JPH0480398B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Technical Field The present invention relates to a word speech recognition device with improved performance for unknown input speech.

<Background> In general, a word speech recognition device amplifies the uttered word speech with an amplifier such as a microphone amplifier, and then uses feature parameters that can express the characteristics of the speech in an acoustic analysis section, such as BPF group (Bant Pass). filter group)
is analyzed into the power spectrum, autocorrelation function, and number of zero crossings. After this, the feature extraction unit determines the word section and compresses the feature parameter time series within the word section in the time axis direction from the perspective of reducing the amount of pattern memory and reducing the calculation time during matching. It is created as an input pattern using a predetermined algorithm.

After this, the input pattern is registered in the same way as before (targeting a specific speaker),
Alternatively, matching is performed with a standard pattern created by many speakers (targeting unspecified speakers), and the most similar pattern is output as a recognition result.

In this standard recognition, erroneous recognition may occur for sounds other than the words to be recognized, especially for voices carelessly uttered by the user, coughing, etc., and also for sounds suddenly generated due to ambient noise. The inconvenience of getting used to it is common.

<Purpose of the Invention> The present invention is intended to improve such inconveniences, and by adding time length information for each word to a standard pattern and performing pattern matching processing with word length restrictions, the characteristics of speech can be improved. Although it depends on the expressiveness, if the discriminative ability for the phoneme of the feature is low, very similar patterns will be created even for words with different durations other than the recognition target word, and this will cause misrecognition.
Reduce this misrecognition.

As the standard pattern capacity is reduced, if the standard pattern length is fixed at a constant length regardless of the difference in word length, the partial loss of features will increase as the word length increases, leading to a decrease in performance. . Misrecognition caused by this is reduced.

In addition, if the variance in word length between recognition target words registered as standard patterns is large, that is,
If there are many target words with different lengths, matching can be limited to only standard patterns whose input word lengths are within the allowable word length of each standard pattern.
This is a preliminary selection operation that narrows down the standard patterns to be matched to a small number by setting a limit on the word length, and is useful for reducing the matching processing time.

<Example> An example of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention.

Word sounds uttered into the microphone 1 are amplified by a preprocessing unit 2, and are subjected to processing such as pre-emphasis as necessary. If the processing after this stage is digital processing, the audio waveform is converted into a digital signal in this preprocessing section 2.

Subsequently, the feature analysis section 3 analyzes quantities that can represent the features of the voice, such as short-time power spectrum, autocorrelation function, and number of zero crossings.

From this feature time series, the word section extraction unit 4 extracts the feature time series 101 and duration information 102 of the word section. The feature time series 101 of the word section is compressed in the time axis direction by the pattern creation unit 5, and is created as a pattern 103.

The pattern 103 and the duration 102 are sent to the standard pattern storage section 6 or the input pattern storage section 7 by movement of the changeover switch _SW1 or _SW2 .

Normally, in a recognition device targeted at a specific speaker, in order to register the recognition target word in advance, S1 is closed and the recognition target word is registered in the standard pattern storage section 6. In the case of recognition processing, S2 is closed and the input word pattern is temporarily stored in the input pattern storage section 7.

When targeting unspecified speakers, the changeover switches SW ₁ and SW ₂ are not needed, and the standard pattern storage unit 6 stores feature patterns extracted from words uttered by multiple speakers and standard word duration lengths in advance. Stored. In this case, this storage section 6 is composed of a ROM or the like.

During recognition processing, the pattern matching unit 8 compares and matches the pattern stored in the standard pattern storage unit 6 with the current input pattern in the input pattern storage unit 7, and determines the number etc. of the most similar standard pattern. is output as the recognition result 104.

The control section 9 controls each section and between each section, and has an external input section 10 that can adjust the allowable range from the outside by calculating from the duration of each word in the standard pattern. External input section 10
The allowable setting code 105 input from
It is expressed as a lower limit ratio (%).

In FIG. 2, the dotted line area in FIG. 1 is constructed with a 1-bit microcomputer 11. The microcomputer 11 has a CPU 12, a ROM 13,
If the standard pattern includes the RAM 14 and is aimed at unspecified speakers, the standard pattern will be stored in the ROM 13.

