JPH01255000A - Apparatus and method for selectively adding noise to template to be used in voice recognition system - Google Patents

Abstract

PURPOSE: To obtain a speech recognition system adapted automatically to a noise environment by providing the device with a first means which applies a signal to indicate a noise signal and a means which is coupled to the first means and forms a template to be modulated according to the predicted noise signal in response with the predicted noise signal. CONSTITUTION: The central position of a switch 13 is a modulation template mode position and the output of a spectrum analyzer 12 enters as estimated value noise statistical module 162. The function of the module 162 is to basically execute a noise analysis or to estimate noise statistics by processing the noise. As a result, the template is formed by selectively adding the noise and the speech recognition is executed. Namely, the recognition is executed by changing over a switch 100 according to whether the template of the state having the noise or the template of the state having the noise of an extremely low level or having no noise. As a result, the use in the state having the noise is possible.

Description

[Detailed Description of the Invention] [Field of Industrial Application] This invention relates to speech recognition systems in general, and more particularly to speech recognition systems in which templates are used and each template is generated by selectively adding noise to increase the probability of speech recognition. Regarding the system.

[Prior art]

General speech recognition methods have been greatly developed in recent years and are used in many forms. The idea behind voice recognition is
This means that the information obtained from the speech sounds can be used directly to drive a computer or other means. Basically, in the prior art, the key element for the recognition of speech information is the distribution of energy with respect to frequency. Formant frequencies are those whose energy peaks are particularly important. Formant frequencies are the acoustic components of the oral cavity and are controlled by the tongue, jaw, and lips. For the listener, determining the first two or three formant frequencies is usually sufficient to identify the vowel. Prior art machine recognition thus includes some sophisticated means of determining the width or power spectrum of the incoming speech signal. The first step in speech recognition is preprocessing, which transforms the speech signal into recognizable characteristics, parameters, and reduces the data flow to manageable proportions. A ninth way to accomplish this process is to measure the zero-crossing rate of the signal in several broad frequency bands to provide an estimate of the formant frequencies in this band.

Another means is to represent the speech signal by a meter whose spectrum best matches the spectrum of the input speech signal. This method is known as Linear Predictive Coding (LPC). Linear predictive calling, or LPC, is characterized by its efficiency, accuracy and simplicity. Recognition features extracted from speech are typically averaged over 10 to 40 milliseconds and
= sampled at 100 times/second.

The parameters used to represent and recognize speech are directly or indirectly related to the amplitude and power spectrum. Formant frequencies and linear prediction filter coefficients are examples of parameters that are indirectly related to the speech spectrum. Other examples are sepstral parameters and log area rate parameters.

[Problems to be solved by the invention]

In many other cases, the audio parameters used can be derived from the spectral/4' parameters. The present invention relates to selectively adding noise to spectral parameters that produce speech recognition parameters. The invention applies to all forms of speech recognition using speech nodal parameters derived or capable of being derived from spectral parameters.

In any case, many common speech recognition methods in the past have used templates to perform matching. In this method, words are usually represented in the form of /4' parameter sequences. Recognition is accomplished by comparing unknown template tokens to stored templates using similar predefined methods. Time alignment algorithms are often used to account for the variability in the rate of word production. Thus, template matching systems can achieve high performance with small sets of phonetically distinct words. Some researchers are trying to accurately classify speech from a wide range of speakers! P, questions the ability of such systems to perform.

J, S, Perkel
and fE H Clara)' (D-H, Klat
The paper ``Achieving Precise Speech Segregation: Two Template Pairs of Characteristics''(''Variability and Invariance in the Spite Process'', edited by K/1. F., New Jersey, published by Rolle/Suerbaum Symphony Orchestra Associates, 1985) , Publishers R.A. Cole, R.M. Starr/and M.J. Lasry).

Therefore, as an alternative, many have proposed feature-based methods for speech recognition, such as first identifying the backdoor characteristics that capture phonetically relevant information within the speech signal. An algorithm can be configured to extract features from the speech signal based on this knowledge. Classification is then performed to combine all the features and arrive at a recognition decision. There is an argument that feature-based mental systems have better performance in performing precise speech discrimination than template matching techniques, and are therefore superior to template matching techniques. In any case,
Template matching is a commonly used method for pattern recognition, whereby an unknown is compared to a prototype to determine which is the closest match.

With this determination, multiple changes in classification refinement can also be performed by template matching by feature-based speech recognition using the US model. In this case, only statistical classifiers use the feature vectors and turns. Similarly, when viewed as a total feature of spectral amplitude and LPC coefficients, the vector-based technique is also a feature-based method.

The use of template matching and feature-based systems actually represents different points on a continuum. One of the most important problems with template matching methods is that
There is a difficulty in limiting distance measurements that are sensitive enough for precise speech classification but insensitive to unrelated spectral changes.

One manifestation of this problem is due to the excessive weight given to unnecessary frame/frame changes in the spectrum of long invariant vowels. Accordingly, such problematic prior art provides a number of distance temperaments that are sensitive to phonetic distance and insensitive to irrelevant phonetic differences. For example, ICASSP-82
journal (IEEE Catalog ACH1746-7, No. 1
See Paper 1 'Prediction of Receptive Speech Distance from the Critical Band Spectrum' (by De-Aite Kratt), published on pages 278-1281, 1982).

In any case, the gross seeding zooop I EEE (1985
November issue, ■, 73, 11, pages 1537-1
See page 696). In this document of IEEE, Man/
Machine Speech Comic Unicage.

