JPH0573035B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION <Technical Field> The present invention relates to an audio information processing device using zero crossing information.

<Prior art> The number of zero crossings of a waveform (hereinafter referred to as ZCC) can be realized with a simple circuit using a comparator, and
Since it also expresses a certain amount of information about the spectrum of speech, its properties have been studied and used as a means of speech analysis. In reality, if the zero crossings of the input signal are simply detected, the zero crossings will be detected even in parts where no audio is included due to noise contained in the input signal. Therefore, pure ZCC
ZCC alone is not effective for speech recognition, especially for separating speech (voiced) sections from silent sections.
Provide reliable information.

In order to avoid counting due to noise during silent periods, conventional methods include providing hysteresis to the comparator or shifting the reference voltage of the comparator from the average value of the input signal (for example, Japanese Patent Application Laid-Open No. 1989-1999). (See Publication No. 116595).
According to this, the comparator can be set so as not to respond to a noise portion whose amplitude is smaller than that of voice, that is, a silent section. The information obtained by this method will be referred to as the level crossing count (LCC).

Strictly speaking, LCC is different from ZCC, but since it almost matches ZCC in areas where the amplitude is large to some extent, such as in voice sections, it may be treated as information similar to ZCC in speech recognition, etc. However, if the reference voltage (or hysteresis, hereinafter referred to as a cross level) of the comparator is made too large, the comparator may not respond to consonant parts with small amplitude, and zero crossing information may be lost. In particular, since the energy of fricatives exists between 3K and 6KHz, the ZCC shows a very high value and there is a noticeable difference from other phonemes, but because the amplitude is small, when the cross level is large, the LCC may be a small value.
On the other hand, if the cross level is too small, the comparator will react due to noise during the silent section,
It becomes difficult to separate it from the voice section. These are
This is a problem with LCC, and when LCC is used as a feature quantity for speech recognition, etc., errors in detection of the beginning and end of words and errors in determining consonants occur, resulting in errors and rejections.

<Object of the Invention> The present invention automatically corrects the cross level of a comparator according to the signal size of an ambient noise section that does not contain audio, thereby eliminating the above-mentioned conventional audio problems. The purpose is to provide an information processing device.

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

FIG. 1 is a block diagram showing an example of the configuration of a speech recognition device using LCC. In the figure, an audio signal input from a microphone 1 is amplified by a preamplifier 2, passed through several bandpass filters (two in this example) 31 and 32, and then input to comparators 41 and 42, respectively. be done. Each of the comparators 41 and 42 compares the output voltage with a reference voltage which is the output of the cross level converters 51 and 52, and inputs the result to the microcomputer 6. The results are displayed on the output unit 7 or the like. Here, the reference voltage may be set to 0, and the converters 51 and 52 may be connected so as to change the hysteresis of the comparators 41 and 42. Inputs to the converters 51 and 52 are provided by the microcomputer 6.
is sent from the output port of In this example, each
It is possible to change in 3 bits, or 8 steps.

FIG. 2 shows a specific example of the circuit configuration of the cross-level converters 51 and 52. The integrated circuit IC in the figure is
It is configured as an 8-channel multiplexer, and has 3 bits (0 to 7) to the control terminal.
The corresponding switch of each channel is turned on by the digital signal. As a result, the tap of the externally connected resistor R is switched, and the cross level of the comparators 41 and 42 is corrected.

FIG. 3 is a block diagram functionally showing the processing flow of this recognition device. In the figure, an audio signal input from a microphone 1 is analyzed and processed by an acoustic analysis section 11. This acoustic analysis section 1
1 are the bandpass filters 31 and 32, the comparators 41 and 42, and the cross level converters 51 and 5 shown in FIG.
Corresponds to 2. The analysis process is, for example, outputting a time series of feature vectors for each unit time (frame). The obtained vector time series is
Next, the pattern conversion section 12 enters the pattern conversion section 12, which uses a method such as segmentation to express the phoneme as a series of phonemes or labels corresponding to the phoneme while referring to the standard pattern 13 such as phoneme. Detect.

