JPS61273600A - Voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Application of the Invention] The present invention relates to a speech recognition device that targets monosyllables constituting a speech signal.

[Background of the invention]

Conventionally, for audio signals that were uttered by dividing each -word,
In the so-called monosyllabic recognition method, which recognizes each monosyllable individually, an envelope signal is formed from the speech signal of each word, the rising region of the envelope signal is used as the region from which consonant information is obtained, and the envelope signal is The area from the time when the maximum amplitude value is obtained is defined as the area from which vowel information is obtained.
The spectral shape of the original audio signal in each region,
Alternatively, a method of determining changes in the shape of the spectrum has been used. Then, the degree of similarity with standard characteristic values or patterns set in advance is calculated, the closest monosyllable is selected, and this is identified as the monosyllable of the input speech.

There are five types of vowels in Japanese: ``aiueo.'' When a vowel itself is uttered without a consonant, the spectral shape of each vowel differs greatly, so it is easy to recognize the vowel based on the spectral shape. However, monosyllables with consonants, % with a voiced sound "zajizuzezo"
etc., the spectral shape of the vowel region is significantly deformed compared to the shape of the vowel itself, even though K is the same vowel. This is one of the factors that worsens the vowel recognition of the entire monosyllable.

In order to obtain as much of the same information as possible for the same vowel regarding the above-mentioned transformation of the vowel region spectrum, various calculation processing methods have been conventionally applied to the spectrum.

One of the representative methods is described in Technical Research Report of Institute of Electronics and Communication Engineers, EA-61-68. P 57-64 (198
In 5), as described in the paper ``Recognition of Japanese plosive consonants using local peaks of the spectrum'' by Kin, Makino, and Kido, a straight line determined by the least squares method is drawn on the spectrum, A method has been proposed in which this straight line is used as a threshold and the position of a local peak in the spectrum exceeding this value is determined. This method has the advantage that even when the slope of the entire spectrum changes in various ways, it is possible to easily obtain the position of the peak, that is, the frequency information at which the pseudopolmant appears.

Another method is the Electronic Technology Research Institute report, +41
841. Pi~122 (1984) As described in the document by Kuninaka [Basic research on automatic recognition of speech based on phonemic units] in K., a peak location with a predetermined frequency width is determined from the spectrum shape, A method has been proposed in which the pseudo formant frequency is determined by emphasizing the part. If the emphasis at the peak location is appropriately increased, there is an advantage that frequency information at which pseudo formants appear can be easily obtained even when the slope of the entire spectrum varies in the same way as described above.

However, all of the above methods require complex calculations based on spectral amplitude data, and therefore, from the viewpoint of shortening the recognition calculation time, the calculation time is considerable.

Moreover, within the range of so-called nine sounds, changes in the spectral shape between the same vowels are small, and the spectral shape itself is useful information for identifying monosyllables. Namely, pattern ma.

The recognition calculation based on checking works effectively. Therefore, the method of eliminating the spectral shape as described above causes a problem in that useful information is lost. Every time a person utters a word, the conditions under which it is uttered differ even if it is the same word, and the observed spectrum may show the characteristics of the word clearly, may be unclear, and may have clearer characteristics. The reality is that the frequency range that appears in the spectrum also changes from time to time. Therefore, in order to perform accurate recognition calculations, it is necessary to take detailed measures based on the actual situation.

[Purpose of the invention]

An object of the present invention is to eliminate the above-mentioned conventional problems by providing a signal processing circuit after an analysis circuit that analyzes the frequency of a speech signal, thereby obtaining information regarding a plurality of vowels and providing multifaceted information. An object of the present invention is to provide a speech recognition device that enables accurate recognition calculations.

[Summary of the invention]

In the recognition device of the present invention, a voice signal is first analyzed by a bandpass filter, and the output signal is converted into an envelope signal by an envelope forming circuit. The center frequency of the bandpass filter is swept from low to high frequencies within a predetermined period of time to form a signal representative of the frequency characteristics.

A high-pass filter is provided to remove the low frequency components of this signal, and the signal passed through this filter is used as the input signal spectrum for performing recognition calculations.

