JPS6165300A - Voice signal processing circuit - 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 an audio signal processing circuit that automatically tracks the noise level and changes the audio detection level according to the magnitude of the surrounding noise level to detect fiffi audio sections. The purpose of this invention is to provide a circuit for detecting.

[Background of the invention]

When automatically recognizing a voice or reproducing a voice by p^rf editing, it is necessary to automatically find only the section where the voice exists from the JJ JET issue.

r''? A common method for voice detection is to find the short-term power of the input word, and to set the period in which this power exceeds a predetermined threshold as a speech section, and the other sections as silent sections.However, Under various noise environments, the noise level differs or fluctuates.This yaw L ''+, ii M level changes 4
σ71. If the noise is +1, for example, a fixed threshold value is used for determination, it is difficult to separate the noise and the sound 111 and accurately detect only the sound superimposed on the noise.

Therefore, the next consideration is the floating threshold method, which creates the sound level of 11 backgrounds and changes the voice detection threshold based on this. As a method to automatically detect background noise level,
For example, there is Japanese Unexamined Patent Publication No. 57-1445971j. In this method, the sound point signal waveform including the audio portion is temporarily stored in an input buffer memory, and the background noise level is estimated from this. Then, based on this, the input signal waveform of the input cover sofa memory was judged again to detect the audio portion. To estimate the noise energy level, the cumulative frequency of the short-term energy level of the audio signal waveform was determined as a histogram, and the energy level indicated by the maximum error was defined as the sudden sound level. This method estimates the background noise level from the entire input waveform including speech, so it is robust against noise level fluctuations, but it requires a buffer memory, and after estimating the background noise level, the waveform in the buffer memory is Since the determination is made again, real-time processing becomes difficult, and there is a problem in that processing such as speech recognition for read-behind is significantly delayed.

[Purpose of the invention]

The purpose of the present invention has been made in view of the problems of the above-mentioned conventional technology, and is to automatically detect the noise level in real time,
An object of the present invention is to provide a circuit that can accurately detect sound superimposed on noise.

[Summary of the invention]

In order to achieve this objective, the present invention utilizes the short-term energy of the human input signal, calculates the cumulative frequency corresponding to the < noise level as a histogram, and calculates the maximum frequency power level of the histogram as the noise average level. 17 so that the cumulative frequency is detected as 17, and when the cumulative frequency exceeds a predetermined value, all cumulative frequencies are subtracted within a range in which the cumulative frequency does not become a negative value, so that the cumulative frequency does not always exceed the predetermined value. This takes advantage of the fact that while the noise power remains within a certain range as long as it is stationary, the sound part takes on various levels, and also because the noise level is successively estimated from past power levels. MF allows voice detection in real time.

[Embodiments of the invention]

The present invention will be explained below using examples. FIG. 1 is an example of speech recognition compensation for a specific speaker. First, the speaker's voice is registered in the standard voice storage section by recording and recording. When registering, it is necessary to automatically detect the voice part from the human signal waveform.

First, an input audio signal from a microphone or the like is input to input section 1.

The analysis section 2 analyzes the frequency spectrum of the captured charcoal.

It decomposes and outputs feature parameters that are the basis for recognition and voice usage parameters that are the basis for voice detection. More specifically, the human power unit 1 uses a 5.6 KHz low-pass filter to extract only the voice band signal from the human voice signal. a, then sampled at 8 KHz, converted into a 12-bit digital audio signal by analog-to-digital conversion, and sent to the analysis section 2. In the analysis section 2, the audio signal from the human power section 1 is subjected to frequency spectrum decomposition using 16 band-pass filters, smoothed using a low-pass filter, and then
For example, 16 feature parameters are output by sampling at intervals of 20m513c. In addition, in the analysis section 2, the input section 1
The audio signal is divided into 20m5ec, the short-term energy during this time is counted, logarithmically converted, and output as audio detection parameters. This time unit of 20 m5ec is the timing at which the feature parameters and voice detection parameters are output from the analysis section 2, and is also the basic unit of analysis, which is called a frame. The noise level detection unit 5 inputs the audio power in units of frames, and uses the patented method described below to
The noise level is estimated on a frame-by-frame basis using the accumulated past audio power. The voice detection unit 4 & Mako 1 determines whether the frame voice is a U sound or no sound using both the estimated level of scratching sound and the voice power. Discrimination of sound/silence is 1
i;Ii Sound 1 at a certain level high based on the noise level
119, set the j value and check whether it exceeds this threshold. Next, use this information to detect the voice section. When it is considered your voice.

