JPS6146998A - Voice head detector - 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] [Industrial Application Field] The present invention detects and outputs the time point corresponding to the start of a voice section by performing sampling and other processing on amplitude changes with respect to time changes in a voice waveform. The present invention relates to a voice start edge detection device.

[Prior art] Conventionally, the voice start edge detection device of this scale first samples continuous changes in amplitude with respect to time changes in the voice waveform, converts it into a string of multiple values, and then squares each value. and convert it into a sequence of short-term energy values of the voice. Next, in this short-term energy value column J5, an interval exceeding a preset threshold corresponding to the lower limit of voice short-term energy was detected, and this was determined as the starting point of the voice interval. .

[Problems to be solved by the invention] In general, the short-term energy of speech does not necessarily increase or decrease smoothly even in one speech section, for example, in bursts (
It is possible that something that was gradually increasing, such as burst), suddenly decreases to a very low level and then increases again to a high level.

For this reason, the short-term energy of the voice will go up and down the preset threshold multiple times, and if you unambiguously decide which interval is the start of the voice interval, you will not be able to accurately determine the voice interval. This is inconvenient because it may not be possible to detect the starting end.

In addition, if unexpected noise increases by d before and after the voice section due to the speaker's movement or changes in the surrounding situation during voice input, the energy will exceed the threshold for a short time even though it is outside the voice zone. This can also occur, which is an obstacle to accurately detecting the beginning of a voice section.

[Objective of the Invention] The object of the present invention is to solve the above-mentioned conventional problems, and to solve the problem of irregular reversal of short-term energy such as voice including burst, or temporary mixing of unpredictable noise. The impact of these energy fluctuations is kept to a minimum,
Moreover, it is an object of the present invention to provide a speech start end detection device that can detect the start end of a speech section through simple and short-time processing.

[Means for solving the problem] In the present invention, when a high energy section is detected in which the sequence of short-term energy values of the input voice exceeds the first threshold value continuously for a predetermined period of time or more, A second set in which the sequence of short-term energy values is set to be less than or equal to the first threshold before the high-energy section exists.
Detect the point in time when the value is below the threshold. Next, a period from this point to a point a predetermined time later is set as a starting edge detection section.

Next, detect a point in time when the series of short-time energy values falls below the third pA value, which is set below the second @ value without exceeding the second [I, within the start detection interval. The detection is performed in a temporally backward direction, and the detected time is taken as the start of the voice section.

[Operation] First, by confirming the existence of the high-energy section using the first rii value, it is possible to determine whether the energy of a voice normally having a predetermined level or higher does not persist for a predetermined time or longer.
In other words, a section where the noise level is only temporarily high is excluded from the beginning as not a speech section. In addition, by setting the point in time when the voice falls below the third threshold, which is set lower than the second threshold, as the start of the voice interval without exceeding the second threshold again, the voice can be stopped even though it is within the voice interval. This suppresses misidentification that occurs when the short-time energy temporarily drops to the same level as the noise-only section. In addition, the start end detection section is set to limit the range in which processing for detecting the start end is executed.
This suppresses misidentification of the beginning of a voice section due to fluctuations in the noise level outside the voice section.

[Embodiment] An embodiment of the present invention will be described below with reference to FIGS. 1 to 5.

In FIG. 2, a microphone 1 that picks up the voice produced by a speaker is connected to an A/D converter 3 via an amplifier 2. Here, the amplifier 2 amplifies the level of the sound picked up by the microphone 1 to a level suitable for subsequent processing.

Further, the A/D converter 3 is connected to a central processing unit (hereinafter referred to as CPU) 4. Further, connected to the CPU 4 are a ROM (read-only memory) 5 and a RAM (extra-writable memory) 6, in which programs and the like for each process are written.

Kokote, RAM 6 Ha, A/Dvl exchange? +t in A3
An amplitude buffer 6a into which the amplitude values of the sampled audio waveforms are written one after another, a square buffer 76b into which the square values of the amplitude values are written, and a short time period calculated based on the square values. an energy buffer 60 into which energy values are written; a first threshold buffer 6d into which two time points at which the short-time energy (10) falls below a preset first threshold value are written; It functions as a second threshold buffer 68, etc., in which each point in time when the second threshold value is first lowered in both the backward and forward directions is written.

In the above structure, the speaker then repeats the action in Japanese.
Let us explain the case where the person pronounces °゛. The sound of "ga" pronounced by the speaker is first collected by microphone 1, amplified to an appropriate level by amplifier 2, and then sampled by A/[) converter 3 to calculate the amplitude of the speech waveform over time. The data is converted into data as a string of a plurality of values indicating changes and input to the CPIJ4. Here, the A/D conversion shown in step 21 in FIG. 1 is performed at a sampling frequency of 8 KHz in accordance with the sampling theorem in order to obtain information on audio waveforms up to 4 KH2. Therefore, 8,000 pieces of audio waveform amplitude data are obtained per second, in other words, one piece of amplitude data is obtained every 125 microseconds, and is sequentially loaded into the amplitude buffer 6a. If adjacent amplitude data are connected with straight lines and illustrated with time on the horizontal axis and voltage on the vertical axis, a waveform diagram as shown in FIG. 3(a) will be obtained. Further, this A/D conversion is started when the CPU 4 detects that the switch of the microphone 1 is turned on, and is executed while the switch of the microphone 1 is turned on. Therefore, in the waveform diagram, a section containing only noise appears before the speech section. Let A(Q) represent the q-th amplitude data among the amplitude data every 125 microseconds. Here, q is an integer from 1 to n, and n is the total number of amplitude data. Furthermore, when q increases by 1, 125 microseconds elapse.