There are various papers available on the system.
1 Uneno' for specific related issues! ! - It is something that is associated with Hiroi. As can be seen here, an important aspect relevant to any speech recognition system is the performance of its distributional task, the ability of the ninth system to recognize speech for all types of environments.

As mentioned above, templates are used in many speech recognition systems. Basically, in such a system, the utterances are converted into a metric sequence and stored in a computer. The speech waves are conveyed from the speaker's mouth through a microphone to an analog/7' digital converter, where they are filtered and digitized together with, for example, any background noise that may be present. The digitized signal is then further filtered and converted into recognition parameters, in which form it is compared with a stored speech template to select the most likely of the words spoken. Yet another example of such a method is the IEEE Standard (published in April 1977, Vo124.I64).
). Please refer to the article "Performing Speech Recognition" by T. Walch (pp. 55-57).

As this paper shows, the application of speech recognition systems is constantly expanding, and as the paper points out, many models are already available for use in various applications. The formation of templates is also well known in the prior art. Such templates are used in many different types of speech recognition systems.

An example of a system is a "keyword recognition system" in the paper "An Efficient Elastic Template Method for Determining Given Words in Continuous Speech" by Jay Nis Pridle (J.S. Br1dle).
(April 1973, 'Spring Conference of the British Phonetic Society'
, pages 1 to 4). In this paper, the author discusses searching by extracting elastic templates from a 4-panel display of utterance examples of keywords. Similar parameterizations of incoming speech are successively compared with these templates, and the speech and templates are derived to estimate the similarity between keywords.

If the incoming speech segment closely approximates the corresponding template, the word is determined to have been spoken by the recognizer.

Word templates are termed "elastic" because they can be expanded and compressed in time to account for changes in speaking rate and rate of pronunciation of words.

Keyword recognition is similar to traditional speech recognition. In the former, the template is only memorized for arbitrary words, i.e. "key" words that are to be recognized within the context of the sound, whereas in the latter the template is memorized for all expected speech. The template is stored.

This kind of system is all about keywords! ! ! ! Whether it is a recognition system or a traditional speech recognition system using templates, the system is capable of solving the same problem, i.e., recognizing words uttered by different individuals or under different conditions by the same individual. I run into the problem of not doing it.

It is therefore an object of the present invention to provide a ninth improved apparatus and method for automatic speech recognition systems.

It is further an object of the present invention to provide a speech recognition system that automatically adapts to noisy environments.

[Ninth means of problem solving]

As can be seen from the appended claims, many speech recognition systems have reduced performance in noisy conditions.

This is particularly the case when the template is derived from speech with little or no noise, or where noise of a different nature is present at the time the recognition is performed. Conventional methods, which reduce the difficulty, require generating a new noisy new template. This generation requires new speech and noise collection. In the system of the present invention, analytical noise is added to the template, which improves the probability of recognition and substantially increases the performance of the system, without the need to collect new speech to generate the template.

The system of the present invention provides as an output the magnitude of the utterance's spectral value and provides an output when a stored template is compared with the processed spectral value and a good comparison is obtained indicating the presence of speech in said utterance. a first means for coupling to the spectrum analyzer and providing a signal indicative of an expected noise signal of the incoming signal, comprising a spectrum analyzer and a third device for generating the stored template; , means for generating, in combination with the first means, a template' in response to the expected noise signal, which is modulated according to the expected noise signal.

〔Example〕

As shown in the figures, the invention applies to all recognition systems that use real spectral lines or i4 parameters derived from spectral lines. The latter requires storing templates in two forms: noise analysis addition ninth spectral line and motion template.

Referring to FIG. 1A, a ten-step diagram of a speech recognition system using a spectrum-derived recognition meter according to the present invention is shown.

A microphone 10 is shown and is used by a person using the system to input speech. Microphone 10 converts audio waves into electrical signals, which are amplified by amplifier 11 .

The output of amplifier 11 is coupled to spectrum analyzer 12. Spectrum analyzer 12 is a wideband or narrowband spectrum analyzer with short-term analysis capabilities. The functions and configuration of spectrum analyzers are basically well known;
It can be configured in a number of ways.

The spectrum analyzer 12 divides the speech sound into b frames and outputs at its output an indication of the parameter variations of each frame. The particular type of audio analysis performed by spectrum analyzer 12 is not critical to the invention, and many known audio or spectrum analyzers can be used. An example of such is U.S. Patent Application No. 439,018 (1982).
Filed on November 3, 2015, C. Bensko et al.) and No. 4734
No. 22 (filed on March 9, 1983, No. Bensko, etc.)
It is stated in the specification. Both applications are assigned to IT Corporation, the assignee of the present invention, and are incorporated herein by reference.

Also referenced is US Patent Application No. 655,958 (filed September 28, 1984, inventor N. L. Higgins et al., entitled Friend Keyword Recognition System and Method # using a One-Template Sequence Model).

The spectrum analyzer 12 is equipped with a 14 channel band i4 filter array and the frame size used is 20 milliseconds or more. These spectral parameters are processed as shown in Figure 1A. As shown, spectrum analyzer 1
The output of 2 is coupled to switch 13, which can operate in recognition, 7 ohm template, or modulation template modes.