Further, when the silent section continues for a predetermined time or longer, the sound section is assumed to have ended and is outputted to the standard speech storage section 6 as a speech section. The standard voice storage unit 6 inputs the feature parameters from the analysis unit 2 in frame units, and inputs the characteristic parameters from the analysis unit 2 to the voice detection unit 4.
Gate the speech section from , and store only the feature parameters within the speech section. The above is the recording work. After registering the predetermined voice in this way, the recognition phase begins. When the recognition activation enters the collation unit 5 and standard voice storage unit B, the input signal is analyzed and noise level detection and voice detection are performed in the same manner as in registration. The matching unit 5 takes in the feature parameters obtained by analyzing the input signal and the speech section information, and uses the feature parameters of valid speech sections as the object of matching. Then, the already registered voices are sequentially inputted from the standard voice storage section 6), and the name of the best standard voice is outputted as a recognition result by soft comparison with the human voice. The above is the flow of recognition. Since the noise level and voice are detected in real time in this way, the matching process can proceed in parallel from the time the start of the voice section is detected, making the recognition process extremely fast.

Figure 2 shows an example of speech analysis and synthesis using this patent. Since the input audio is compressed and stored by starting the analysis, it is necessary to automatically detect the audio portion from the input signal waveform in this case as well. An input audio signal from a microphone, etc. is taken into the input section 1, and a 12-bit 11" CM digital signal is output as in FIG. The audio power is output to the noise level detection section 6.The analysis section 2 also outputs the sound power to the noise level detection section 6.The analysis section 2 outputs the sound power to the noise level detection section 6.
The waveform data is compressed into 4 bits using a POM (referred to as POM), and the 4-bit waveform data is output to the compression section 7. In the noise level production section 3, audio power is input in frame units of 20m5θC,
After estimating the noise level, the voice detecting section 4 detects an effective voice section using the voice power in combination, and outputs the section position 11 to the compression section 7. The compression section 7 further compresses the 4-bit speech waveform based on the speech section information from the speech detection section 4. Compression is performed in the force/sound section.

If the voice section includes a voice section and a silent section that does not contain voice information, the silent section is deleted and only the length of the silent section is stored. The voice to be memorized may be a single (1f1) or a sentence. In the case of a sentence, the sentence ending condition may be set to 1 second, for example, and the sentence is considered to be finished when silence continues for 1 second. This compressed p a
The M waveform data is labeled and recorded in the storage section 8. Whenever the recording is finished, the synthesis is activated.

Can be restored to the original waveform. The synthesis section is now in the storage section 8.

When the type of voice is manually entered along with the motion, the storage section checks the address where the voice is stored, decompresses the compressed 20M waveform data from there, and sends it to the length/synthesis section 9.

The decompression/synthesis section 9 performs decompression processing that is the opposite of ADPCMIIE compression, and also restores the silent sections that have been deleted and compressed to restore the original waveform and output it. The above is the flow of audio analysis and synthesis, and noise level detection and audio detection are performed in real time when analyzing and compressing recording.

Next, an embodiment of a method for estimating the noise level in the noise level detecting section 3 of the present invention will be described in detail. FIG. 4 shows a flowchart, and FIG. 3 shows a Hasugi diagram for estimating the noise level. First, the cumulative frequency corresponding to the level of voice power sent from the analysis section 2 is counted up by +1,
Modify the histogram. This is done in frame time units. If the count continues to increase, the cumulative frequency will eventually overflow.