Here, in the case of word speech recognition, the word pattern of the input speech is sent to the matching section 15, recognition is performed with reference to the standard word pattern 16, and the result is displayed or transmitted in the judgment result output section 17. . The noise adaptation unit 18 uses the results of the word detection unit 14 as a reference, and based on the time series of the feature vectors of the noise interval, which is the part where there is no speech interval, the acoustic analysis unit 11
Correct the cross level. If it is determined that the environmental noise is too large to be recognized, a display 19 indicating that input cannot be accepted is displayed.

The noise adaptation unit 18 will be explained in more detail. In order to realize the adaptive algorithm,
There are the following requirements.

1 From the standpoint of practicality, the faster it takes to converge to the optimal cross level, the better.

2. The amount of processing required for adaptation must be within the range that can be achieved by a microcomputer.

3. Determine weak fricatives as noise and do not perform adaptation.

4. Sudden noises must be ignored.

5. When a long utterance is input, the recognizer will output a reject due to the duration limit, but this must be distinguished from loud noise.

6 Physiological noises such as breathing sounds and tongue clicks are sometimes present before and after vocalization, and if adaptation is made in these noise intervals, the cross level will be higher than the expected cross level, so it is better to avoid adaptation in this interval. desirable.

Among the above items, there are conflicting requirements, and although it is impossible to satisfy all of them, it is necessary to realize an algorithm that satisfies all of them to some extent. The important point is that when the cross level is appropriate or excessive, silent parts (hereinafter referred to as S) can be extracted by pattern conversion, and even when the cross level is too low, S can be extracted if the amplitude of the ambient noise is small. can be extracted. However, when the cross level is too low and the ambient noise is higher than a certain level, S is not necessarily extracted, so this case needs to be treated specially.

The former, that is, S is extracted and the beginning and end of the word are detected, is case A, and the latter, that is, S
Case B is a case in which a reject occurs because the duration is too long without being extracted. In this way, the adaptive part is divided into two parts, A and B. FIG. 4 shows a flowchart of the noise adaptation section 18. In the figure,
Information on the beginning and end of a word is sent from the word detection section 14. In case A, adaptation is performed for the silent interval between utterances, and in case B, after the beginning of a word is detected (however, it is not determined in this example whether this is speech or noise). The cross level is changed when the duration exceeds a certain reference value.

In the figure, if the beginning of a word does not exist, the case A routine is entered because it has not been uttered. If it exists, it means that the utterance was made, and we will check the result of the word ending determination. If the end of the word is not detected, the process exits the noise adaptation unit 18. If detected, it is considered that the device is operating normally, and the case A routine is entered. Ordinary word speech recognition devices set minimum and maximum values for the duration of uttered words, and reject input signals outside the range.
If the maximum value is exceeded, case B is entered.

In the case A routine, from the end of the word or reject, there is a reference time T _B (approximately 0.5
2) Perform adaptation A after the continuation. The reason why T _B is placed instead of immediately after the end of the word is because of the requirement 6) mentioned earlier.
This is to avoid physiological noise in the term. In adaptation _A , the sum of feature vectors (X
), the value is mapped by a certain function or table T, and the current cross-level
Add it to LV and make it a new cross level LV′.
That is, LV'=LV+T(X)...(1) Once adaptation A is applied, feature vectors are not collected for a certain number of frames (for example, 16 frames) from that point on. This is to satisfy requirements 4 and 3. Thereafter, collection, adaptation A, and pause are continued in the same manner until the beginning of the word is detected. Here, the maximum value of the cross level is set in advance to the maximum level that does not impair recognition performance, and if LV' exceeds the maximum value, a message indicating that input cannot be accepted is displayed. At this time, the patterns input at the same time may not be sent to the matching section 15 in FIG. 3. In equation (1), if LV is appropriate, T(X) takes a value close to 0, and if LV is excessive, T(X) becomes negative and LV' decreases. T if LV is too small
(X) becomes positive and LV' increases.