When the switch 13 is placed in the 7 ohm template mode, the output of the spectrum analyzer 12 will be in the 7 ohm template mode.
14. module 1
The purpose of 4 is to assist in the formation of a template from the output of spectrum analyzer 12. These templates are formed in module 14 and are in the form of template spectral lines, and many methods of forming such templates are well known. Basically, in the form template mode, the output of the spectrum analyzer 12 is processed by the monologue 14 to provide a template for the mental utterances made by the caller through the 10ft microphone. The caller speaks the words to be recognized and a template is generated that basically shows the words spoken. These templates are used by module 15 to derive t! from the spectrum generation templates. (recognizing the parameters) and generating the final template with low or no noise as indicated by module 16.

The clean h-state template is then stored, as directed by module 16, to represent a particular utterance, such as a word, phrase, etc. uttered by a particular speaker.

The stored templates are coupled by switch 100 to processor 160 where recognition algorithms are executed. Therefore, the processor 160 operates in recognition mode to generate the unknown speech in a noise-free manner and the module 1
Compare with the template stored in step 6. The first provision
As shown in figure A, in the form template mode, a vector format template is given and the template L/-y! parameter is obtained, and this template parameter is then used to form the template in no-noise or low-noise conditions. As will be explained below, processor 160 can operate with templates stored in module 16/C in a low or no-noise manner.

The functionality of processor 160 is well known and operates essentially to match based on various distance measurements and other algorithms. Once such a match is made, an indication is given that this is the correct word and that this word or sound will be the output of the system.

Switch 13, when placed in recognition mode, couples the output of spectral analysis a1z to derivation/IP parameter module 16, which module 161f-! 9 parameters are stored in the module 16 and compared to the stored template, for example as described above.

Switch 13 can also be set to the central position, as shown in FIG. 1A. The center position is the modulation template mode position and the output of the spectrum analyzer 12 enters the estimate noise statistics module 162. The function of module 162 is essentially to perform noise analysis or process the noise and estimate noise statistics.

This is a key feature of the present invention, and allows the present invention to selectively add templates to miscellaneous t1. form, perform speech recognition, and make improvements in such recognition in the presence of such complications.

Accordingly, the functionality of the estimate noise statistics module 162, which will be further described below, is combined with the module 14 to -
and modulating the spectral template formed in module 164 that receives information from the module. An output module 165 of module 164 derives recognition parameters that are used to form a template for use in noisy or low noise conditions as indicated by module 166. To this end, recognition is performed by the system shown in FIG. 1A by toggling switch 100 between a noisy template, very low level noise, or a clean template.

As briefly described above, in the recognition mode, the spectral parameter output of the spectral analyzer 12 is provided to the input of the processor 160 by a derivation parameter module 161. Processor 160 typically executes algorithms, but this is also not critical to the invention. processor 1
60 determines a sequence of stored templates;
Make the best match so that the incoming speech can be recognized. The output of the processor is therefore essentially a series of template labels, each label representing one template in the best matching template sequence.

For example, each template is assigned a number and label. This number may be displayed as a multi-pit. This output is provided to a template search system provided in processor J60, which becomes a comparator with storage for template labels, for example, if there is a multi-bit representation. Processor 1-60 is therefore operative to compare each incoming template label to a stored template. Subsystem processor 160 then provides an indication that a particular word or phrase was uttered as well as the word or phrase itself.

In either the 7 ohm template mode or the modulated template mode, the user speaks and a recognition meter is derived from the spectral output of the spectrum analyzer 12. In the modulated template mode, the system generates various templates for use in conjunction with the system in the recognition mode, which templates are modulated by selective addition of noise by the estimate noise statistics module 162 as described above. Ru. This selective addition of noise by module 162 provides more reliable system operation, as will be discussed further below.

Referring to FIG. 1B in its entirety, a recognition system using natural spectra as recognition parameters is shown. In each case, parts of the same function are designated in FIG. 1B by the same reference numerals as in FIG. 1A. As can be seen, microphone IQ is coupled to the input button of amplifier 11, and the output of amplifier 11 is coupled to the input of spectrum analyzer 12. The output of vector analyzer 12 is again coupled to switch 13, which is capable of operating in form template, modulation template, or recognition mode.

As seen in FIG. 1B, in the 7 ohm template mode, the module 170 forms the template with low to no noise.

This modulus 170 directly gives 9 parameters, which are the spectrum of nature. This 7 ohm template is then stored and coupled to a module 171 which converts the spectral template into 4 of the effects of the estimated noise statistics generator 122 which essentially functions as a noise module 162, e.g. - modulate the spectral template derived from the rule 120; The output of modulation spectral template module 171 is coupled to module 173, which stores the template for use in noise conditions. Processor 177 is also shown in this figure and operates on either templates stored in module 17θ or templates stored in module 173.

In any case, prior to further processing, it is known how to generate templates according to the prior art. There are several ways to generate templates. The method of performing the template rawhide operation is automatic and usually uses a multi-step or two-step process. In one such method, speech data from training utterances (template mode) is divided into segments. These segments are then provided as input to a statistical cluster analysis, which selects the salset of segments that maximizes a mathematical function based on distance measurements between the segments.

Segments belonging to the selected Saraset are used as templates.

Such techniques are described in the above-mentioned US patent application Ser. No. 655,958. In any case, various methods of measuring distance are known and described in some of the references cited in the Background of the Invention. A widely known method for measuring distance is Mahalanobis distance calculation.

An example of this method is given in U.S. Patent No. H003971 (
The invention is entitled "Multiple Nine-Layer Speaker Recognition System and Method," January 16, 1987, assigned to Wrench et al.). This specification provides examples of various other techniques used in caller recognition systems, including:
Some of the algorithms used in this system are described in detail. In any case, referring to FIG. 1, the main features of the present invention relate to the speech recognition system shown in FIG. Determine if the word was spoken. This method is suitable for keyword recognition systems, speech recognition systems, speaker recognition systems, speaker verification systems, language knowledge systems, or any system that uses templates or a combination of templates to make decisions about spoken sounds. It can also be used in such systems.