For speaker-specific word speech recognition, boards and LSIs have been developed with the aim of reducing costs due to advances in digital signal processing technology and LSI technology in recent years, but speaker-independent recognition requires large and expensive equipment. It's staying.

In word speech recognition for non-specific speakers, there is no feature parameter that can be applied to all phonemes as a feature parameter that has little variation depending on the speaker.However, when performing rough classification, zero-crossing is a parameter with relatively few individual differences number is known. Taking the number of zero crossings as a parameter,
As a result of various studies, we found that by adopting a short-time constant level crossing analysis method, it is possible to realize an unspecified speech recognition device using a CMOS 1-chip LSI, excluding the analog circuit section as described above. can.

Figure 3 shows the definition of the number of short-term constant level crossings.
This feature quantity is obtained by counting the number of times the audio waveform v crosses a certain threshold level _th for each short time frame (period τ). By adjusting and setting the threshold level to a value slightly larger than the steady ambient noise, it is possible to distinguish between word speech sections and silent sections.

The feature amount can represent the relative strength of the speech spectrum, and Figure 4 shows the feature time series obtained by constant level intersection analysis (when a stop occurs).
shows. (a) is a speech waveform, and (b) is a constant level crossing series. As seen in the figure, the value is low for voiced sounds, and for unvoiced sounds, especially fricatives such as "S" and "SH",
It shows high values for affricates such as “TH”. In the zero-level intersection analysis method, the feature analysis unit 3 extracts the number of constant-level intersections for each frame as a feature.

In addition, in FIG. 4b, point A is the beginning of a word, and point B is the end of a word. Based on this, the word segment extraction unit 4 extracts word duration information 102, and gates the feature time series to form a pattern creation unit 5. It will be supplied to In the case where a one-chip microcomputer 11 is used as shown in FIG. 2, this processing is performed inside the one-chip microcomputer 11.

FIG. 5 shows an example of a standard pattern stored in the standard pattern storage unit 6 or the ROM 13 of the 1-chip microcomputer 11, in which a feature series X and a standard word duration Y are stored for each word. Furthermore, as shown in the figure, information such as a result output code Z to be output when this pattern is selected may be included.

FIG. 6 is a flowchart showing the main part of the recognition procedure, and the dotted line indicates the matching procedure based on the word length tolerance.

After starting, the input status of the allowable setting code 105 from the external input unit 10 is first seen, and the upper limit and
This code is read and stored for the lower limit calculation. Thereafter, the beginning of the input word as at point A in FIG. 4 is detected, and if the beginning is detected, the time to the end of the word at point B in FIG. 4, that is, the duration of the uttered word is measured. Next, an input pattern is created based on the feature time series of this word section, and the input pattern storage unit 7
(In Figure 2, a 1-chip microcomputer 11
RAM 14). The input pattern includes the word length information measured above. Then, matching processing is started using the word length tolerance value within the dotted line in the flowchart.

First, the standard word duration length Y of the standard pattern stored in the standard pattern storage unit 6 (in the ROM 13 of the 1-chip microcomputer 11 in FIG. 2) is loaded, and the upper limit of the permissible word length value is taken in with the permissible setting code 105.・Calculate the lower limit. Based on this, it is determined whether the input word is within the allowable value, and if so, the inter-pattern distance from the standard pattern is calculated. After calculating the inter-pattern distance, or if the input word is not within the allowed value, jump this inter-pattern distance calculation, judge whether the standard pattern is finished, and start the next comparison. The standard word duration length Y of the standard pattern is taken. The above steps are repeated for each word until the standard pattern to be compared and matched is completed.

Here, if the discriminative ability for the phoneme of the feature is low, very similar patterns will be created even for words with different duration lengths other than the recognition target word, but by limiting the word length as described above, Calculation of inter-pattern distances is omitted, reducing misrecognition.

In addition, along with the reduction in the capacity of standard patterns,
Even if the standard pattern is fixed to a constant length regardless of the difference in word length, the partial loss of features will increase as the word length increases, leading to a decrease in performance. By using that as information for comparison and verification, misrecognition of this can be reduced.

Further, by setting a limit on the word length, it is possible to perform a preliminary selection operation to narrow down the standard patterns to be matched to a small number, such as by omitting the calculation of the distance between patterns, which is useful for reducing the matching processing time.