FIG. 2 is an example of a spectrum in the vowel region of a monosyllable. Here, 0etB represents the average level of the spectrum. Change the frequency range of the spectrum to the low range (25045
0ffz), midrange (6301250Hz).

High mid range (t25j 2.5 & ffz), high range (2,
By dividing the frequency range into 5kHz-5kHz and comparing their average levels, it is possible to characterize the five vowels of ``Aiueo'' to some extent. The low frequency defined here is the region where the first pseudo formant appears, the middle frequency is the region where the second pseudo formant appears, and the high frequency is where the higher order pseudo formant appears.

Since the formant frequency differs depending on the type of vowel, the differences between vowels can be characterized to some extent even from the above-mentioned relatively simple viewpoint.

In the monosyllabic spectrum, the average amplitude value within the above frequency range is determined, and the horizontal axis is the average amplitude in the low frequency range, and the vertical axis is the average amplitude in the middle frequency range. When this is expressed in a graph, FIG. 3 is obtained.

The two-dimensional plane in the figure is divided into 8 regions, but these regions are divided into 70 monosyllables (excluding diphthong monosyllables such as "kiya" and "kyu"). A large amount of data in the above 2
It is plotted on a dimensional plane and set as an area that satisfies all data. According to the diagram, “A”, “
oh".

It is divided into an area where there are only one vowel, such as ``e'', an area where there are multiple vowels, and an area where all vowels are present. − Samples existing in the region represented by vowels are
Samples that exist in the region represented by all vowels have distinct characteristics only in the relationship between the low and mid frequencies, but in contrast, samples that exist in the region represented by all vowels do not show distinct characteristics when treated simply as described above. Conceivable.

Next, regarding the region in which multiple vowels exist in Fig. 3 above, if the average amplitude in the middle range (6301250 Hz) and the average amplitude in the high middle range (t25 & -2,5 & Hz) K are expressed on a two-dimensional plane, Two types of graphs as shown in Figures 4 and 5 are obtained.

The areas belonging to these two types are as follows.

Type 1: "Aeo", "Aueo", "Aueo" Type 2: "Iue", "Aueo" The graph of Type 1 is, for example, in the area of "Aeo" in Figure 3. , is a graph obtained based on samples existing in this area, and as shown in Figure 4, there are "Ah", "Ao", and "Aeo".

It is divided into four areas: ``e''. If we replace "e" with the "ueo" area, we will get a graph that is common to areas belonging to type 1. A type 2 graph is similarly obtained as shown in FIG.

In the end, the area in which multiple vowels exist in Figure 3 is 2
Therefore, eight areas are aggregated into five types. However, by dividing the region into smaller parts, we gain an important clue as to the nature of the input monosyllable.

As mentioned above, by using a two-dimensional map with relatively simple variables, the vowels of samples that have clear characteristics with respect to the input monosyllable spectrum can be immediately identified.
Two to four candidate vowels can be set for samples that do not have such vowels. For this candidate sample, pattern matching is performed as a recognition calculation. Since the characteristics of the sample spectrum in each region above are relatively similar, if a reference p4 turn is created for each region in advance, Since mutual comparison can be performed using relatively detailed features, the accuracy of /paddle turn matching calculations is improved.

Next, in order to speed up the calculation using the above-mentioned map, in the present invention, while obtaining time-series data of the spectrum, a signal processing circuit centered on four band-pass filters connected in parallel is used. The time-series data is signal-processed.

A schematic block diagram of this signal processing system is shown in FIG. In the figure, a monosyllabic signal input through terminal 1 is input to a bandpass filter 2 whose center frequency is variable.

This bandpass filter 2 forms a time series signal by sweeping the center frequency from a low frequency range to a high frequency range. When the time axis required for one sweep of this signal is replaced with the frequency axis, the frequency characteristic becomes that of a monosyllabic signal. This signal is converted into an envelope signal by an envelope forming circuit 3, a low-frequency component is removed by a high-pass filter 4, and a signal component in a set frequency band is processed by a group of four band-pass filters 5 connected in parallel. Extract.