Before explaining the structure and method of the present invention, the principle and idea of the invention will be explained.

The inventors have recognized that when the signal-to-noise ratio of the template is the same as that of unknown or spoken speech, recognition performance is better than using noisy or noisy templates. My friend. Therefore, if the signal-to-noise ratio of the audio signal is considered predictable, the template is 1-modulation can optimize recognition performance.

Therefore, in order to put the present invention into practice, the following considerations must be made. The first is to predict the S/N ratio of the incoming speech, and the second is to modulate the template to be # as if it were 1.

Predictions are based on both theory and experience. Often low level or absolute in the case of constant noise;
One can expect the speaker to speak at a relatively constant level, either at a relatively constant level that is greater than this noise. The speech and noise levels can then be used to predict the signal-to-noise ratio of unknown speech. This is done by using a speech and noise level tracker module, as explained below.

At a certain distance, both the speech level and the noise level of each filter channel change slowly enough so that the current value is a valid estimate of the near future value.

Modulating a noisy or relatively noisy template to make it "as if" it were created from noisier speech is based on both empirical and theoretical considerations.

After research, it was determined that the noise and speech power added in each individual filterbank channel was a good approximation. A more accurate approximation has a speech-like Chiheiman distribution with a large number of degrees of freedom regarding the filter bank channel band. From the above considerations, a more accurate estimate of the expected value of a known speech power combination with noise of known statistical properties can be made. This increased accuracy in “adding noise” also increases the accuracy of the generated template, but ′
It does not significantly increase recognition accuracy beyond that obtained using the "power addition" rule.

The following describes the Nokwar addition rule, although the process could theoretically be made more accurate by substituting another method of estimating the expected value of the speech and noise power combination. This substitution does not alter the spirit or substance of the invention.

Furthermore, it is observed that both internal electronic noise and quantum noise combine with acoustic noise and signals with respect to the "Noyer addition" rule. Although these noises are smaller than the associated acoustic noises, the application is possible. Accordingly, we can use the results of the "Noyer addition" per K that constitute the various models, thereby clarifying the research work in a continuing effort, and use the numbers derived from the valid model. This will be explained below.

It has been shown that the template resulting from the noisy software is equal to its average value and performs very well with respect to producing reliable validation outputs. It is therefore not necessary to anticipate the frame-to-frame variability of the noise, and it is sufficient to use the average value. are the parameters generated from the same speech/4' word that is efficient in .

The channel noise power value from the system is an estimate of the noise power and can be taken in relation to an average noise power that can be determined mathematically. Therefore, in order to fully understand the process and its validity, it is explained below.

It is first pointed out that the probability distribution of the output of a single separated Fourier transform (DFT) of a speech signal whose additive zero mean has been degraded by the Woos noise can be easily calculated. The next important factor to consider in extending the model of how speech and noise combine to allow each channel of a bandpass filter bank to be applied is that the bandpass of the channel is
This means that it is or can be much larger than the T channel. Therefore, the noise matrix and the number of channels that make up the channel can be estimated by observing the output of the band t4 filter in the absence of speech and in the presence of noise.

The next step is to improve the speech recognition template formed in the absence of noise for use in noisy conditions by modulating it to be equal to the expected value in noisy conditions. . The method used therefore allows each speech sample and bandpass filter channel implemented in the clean template to substitute an expected value of the clean template that is modulated by the presence of current noise.

Therefore, by measuring the average and variation in the output of a pantone or filter channel, the characteristics of the channel as it passes through Gaussian noise can be estimated. Fundamentally, as can be seen from the above (and most of the above points have also been mathematically proven), there is both a theoretical and an empirical basis for carrying out the invention. Basically, the characteristics of the present invention are that the noise is analytically added to the template formation, and the formed template? The operation increases the credibility of the voice recognition system.

There are two ways to add noise to a friend template data collected in a noisy environment, thereby creating a new template data for use in a noisy environment. The exact method adds noise to each tensile node and then averages the results.

An approximate method is to form the base form data by averaging the clean tokens, and then add the appropriate noise to the current state using "Noyer addition" or some other convenient or more accurate rule. -Modulate the evening. Exact methods require maintaining all templates and surrounding tokens, and require excessive memory. Approximate methods give essentially the same template and recognition results. There are main ideas in the implementation. This means that the template data is noise-free with respect to the environment in which it is used.

Referring to FIG. 2, a detailed block diagram of the template formation used by adding noise to a pace 7 ohm template is shown. A paceform template is itself an average formed by a set of words called "tokens." Each token consists of nine parameters taken from one pronunciation of a given word. 1
One or more tokens are arranged to form a paceform template. The base 7 ohm template is silently formed and stored either in the module shown in FIG. 1A or in the module 170 shown in FIG. 1B. FIG. 3 is a table defining the values shown in FIG. 2.

FIG. 2 again shows the microphone 10 into which the speaker speaks. The output of the microphone is coupled to the input of amplifier 11, the output of amplifier 11 a B
It is coupled to a spectrum analyzer 12, illustrated as a PF, or bandpass filter. switch 13
is at the modulation template position. Spectrum analyzer 12
The output from the pantone filter is a vector of values of magnitude 4, which is applied to module 2o, which averages the frame pairs.