Therefore, we set a limit so that it is less than a predetermined amount 1. If the frequency becomes 0 or more, all cumulative 1 f 1 degree minute is 1 1
Count down. However, when it becomes a negative value due to countdown, it is reset to zero and controlled so that the frequency is always within O-a. The noise level can be estimated by detecting the maximum value of the cumulative frequency and using the audio power corresponding to this value. By controlling the cumulative degree in this way, you can not only prevent overflow, but also ensure that even if the noise level changes, the maximum part of the 0 degree will eventually move to the current noise level and accurately track the noise level. This situation is shown in Fig. 3.The horizontal axis represents time and the vertical axis represents the audio power level.At t=Q, all cumulative frequencies are reset to zero.

The histogram at point A is as shown in No. 31J(A).

The vertical axis represents the voice power level, and the horizontal axis represents the cumulative frequency. The voice power is dispersed in the background δ sound, and it can be estimated that the level of the h'5 sound is P, 7.

Therefore, the voice detection threshold can also be set correctly, and thereby accurate voice sections can be obtained. When analyzing and recording voice 1, the voice detection threshold becomes even more important. Although the recording target is a voice section, silent portions below the voice detection threshold are deleted as described above. This makes it possible to compress and record only valid audio.

In the histogram (B) at point B, the audio portion is zero. This is because the background noise level reached a predetermined amount of 0 and other noise levels decreased.

However, it never becomes a negative value. The dawn sound level is decreasing from point B. Both the estimated noise level P1■ and the voice detection threshold remain at their previous values for a while, but at time 0, looking at the histogram (a), the cumulative frequencies of the previous noise level and the current house sound level match, Only then can the noise level be determined correctly. Although voice 2 cannot be detected using the previous voice detection threshold, it can be detected correctly by automatically tracking and correcting the noise level in this way.

As can be seen from the histogram (1) of the navigation point, the cumulative (゛?1) degree of the previous noise level PV has decreased, indicating that automatic tracking is being performed.

Note that the present invention is not limited to the above-mentioned embodiment. In this embodiment, the short-range sound energy pair: "'t f + α is used as the voice detection parameter, but the energy may be used as it is, or the voice It is also possible to combine the power pond with a high frequency 2/low frequency power ratio that represents the ratio of high frequency components and low frequency components. The high/low frequency power ratio of this noise can be estimated using a histogram in the same manner as the voice power.

〔Effect of the invention〕

According to the present invention, the problem of noise level fluctuation is solved, the noise level can be automatically tracked in any noise environment, superimposed voices can be accurately detected and rewritten, and voice recognition and analysis can be performed. Its practical advantages in synthesis, etc. are extremely large.

[Brief explanation of the drawing]

Figure 1 is a block diagram of speech recognition using the present invention, Figure 2 is a block diagram of voice analysis and processing using the present invention, Figure 3 is a diagram showing an example of voice detection in Figure 11 and Figure 2. , Figure 4 is a noise level detection flowchart of Figures 1 and 2. OP...Short interval audio power...Estimated background noise level S...Cumulative frequency Attorney Patent Attorney...1 ．． (1 (ii Akio No.
1 Volume 1, Volume 1, Volume 1, Volume 4, Volume 1

Claims (1)

[Claims] 1. Means for extracting short period audio power from an audio input signal; Noise level detection means for estimating a background noise level from the extracted audio power; and Audio detection means for detecting a speech section from the noise level and audio power. In the audio signal processing circuit, the noise level detection means obtains the frequency with respect to the noise power level as a histogram,
The power level of the maximum frequency in the histogram is estimated as the noise level, and when the cumulative frequency exceeds a predetermined amount during histogram calculation, all cumulative frequencies are subtracted, and if the cumulative frequency is a negative value, it becomes zero. An audio signal processing circuit characterized in that the frequency is suppressed to a positive value within a predetermined amount by being reset.