In the routine of case B, even after the rejection occurs due to the duration being too long, the utterance continues for a certain period of time T _N (approximately 1 to 1.5 seconds).
That is, when the ending of a word is not detected, adaptation B is applied. In adaptation B, since feature vectors cannot be collected from the section S, an operation is performed to add a certain amount K to the LV.

LV'=LV+K...(2) After applying adaptation B, if the ending of the word is detected, the process returns to the case A routine. Of course, the input pattern at this time is rejected. on the other hand,
If T _N elapses without a word ending being detected, adaptation B is performed again. After this, repeat this operation. If LV′ exceeds the maximum value of the cross level, case A
It is similar to The explanation for case B above is a case where the recognition device does not have a function to determine whether the target input signal with an excessively long duration is due to vocalization or pure ambient noise. Adaptation B may be activated immediately after the reject occurs.

Figures 5 and 6 show how adaptation A and adaptation B occur. 100 represents feature vector collection, 101 represents adaptation A, and 102 represents adaptation B processing operation. In Figure 5, 100 _minutes have passed after the end of utterance.
collection, 101 adaptations A are occurring. There is a sudden noise in the middle, but because the duration is short, it is rejected, and from the end of the noise, after T _B has passed, case A has occurred. In FIG. 6, the large-amplitude noise continues for a long time, a reject occurs, and then _TN continues, so the adaptation B of 102 is operating. After that, S appears and is determined to be the final word, making case A.

Although the speech recognition device has been described above, the present invention can also be used in a speech synthesis device that uses LCC to determine voiced, unvoiced, silent, etc., and is useful for improving the sound quality of synthesized sounds and reducing the bit rate.

<Effects of the Invention> As detailed above, the present invention is characterized in that the cross level of the comparator is automatically changed according to the ambient noise.For example, it takes considerable skill to manually set the cross level. However, the benefits of this automation are enormous. Further, according to the present invention, the amount of processing is small;
Furthermore, since it can be realized with a simple circuit, it can be incorporated into a speech recognition device or a speech synthesis device using a microcomputer, thereby providing a useful speech information processing device.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a block diagram showing a specific example of the main parts of FIG. 1, FIG. 3 is a diagram functionally showing the flow of processing, and FIG.
This figure is a flowchart for explaining the main part of FIG. 3 in more detail, and FIGS. 5 and 6 are diagrams showing adaptation examples together with waveforms. 41, 42... Comparator, 51, 52...
Cross level conversion unit, 6...Microcomputer, 11...Acoustic analysis unit, 12...Pattern conversion unit, 13...Standard pattern such as phoneme, 14...Word detection unit, 15...Matching unit, 16... Standard pattern of words, 18... Noise adaptation unit, 100... Feature vector collection, 101... Adaptation A, 102...
Adaptation B.

Claims (1)

[Scope of Claims] 1. An input means for inputting an audio signal; an acoustic analysis unit for outputting the audio input signal input from the input means as a feature vector time series for each unit time; a pattern conversion unit that expresses the phoneme as a phoneme or a label sequence corresponding to the phoneme while referring to a standard pattern such as; a word detection unit that detects a speech interval of the input signal based on the label sequence; a matching section that performs word speech recognition by referring to a word standard pattern based on the detection result; a judgment result output section that outputs the matching result of the matching section; means for comparing a reference value biased by a quantitative amount; means for correcting the reference value according to the magnitude of a signal in an ambient noise section that does not include speech; and a function for detecting a speech section; When the end of a word of the voice input signal is detected, the section of ambient noise is analyzed after a certain period of time has elapsed from the end of the word, and the correction of the reference value is started according to the result, and the duration of the voice input signal is too long. 1. A speech information processing device using zero-crossing information, comprising: a noise adaptation unit that increases a reference value at that time by a predetermined amount when a predetermined period of time or more elapses without a word ending being detected.