Frame pair averaging is a well-known technique and is essentially performed by many intelligent circuits.

The output of module 20 is the result of averaging successive pairs of inputs from spectrum analyzer 12, and module 20 halves the effective frame rate.

The output of module 20 is scale pit module 21
and the squared component module 22.

The squared component module 22 provides a vector output that is essentially equal to the squared value of the average frame versus the output of the module 20.

The output of the scale pit module 21 essentially serves to double the average of successive pairs performed by successive shifts, making it possible to adapt the vector maximum component to a 7-bit scale. To this end, module 2 includes a shift register, which essentially performs a number of right shifts to perform the operations described. The output from the scale pit module 21 is directed to a logarithmic converter 23 which produces a scaled log spectral parameter vector at its output. This parameter vector is then carefully averaged by module 24 over a given set of template tokens to provide at output a scaled log spectral parameter that essentially provides one parameter of the paceform template. The output from the square component module 22 is combined with the input of the relative energy module 25 and the speech and noise level tracker λ.

The output of the relative energy module 25 is a square meter indicating the relative energy determined, for example, by averaging the energy from the output of the square component module 22. This is averaged over the template talk/s by module 36 to provide an average indication of the output vector, which is the relative energy parameter needed to provide another paceform data value. The output from the speech and noise level tracker 26 indicates the energy level, which is averaged again by the module 27, as described below.

- provides energy levels of yet another paceform characteristic at the output of the module. The speech and noise level tracker provides two additional outputs, as further described, one of which is a logarithmic indication of speech level averaged over word time and channel. This is the counting circuit attached to the. The other is a vector of noise levels in each channel, averaged over time, but not related to the channel.

This is also a vector attached to the word recognition unit. The output from module 27 is therefore provided to a first adder module 30, which is shown receiving additional output from the speech and noise level tracker. The output of adder 30 is applied to one of the inputs of adder 31, which receives at its other input the output derived from scale bit module 21. The output of scale bit module 21 is multiplied by a factor K via module 32, and K
is shown in FIG. 3 as being equal to 18,172.

This value is then averaged by a module 33 to produce a logarithmic paceform value which is applied at its output to the other input of adder 31. The output of adder 31 is given to adder 32. Summer 32 receives as another input the output from speech and noise level tracker 26, which is again a vector of noise levels in each channel. This output is applied to one input of functional module 4Q, and module 40 receives the output from module 23 at its other input. The output from function module 40 is a scaled log spectral parameter vector of the noisy template. This is fed to a functional module 41 which provides at its output a recognition parameter vector which is the Meru cosine transformation matrix of the particular utterance. Accordingly, the output from the module 41 and the output from the tracker module 26 are used to generate the motion template data file 4.

As mentioned above, the outputs related to the block diagram in Figure 2 are
As shown in the figure. As can be seen from FIG. 3, the effective spectral magnitude of the pace 7 ohm template derived from FIG. 2 is essentially given by the following equation.

m = 2 ・XPb(t) The effective/yower is given by the following equation:

P = storage Lm = 22111@XPb (2211) Please refer to Figure 3 for the definition.

Before adding noise, the noise of each frame is modulated so that the average speech level of the template shown at the output of module 22 in FIG. This is the same as the current speech level indicated by the output of 26. Since its value is within the recognition unit (9,331 dB), the effective power of the paceform changes and the tracker 2
Indicated by the output of 6. In this regard, the current noise level is added to obtain the effective noise level of the noisy template, and the effective magnitude of the noisy template can be indicated at the output of module 41.

All motion recognition parameters are therefore Melukosine transforms of the logarithmic spectral parameters, which are measures of relative energy. The above will be clear and mathematically clear to those skilled in the art when looking at the entirety of FIG. 2 together with the definitions in FIG. 3.

Thus, by using the same exact technique, a template can be formed by adding noise to the template tokens and then averaging. Essentially the process for doing this is the same as shown in FIG. 2, thereby producing the same exact output as shown in FIG. 2, except that the averaging is done after functional unit 40. Given.

In FIG. 4, a detailed block diagram of nine conventional systems using template formation techniques as described above is shown.

In Figure 84, the same reference numbers are used to indicate parts with the same function. Sum, kneader/decoder (C0DIC) -v-Joule and linear circuit 47 as seen in Figure 4
An AGO 1 or automatic gain control module 45 is arranged coupled to one input of the adder 46, with the output of the adder coupled to the adder output. A coder/decoder module is essentially an analog/digital converter, followed by a digital/analog converter. The output of the Kobukku is provided to a synthesizer or pantone filter bank or spectrum analyzer 12.

The output from the spectrum analyzer 12 is sent to an average frame pair module 20, which is associated with a scale module 21 and a speech and noise track tracker 26, again referred to as ff1K.

The output lines shown on the right side of FIG. 4 provide various motion template data values that are used to form the noisy template.

The main functional module is the speech noise tracker 26, which will be discussed further below. Also, in FIG. 4, the inputs to the microphone 10 are marked with symbols Na and Sc, which are important signals and miscellaneous t@.

The subscript C'' indicates that these expressions represent the average spectral magnitude over the passband of each of the filter bank channels forming the spectrum analyzer 12. has the value of
Each value represents each filter in the filter bank.

Therefore, Sc is the spectral magnitude of channel C of the audio speech signal, and Na is the spectral magnitude of the acoustic noise of this channel.
It has a root mean square magnitude. The output from adders 50 and 46 is the electronic noise magnitude, which is A
The front of the GC gain control 45 is injected after. C0DE
The output from C47 includes the magnitude of the quantization noise introduced by C0DEC. , and the output of the average frame pair module 20 is the result of averaging successive pairs of vector magnitude values.

