part and a fluctuating part, as shown in the following equation:

n(t) = n + n(t)
(20) Combining equations (7) and (20), we get
The firing rate is obtained as follows.

f(t)−” (S o +D−L ) (n +n(
tJ) (21) The So-n term is constant, but all other terms are either varying parts of n or (D-
L). Since further processing concerns only the square of the difference between the output vectors, the constant term is ignored. The following equation is obtained from equations (21) and (19).

Considering equation (9), the "next state" is as follows.

n(t+Δ1)=;(1+Δt) +n(t+Δt)
(25) n('t+Δt)=n(t)+A-(So
+Sh+I)L)・(n + n(tl)
(24) n(t+Δt)=n(tl−
(sh-n(t)-(so+Ao-L)-n(t)(A
o-L-D)/(So+sh)+Ao-(So-A
o)+(Sh-Aon)/(So+Sh)
(25) Equation (25) becomes as follows if all constant terms are ignored.

n(t+Δ1)=circulationtbar1-8o・Δt>-'nt>
(26) Equations (21) and (26) constitute the output and state update equations applied to each filter during each 10 ms time frame. The result of using these equations is a vector of 20 elements every 10 ms, each element of which corresponds to the firing rate of a respective frequency band in the mel-scaled filter bank.

Regarding the embodiment designated F3, the flowchart of FIG. 1 shows that f
, dn/dt and n(t+Δt).

F4. General features of the auditory organ model The auditory organ model of the present invention realizes the following features in a good manner as described above.

(a) The auditory nerve responds to acoustic signals as if passing through a critical bandwidth filter.

(b) In response to zero acoustic wave force (silence), nerves fire at a spontaneous rate.

(C) When responding to loud sounds, the firing rate is large;
It decreases with a time constant of about 50 milliseconds.

(d) Neural firing rate in response to large acoustic wave signals turning off decreases with a recovery time constant of approximately 50 ms.

(e) Steady-state responses to soft and loud sounds are simple functions of loudness, as defined psychophysically.

(f) The balance between transient and steady state responses is adjusted to emphasize the transient response of the system.

(g) The model operates semi-independently for each critical band.

F5. Alternative Embodiments First, although it is desirable that the loudness be in sones or other compressed forms, other measures of loudness, i.e., the intensity of power, may be sacrificed, perhaps at the expense of some of the benefits gained from utilizing sones. It is also possible to give Eq. Second, it is desirable but not required to define a frequency band as a critical band of basement membrane 210. Therefore, a mel-scaled filter bank of 20 or more channels may be desirable, but is not required.

Sixth, each equation term has a unique value (i.e., τ
=s, (!1ie(!%τLmax=3csec, Ao
= 1, R = 1.5 and L = 20) can also be set to other values, and the terms 5O1sh and D can be set to different f values when the other terms are set to different f values. Value 0, 08
88.0.11111 and C1,00666.

The present invention can be implemented by various software or hardware.

Effects of the G0 Invention The auditory organ model of the present invention filters the critical band that matches the human ear, and allows the operation of the basilar membrane to be detected. It can be reflected.

[Brief explanation of drawings]

FIG. 1 is a flowchart of the operation of the acoustic processor of the present invention, and FIG.
Figure 3 is an explanatory diagram of the human inner ear, Figure 4 is a block diagram of the volume equalization process, Figure 5 is a diagram showing the volume equalization curve, and Figure 6 is a block diagram of a conventional acoustic processor. is a relationship diagram between Hong and Thorn, and Figure 7 is a flowchart of the power density conversion process. 200... White hair cells, 204... Grooves, 206...
... exogenous cells, 210... basement membrane. Applicant Intasano Naru Bi-copper Machines Corporation Sound Processing Figure 1 J1 In the flow of the world's movements (CPS) In the past, the Quantification Weaving Boat Takashi Conversation Yubuto Machine, Mian Hong and Thorn's Gang Person in charge

Claims (1)

[Claims] A method of processing acoustic wave input in a speech recognition system, comprising: measuring the loudness of each of the acoustic wave inputs in at least one frequency band; determining a modeled neural firing rate based on the loudness measurements in each, and displaying the acoustic wave input as the determined neural firing rate for each frequency band. Acoustic wave input processing method.