The signal passing through path 7 is the output signal of high-pass filter 4 itself. From the output signal of each bandpass filter, the average amplitude value in one scanning time is determined by the average value extraction circuit group 6Vc, and the multiplexer 8 sweeps these signals at predetermined time intervals. Φ converter 9
to form a digital signal. Here, as for the output signal KI$I of the envelope forming circuit 3, the sampling of the signal in the multiplexer 14 operates in accordance with the timing when the center frequency in the bandpass filter 2 is swept from the low frequency range to the high frequency limit frequency. The operating logic of the multiplexer 14 is configured as follows.

[Embodiments of the invention]

The overall configuration and operation of the device of the present invention will be explained below.
This will be explained with reference to FIG. The uttered audio signal is transmitted to microphone 1 in the figure. The signal is inputted via the microphone amplifier 10 to the envelope forming circuit 1] and the band i4x filter 2. When the amplitude of the audio envelope in the envelope forming circuit 1 exceeds a predetermined reference value, the information is transmitted to the controller 19, and the controller starts operating.

, p< 7 The center frequency of the i4tooth filter 2 is variable by the logic circuit in the controller 1sK)%2
00jrz to 5kHz J), at time intervals of 15++s seconds. The frequency bandwidth of the bandpass filter 20 is 1/6 octave. The output signal of the bandpass filter 2 is converted into an envelope signal by the envelope forming circuit 3, and further, the low frequency component of the signal is removed by the highpass filter 4.

This high-pass filter 4 has a cutoff frequency of 250Hz and an attenuation characteristic of 12cLB100t.

The output signal of the high-pass filter 4 is processed by the group of four band-pass filters 5 connected in parallel as described above, and then the average value voltage in each band is extracted from each signal by the average value extraction circuit group 6. Output. These output signals become multiplexer 80 input signals. At the same time, the signal of the bypass fill 4 becomes the input signal of the multiplexer B through the path 7.

On the other hand, the output signal of the envelope forming circuit 1 is an envelope signal of the original audio signal, and this signal also becomes an input signal to the multiplexer 80. The multiplexer 8 controlled by the logic circuit of the controller 19 samples the above six types of signals. Here, path 7
For the signals, time intervals are assigned so that all data for 15 wind seconds are input to the ψ converter 9. A/
The sampling period in the D converter 9 is cL 5 msec, and 30 pieces of spectre data for 15 s are recorded in the memo 1312. The memory 15 is a buffer memory that can store approximately 8 seconds of voice analysis data.

When the memory 15 stores the signal for the first CL36 seconds, the spectral pattern calculation unit 13 starts calculation.

036 seconds includes 24 spectra obtained as time-series data of 15 seconds, and one monosyllable is subdivided into one or two syllables from this spectrum.

Next, calculations in the 1-spectral pattern calculation unit 13 will be explained using the schematic flowchart of FIG. At first,
The envelope signal in the envelope forming circuit 1 is used to detect the maximum value point of a single syllable. This maximum value point provides a criterion for separating consonant and vowel regions. The spectrum of the vowel region is derived by averaging the spectrum for 15 seconds at the time of this maximum value and the spectrum for the next 15 seconds. Since the total power over the entire frequency range of the spectrum changes normally depending on the vocalization intensity, it is necessary to normalize the amplitude value of the spectrum by keeping the total power at a constant value.

The signal amplitude value that has passed through the bandpass filter group 5 is normalized by the parameters used in this normalization calculation. Using this normalized parameter (average level in various frequency ranges), the position of the input monosyllabic vowel in the vowel map shown in FIGS. 3, 4, and 5 is determined. Depending on the determined position, either a vowel is determined immediately or a plurality of vowel candidates are determined.

Next, using the envelope signal in the envelope forming circuit 1, the maximum amplitude value is set to, and a time point having an amplitude close to α5 is searched from the maximum value time point toward the monosyllable start time point, and its spectrum is determined. The difference between this spectrum and the spectrum at the time of the maximum value is calculated.