The effective output signal of vector analysis B 12 is an estimate of the spectral magnitude of the signal at the filter bank input across the passband of the filter puncture, which is directed to each channel in the filter bank. Successive pairs of these values are averaged to produce output from module 20 at a rate of 507 seconds.

Essentially all sets of values for the 14 channels are all shifted to the right by the same number S in module 21, thereby occupying a maximum of 7 bits or less, and the resulting values are logarithmically proportional by the header table. converted to a number. The table can be thought of back to input 127 where the result is 26.2 times the natural logarithm of the input, or the logarithm to base b (b is 1.03888). The 20 millisecond frame values are also used by tracker 26 to generate a measure of peak speech energy and an estimate of the average noise energy for each channel.

The speech level is an estimate of the logarithm of the overall speech energy at the microphone 10 with respect to the base b, which can be any constant.

The effect of AGC gain is not essentially a spectral value that is removed. For example, this gain is a complete view of the passband energy of the entire filter bank.

The utterance level estimate is also the relevant word or phrase, and its time constant is the magnitude of the level at which the short utterance is made. Therefore, there is only one level value associated with each template or unknown segment of the template period.
There is only one. The time of noise estimates from tracker 26 is strictly constrained such that there must be only one noise level estimate assigned to each channel over the length of time that is being spoken. The output value from the speech and noise tracker 26 coupled to the logarithm circuit 54 of FIG. 4 is therefore an average energy estimate of the output of the filter bank. These values are therefore influenced by the AGC gain and are directly proportional to the average spectral energy without logarithmic transformation.

The signal and the various noise sources are statistically independent and their energies are calculated on average. This is not only convenient for determining internal noise sources, but has also proven to be a good approximation of both acoustic noise and signal sources.

Furthermore, there is considered to be a noise value that may be the equivalent noise/yew in the microphone. These values include acoustic noise and other system noise;
A portion is reduced by the gain of AGC45.

Therefore, by using the template averaging process, the scale factors derived from FIG. 4 and shown in FIGS. The same idea can be generated by averaging the logarithmic spectral parameters of all tokens in the S/N ratio and S/N ratio. therefore,
To simplify the overall problem, consider that all templates as well as all template tokens have the same signal-to-noise ratio. This can be done by adjusting the speech level of all tokens to be the same, so that the same S/N ratio will be the same as the noise value of all tokens.

Based on this idea, all forms that average the equivalent values of noise can be created.

Research has shown that when the S/N ratio of the template is the same as that of the unknown speech, as shown above, the recognition performance is better than when the noise is thicker or smaller. There is. Therefore, based on the above technique, we can predict the S/N of the audio signal so that the template will appear as if it were generated from speech with the same S/N ratio as the incoming unknown speech. It is shown that the template can be modulated to optimize recognition performance before it is used.

Two steps are therefore used. One is to anticipate the signal-to-noise ratio of the incoming speech and modulate the template to suit this requirement.

Therefore, as explained below, the speech and noise tracker 26 does not form an estimate of the speech power in each channel because the speech power in each channel varies from word to word depending on the speech content of each channel. Since it is impossible to predict what words will be spoken, the data has no predictive power. What is important is that in a normal process, there is no estimate of the S/N ratio for each channel. Therefore, as mentioned above, the template modulation process does not use a specific S/N ratio for each channel. A template resulting from the noise power equal to its average value performs very well in recognition systems.

That is, there is no need to consider the frame-to-frame variability since it is sufficient to use the mean value of the noise. And the template parameters are 6 base forms 1 combined with the current average noise power
It is generated from the same speech power that is effectively present in the template. Basically, as described above, the speech and noise tracker 26 is a digital signal processing (DSP) circuit that provides a measure of the power level of the speech signal in the presence of additive acoustic noise and the average of any form of bandpass filter bank channel. The algorithm operates to give a measure of noise power. The utterance level measure found indicates the speaker's conversation level that is appropriate for adjusting the S/N ratio for Seido recognition. Other measures of speech level involve rapid changes and/or relative frequencies of vocal and non-speech sound occurrences within spoken speech. Measures found by speech and noise trackers avoid these problems by detecting slightly smooth peaks in vowel nuclei.

More specifically, it detects the slightly smoother Beekz vowel within the more energetic vowel nucleus. Continue to indicate general speech levels by ignoring unstressed consonants and warp peaks in speech intervals that are not vowel nuclei. Trackers are used in the presence of additive noise that is unrelated to the existing speech, where the total amount of noise power typically changes slowly compared to the vowel nucleation rate within the speech (typically 5 to 157 seconds). The tracker operates to recover from faster changes in noise level. Speech and noise tracker 26 uses a logarithmic, or compression technique, which provides a measure of the total amount of speech power in the frequency domain of interest. This measure is initially filtered to rise slowly and then fall quickly; the rise and fall time limits in this case are the instantaneous signal power between the vowel nuclei in the first few milliseconds and the difference between the filter value and are selected such that there are large positive differences in the values and no large negative differences in values.