After such comparison and verification, the one with the minimum distance is searched among the calculated standard patterns. Here, a predetermined threshold value is further defined, and only those that are equal to or higher than the predetermined value are determined to be valid and the results are output. Those below a predetermined value are output as rejects.

In this way, the process returns to reading the permission setting code 103 from the external input unit 10 again. It is possible to input any permissible setting code 105 from the external input unit 10, and if it changes, the changed permissible setting code 105 is read. This sets the optimal tolerance according to the usage situation and speaker, and in actual examples,
For example, even if you add a limit within the tolerance range of ±20 to 30%,
After improving performance for unknown input audio,
It was confirmed that the effect on the recognition rate was also low.

In the above example, a method of searching for similar patterns using the distance between patterns as a measure was shown. If the device has a creation section and the recognition result is something that completely matches this transition sequence, calculate the word length tolerance after finding a match, and check if the input word length is within the tolerance. It is also easy to construct an algorithm that outputs the number of the matched pattern if there is one, and rejects the input word otherwise.

<Effects of the Invention> As described above, according to the present invention, in a word speech recognition device that uses a constant level intersection number analysis method, time length information for each word is added to a standard pattern, and furthermore, the time length information for each word is added to the standard pattern, and the situation of surrounding noise etc. A means is provided for externally setting and inputting permissible setting information of word duration length in order to change the permissible word length value according to the word length, and a calculation means is provided from the input permissible setting information and the time length information for each word. Therefore, the upper and lower limits of the permissible range are determined, and based on these, it is determined whether the input word length is within the permissible value, and pattern matching is performed only for those words that are within the permissible value. For example, if the ambient noise becomes louder and longer input word lengths are extracted, it is possible to prevent the recognition rate from decreasing by changing the tolerance setting information and widening the tolerance range. Furthermore, by adjusting the permissible range by changing the permissible setting information in accordance with the speaking speed of the person using the word speech recognition device, it is possible to prevent a decrease in the recognition rate and reduce the matching processing time.

As a result, it is possible to sufficiently improve the low rejection ability for unknown speech other than the recognition target word that has a similar phonological sequence, which is a drawback of the fixed-level intersection analysis method. It becomes possible to employ a constant level crossing number analysis method that is suitable for integrated circuit implementation and has relatively small speaker variation in the device, and it becomes possible to realize a word speech recognition device with a CMOS 1-chip LSI.

[Brief explanation of the drawing]

Fig. 1 is a block configuration diagram showing an embodiment of the present invention, Fig. 2 is a block diagram when a 1-chip microcomputer is used, Fig. 3 is a diagram showing audio waveforms, and Fig. 4 is a diagram showing audio waveforms a and 1. A diagram showing a comparison of feature series b, Figure 5 is a memory map of the standard pattern,
FIG. 6 is a flowchart explaining the operation of the main parts. 1...Microphone, 2...Preprocessing unit, 3...Feature analysis unit, 4...Word section cutting unit, 5...Pattern creation unit, 6...Standard pattern storage unit, 7...Input pattern storage unit, 8...Pattern matching section, 9...Control section, 10...External input section, 11...
...1 chip microcomputer, 12...
CPU, 13...ROM, 14...RAM, X...
Feature series, Y...Standard word duration length.

Claims (1)

[Scope of Claims] 1. A storage means for storing a feature sequence pattern X and standard word duration information Y obtained by a constant level intersection number analysis method as a standard pattern to be compared and matched for each word; and utterance. a feature analysis means for performing a feature analysis of the uttered input word speech using a constant level intersection number analysis method; measurement means for measuring time length; input pattern creation means for creating an input pattern based on the characteristic time series of the uttered input word sounds obtained by the feature analysis means; and word continuation in the standard pattern. an external input device for inputting permissible setting information for adjusting permissible values based on time length; and each standard word time length information Y in the standard pattern.
means for calculating an upper limit and a lower limit of a permissible word length value based on the input permissible setting information; and a means for calculating an upper limit and a lower limit of a permissible word length value based on the input permissible setting information; and means for performing pattern matching processing with the input pattern only for standard patterns that are within the range of allowable values based on the determination result. word speech recognition device.