This difference spectrum is rounded to obtain consonant information. Next, proceed with the search in the direction where the monosyllable ends,
Derive the spectrum at a time point with an amplitude of α5. This is the end range spectrum. The spectrum of this end region is used to judge whether the vowel in the end region is the same as or different from the vowel at the maximum amplitude of the envelope. syllable,
If it is different, it will contain two vowels.

It is determined that it is a single syllable such as "kyu".Next, the above three types of spectra, ie, the vowel area spectrum, the difference spectrum, and the ending area spectrum, are sequentially recorded in the pattern memory 14 when the calculation is completed. When the recording is completed, calculation in the pattern matching section 15 is started. The pattern matching section calculates the degree of similarity with a standard pattern stored in advance, determines one vowel among the candidate vowels, and then determines the consonant by calculating the degree of similarity with respect to the consonants. Finally, it is determined whether there is one or two vowels, and these results are combined to determine what the input monosyllable is. The obtained result is output from the output terminal 21 via the output buffer 17. The band-pass filter group 5 and the average value extraction circuit group 6 in the circuit block diagram above are composed of the band-pass filter and the integrating circuit shown in FIG.
A system is established.

Here, operational amplifiers 20 and 21 represent band I-e filters, and operational amplifier 22 represents an integrating circuit.

In this embodiment, the frequency analysis method of the audio signal is based on a method of sweeping the center frequency of a bandpass filter, but the method is not limited to this. For example, a group of bandpass filters, a discrete Fourier It goes without saying that the Fourier spectrum obtained by transformation, the spectral envelope obtained by the linear prediction method, etc. can be treated in the same manner.

In addition, the bundle 4 filter group 5 shown in this embodiment is also included. In contrast to the method of obtaining spectrum-related parameters in real time using the average value extraction circuit group 6, instead of using these circuits, the output signal of the high-pass filter 4 is sent to the multiplexer 9.
, temporarily stored in the memory 12 via the A/D converter 9,
It is also possible to obtain the above-mentioned/dimension meter by calculation in the spectrum pattern calculation section 13. However, in this case, negligible calculation time is required, impairing high-speed calculation, but it does contribute to improving the reliability of vowel recognition.

〔Effect of the invention〕

According to the present invention, by providing a signal processing circuit centered on a plurality of bandpass filters for a spectral signal obtained by frequency analysis of a monosyllabic signal, parameters for extracting spectral features can be extracted in real time. This made it possible to determine monosyllabic vowels at high speed. By being able to determine vowels at high speed and improving the reliability of the determination, the time required for single syllable recognition, including consonant determination, will be shortened and reliability will be improved.

[Brief explanation of drawings]

FIG. 1 is a circuit block diagram of the present invention, FIG. 2 is a waveform diagram showing an example of an audio spectrum, and FIG. 3 is a circuit block diagram of the present invention. Figures 4 and 5 are diagrams explaining vowel feature extraction, Figure 6 is a circuit block diagram explaining the present invention in detail, Figure 7 is a schematic flowchart of calculation, and Figure 8 is the main part of the present invention. FIG. 2 is a circuit diagram showing an embodiment of the invention. 2...Band pass filter 3...Envelope forming circuit 4...High pass filter 5...Band pass filter group 6...Average value extraction circuit group 8...Multiplexer 9
...Memory 20.21...Band pass filter 22...Integrator circuit No. 22, No. 312] Fig. 4 Fig. 5 High tube Fig. 6 Fig. 7

Claims (1)

[Claims] A speech recognition system that recognizes the input speech by frequency-analyzing the input speech, generating multiple spectral patterns ranging from consonants to vowels, and calculating the degree of similarity with pre-stored standard spectral patterns. In the apparatus, there is provided a means for frequency-analyzing the audio signal, a means for arranging the frequency-analyzed amplitude or power spectrum values on a time axis to form a single signal, and a means for arranging the frequency-analyzed amplitude or power spectrum values on a time axis to form a single signal, A high-pass filter that removes frequency components and a signal passed through the high-pass filter,
A speech recognition device comprising means for obtaining an average value in a plurality of predetermined frequency ranges, and means for storing the plurality of average values and an output signal of the high-pass filter.