Therefore, the nonlinear function of the difference between the instantaneous signal value and the rapidly falling and slowly rising filter value is then directed into a moving boxcar integration step of appropriate duration, and the resulting value is It rises above the appropriate threshold only in normal or stressed vowel nuclei in speech intervals, and unstressed vowel nuclei become skilagged. Crossing this threshold is used to identify intervals of high signal noise due to speech kernels. Therefore, only the intervals that are identified are used for speaking level tracking. and a second smaller than the speech kernel threshold.
greater than the threshold of ? The values from the Tsuksker integration process are used to identify intervals that retain speech power and noise power.

Only those intervals where the boxcar integral value is less than the second (lower) threshold and the instantaneous power is not greater than the third threshold, which is greater than the fast-down, slow-up filter value, Used as input for noise power tracking function.

Miscellaneous J? The word tracking module includes a digital signal processor that is basically constructed by an integrated circuit chip. Many such chips are fundamentally programmable and can be configured to perform various types of algorithms. The Alprism associated with the noise and signal tracking function operates to determine both signal and noise energy content and operates in the following manner.

First, obtain a numerical value indicating the channel energy. This is done every frame. The total energy is then calculated. The system can then operate to adapt to automatic gain control changes. Once the energy has been calculated, the result is then smoothed over a given period. After t the smoothed energy value is obtained, the logarithm of the total amount of energy is calculated. After calculating the logarithm of the entire energy, Pantono! Perform a boxcar integral or average of the speech level estimates on the input to the filter array. In the next step, an asymmetric filter is used to filter the log energy of the speech detection by monitoring the rise time of the speech signal. The speech signal is comprehensively interpreted, and the incoming signal may be noise, or it may be noise or a processed signal that is not a speech signal, such as heavy exhalation or basically noise rather than information. It is due to some other characteristics of the speaker's voice. This is also a true speech signal in any case.

Therefore, to determine this, the instantaneous value of the logarithmic energy in the smoothed energy is monitored. The algorithm operates to divide the time interval associated with the rise and fall times of the signal into nine given intervals. When the rise is positive compared to negative, certain decisions are made to recognize the characteristics of the incoming signal. These decision traps determine whether it is speech, processing, or pure noise as described above. For example, during a period when the rise is negative, if the rise is continuously negative, it is perfectly considered to be a noise signal. As the noise signal is received, the system continuously tracks the signal by smoothing the noise values, contributing these values to the average noise energy, and using the calculated value to apply this value to the noise estimate. . Next, wrap it all around to form a template. Caution regarding positive transitions is even more difficult.

Positive transitions are noise, processing, or speech. For this determination, we perform and operate on the integral of a nonlinear function. Based on comparing the integral value with a fixed threshold value, it can therefore be determined whether a positive increase represents speech, noise or processing.

In this way, the resulting ratio values from the speech and noise tracker module represent true speech values. The speech and noise tracker program is shown in FIGS. 5A-5C, where the complete program is shown.

FIG. 6 defines the engineering parameters necessary to understand the Zorogramming format shown in FIGS. 5A-5C. To explain further, this process is performed for each single frame and operates as follows. Fifth
In the first step of the method shown in Figure A, the energy in each channel is obtained together with the total energy. This is shown in steps 1 and 2. The energy is then filtered in each channel with automatic gain control scale changes as shown in the third and fourth steps. The next step is to smooth the energy values to obtain a smooth logarithmic value of the energy that is corrected for AGC.

This is shown in steps 5, 6.7. In the next step, in step 8, the speech level estimate My
Take the Cusker average. We then obtain the asymmetric filtered value of the energy and obtain the current energy rise in the filtered value shown in steps 9 and 10.

The program then moves to Figure 5B. The variable r, shown in step 10 of FIG. 5A, is the amount by which the current logarithmic energy exceeds its asymmetric smooth value. The period r of the vowel nucleus is positive and remains positive over the period of the interval kana υ.

This is to give particular prominence to its positive and negative periods,
Special handling is required when it first becomes positive or negative. This is shown in detail in Figure 5B. r
The first time that becomes positive, we record the frame number as the possible beginning of all distinct speech kernels. It then resets the value Pf used to determine whether it is speech and interrupts noise tracking. In any case, while rt1 remains positive, the value py is accumulated and P
Processing and speech flags are all set to determine if the value exceeds a certain threshold. These are shown on the left side of Figure 5B. When r first becomes positive, it resets the noise tracker to the last known noise value and resumes noise tracking after a given delay as to whether speech or manipulation was detected, while the approximated speech level is sufficiently high above the noise level.

If speech is detected during this rise, a frame is recorded in terms of number as the end of a known speech interval.

Continue to track the noise after a predetermined delay while r remains negative. This is all given in an illustrated flowchart that clearly describes the various operations.

FIG. 5C essentially illustrates the generation of output variables used to provide the motion templates shown in FIGS. 2 and 4, for example. Therefore, as can be seen from the above, the main idea of the system of the present invention is to provide a template template, thereby adding noise in the correct and expected manner to create a template with an associated expected signal-to-noise ratio.
form.

The noise level associated with the template represents an estimate of the noise present in the incoming signal. In this way you basically increase the recognition potential of the speech recognition system.
- Generating such a template by adding noise as described above means that the entire template can be compared with the incoming signal to determine whether the signal is actually speech, processing, or noise. It can be used in any speech recognition system that determines the Therefore, by modulating to be equal to these expected values in noisy conditions, which are first formed in the absence of noise, the system provides improved speech recognition templates for use in noisy conditions. works.

[Brief explanation of the drawing]

FIG. 1A is a block diagram illustrating a speech recognition system using a spectrum-derived recognition parameter according to the present invention. FIG. 1B is a block diagram illustrating another speech recognition system using recognition parameters that are spectral in nature in accordance with the present invention. FIG. 2 is a detailed block diagram illustrating a technique according to the present invention for forming motion template data. FIG. 3 is a table of definitions of the various outputs shown in FIG. 2. FIG. 4 is a detailed block diagram of another embodiment of the invention. Figures 5A-5C are detailed flowcharts illustrating the operation of the speech and noise tracker according to the present invention. FIG. 6 is a diagram showing a table of definitions of engineering parameters according to FIGS. 5A to 5C. 10...Microphone, 11...Amplifier, 12...
・Spectrum analyzer, 13.100...switch, 1
4゜15.16.20,21,25,27,40゜16
2.166...Module, 26...Tracker,
160...Processor, 31.32...Adder, 5
4... Logarithmic circuit. Applicant's agent Patent attorney Suzue TakehikoFig, 2゜tFNT, P81 Fi9 3 n, 5 ■

Claims (21)

[Claims] (1) a spectral analyzer which provides as an output the spectral magnitude of an utterance and which compares a stored template with the processed spectral values and provides an output if a good comparison is obtained indicating the presence of speech in said utterance; a speech recognition system, comprising: a device for generating said stored template; and means for generating, in combination with the first means, a template that is modulated in accordance with the expected noise signal in response to the expected noise signal. system. (2) the first means includes speech and noise level tracking means operative to provide at an output a first signal indicative of the power level of the noisy speech signal and a second signal indicative of the average noise power; A speech recognition system according to claim 1, comprising: a voice recognition system according to claim 1; (3) The spectrum analyzer comprises a plurality of bandpass filters arranged in a filter bank array, each filter configured to pass a predetermined spectral component according to the band of the filter. The speech recognition system according to item 1. (4) The speech recognition system according to claim 1, wherein the second means comprises means for generating a template under low noise conditions and modulating the template in accordance with the expected noise signal. . (5) The first means includes an S of the incoming speech signal.
2. The speech recognition system according to claim 1, further comprising means for predicting the /N ratio. (6) The speech recognition according to claim 3, wherein the first means comprises means for measuring the average and change of the bandpass filter to provide an estimated value of the noise passing characteristic of each filter. system. (7) The speech recognition system according to claim 6, wherein the estimation of the noise is performed based on the filter that responds to Gaussian noise. (8) the template generated in the absence of noise is a noise-free token template, means for providing an average value for providing base form data in an output in response to the template; 5. A speech recognition system as claimed in claim 4, comprising means for modulating the form data in accordance with a currently expected noise signal. (9) a spectral analysis which provides as an output the spectral magnitude of the utterance and provides an output if a predetermined stored template is compared with the processed spectral values and a good comparison is obtained indicating the presence of speech in said utterance; a speech recognition system comprising a device for generating the stored template, the system comprising: a device for generating the stored template; A speech recognition system comprising: processing means for generating a template for; and means for comparing the generated template with an incoming signal to provide the output. (10) The speech recognition system according to claim 9, wherein the predicted calculated value of the processing means indicates the presence of Gaussian noise. (11) The processing means comprises means for averaging the noise-free template to provide a base form data output and for modulating this base form data output by adding this data, i.e. the calculated noise data. A speech recognition system according to claim 9. (12) said processing means comprises averaging means for providing at an output an average value of successive pairs of said spectral magnitude values as provided by said analyzer, coupled to the output of said averaging means; and means for converting said field signal of given length into a logarithmic signal to provide said base form data output. 9. The speech recognition system according to item 9. (13) squaring means coupled to said averaging means for providing at an output a vector signal indicative of the magnitude of the square of successive pairs of said mean values; 13. The speech recognition system of claim 12, further comprising means for providing another output of the form data output. (14) The means coupled to the output of the squaring means:
relative energy shaping means for providing a base form energy parameter in response to the vector signal;
14. The speech recognition system of claim 13, further comprising speech and noise level tracking means for providing base form parameters indicative of both speech and noise power levels at the output. (15) A method for forming a template used in a speech recognition system, the method comprising: providing a signal indicating an expected noise level of an incoming signal; and modulating a given template according to the given signal to generate the expected noise level. 1. A method comprising: providing a template having a noise level of (16) The step of providing a signal indicative of the expected noise level comprises the step of measuring the response of a given speech processing channel with respect to noise and estimating the signal based on this measurement. The method described in section. 17. The step of modulating comprises first forming a relatively noise-free base form template and modulating the base form template in accordance with the signal indicative of the expected noise level. the method of. (18) A patent comprising, in the modulation step, forming relatively noise-free base form templates, adding noise to each template, and averaging the added noise template data to form a new template according to the analysis data. The method according to claim 15. (19) The step of providing the signal is to reduce the S/N ratio of the incoming signal to the existing signal by averaging the log spectral parameters of all templates at the same speech level and S/N ratio. 16. The method of claim 15, comprising the steps of: predicting as perceived by a modulation of the power of the power; and forming a modulation template using the average parameter. (20) A method of forming templates for use in a speech recognition system, comprising: modulating the formed templates by adding a noise signal to said templates representing expected values before they are used for comparison; A method characterized in that the template operates as if it were generated from a speech signal that maintains the same signal-to-noise ratio as the incoming signal to be recognized. (21) The step of modulating predicts the S/N ratio of the incoming speech signal by using the current S/N ratio as the expected value based on the current speech level, and the step of modulating the current noise power and speech power. Claim 2 comprising the step of limiting said additive noise signal by averaging.
The method described in item 0.