Next, the process proceeds to step 22, where each of the n pieces of amplitude data A(Q) shifted by 1 in step 21 is squared and n
The squared values S (Q ) are calculated and sequentially written to the squared buffer 76b. If we put this into a formula, S (Q
) −A (Q ).

Next, proceed to step 23, and every 64 squared values 5 (Q),
That is, the sum is calculated every 8 milliseconds, and the short-time energy 1aE(j>) at the center of the interval is sequentially filled into the energy buffer 6C.

When this is expressed as a formula, it is expressed as follows. Here, j takes an integer from 1 to (n/64). In addition, each short-time energy value E(j)
The address of the energy buffer 76C is set corresponding to the time point, and the corresponding time point can be found by specifying the short-time energy value E (j>.The horizontal axis represents time, and the vertical axis represents the time point. The short-time energy idirectE(j)
When expressed with its maximum value as the absolute standard, it becomes as shown in Fig. 3(b). In the figure, the low-level sections before and after are sections of ambient noise other than voice, and it is necessary that voice is input even in sections that are not shown, going back in time or moving forward in time. For example, although there are some fluctuations due to fluctuations in the If level, it continues to remain at approximately the same low level.

Next, the process proceeds to step 24, and in the column of the short-time energy value E(j> calculated in step 23, the preset first threshold T1-0.1 is set for a predetermined time or longer, 120 milliseconds in this example). Sections exceeding the above, Zunawara, 3rd
In Figure (b), the interval t1 to t2 is detected, and this interval is classified as a high energy interval and two time points tl and t2 are detected.
is written into the 11th shout value buffer 6d. Here, the first threshold value T1, which is indicated by a broken line in FIG.
It is direct.

Next, the process proceeds to step 25, and the process goes back in time from the time point t1 to obtain the second rll value T2-0.
004, the start point detection start time t3 of the voice section, and the second
The time point t4 at which end detection of the Zunawara voice section starts is stored in the second threshold buffer 60. Here, the second threshold T2 shown by the broken line in FIG. 3(b)
is set slightly higher than the lower limit of the short-time energy of the voice section so that the short-time energy does not exceed this second W4 value T2 even if the noise level changes slightly outside the voice section. It is.

Next, the process proceeds to step 26, at the start point detection start time t3.
A predetermined period of time in the direction backward in time from 140 in this example.
The range is limited to go back only milliseconds, and in the backward direction, a time point t5 is detected when the second threshold value is not exceeded again and the time point t5 falls below a third threshold value set in advance and shown by a broken line in FIG. 3(b). Then, the data signal corresponding to time t5 is outputted as the start of the voice section. The predetermined time here is set to be long enough to ensure that the beginning of the voice section is included within the predetermined time. Further, the third threshold value corresponds to the lower limit value of the short-time energy of the voice section.

Next, the process proceeds to step 27, where the range is limited to a predetermined period of time, 140 milliseconds in this embodiment, from the end detection start point t4 in the time elapsed direction, and the second 1311 value is thereafter set again in the elapsed direction. The data signal corresponding to the time point t6 is detected as the end of the voice section when the time point (6) falls below the third threshold value without exceeding the third threshold value. ,
It is set to be long enough to ensure that the end of the voice section is included within that time. In this way, the audio section ■1
(t5-t6) is detected, and this section is shown in Figure 3(a).
This almost coincides with the speech section determined from the diagram showing the relationship between time and speech amplitude shown in FIG.

Further, FIGS. 4 and 5 show the results of performing the same processing as described above on Japanese sounds pronounced as "ba" and "ba". In these examples as well, voice sections ■2 and v3 are detected as in the case shown in FIG.

[Effects of the Invention] As described in detail above, the voice start edge detection device according to the present invention is capable of detecting short-term energy obtained by representing the voice energy as a representative burst at each predetermined time, such as a burst. Even if the noise level temporarily drops to the same level as the noise-only section within the section, misidentification of the start of the speech section due to this effect can be suppressed with a simple configuration. In addition, sections where the energy above a predetermined level normally distributed by speech does not last for more than a predetermined time, in other words, where the noise level is only temporarily high, are not considered to be speech sections and are treated as such from the beginning. It is excluded from processing, and the execution range of the processing for detecting the start end of a voice section is limited. Therefore, it is possible not only to prevent an increase in processing time due to the process of detecting the start of a voice section even for sections unrelated to the voice section, but also to prevent fluctuations in the noise level that fall outside the voice section. It is also possible to prevent an error from occurring in detecting the start of a voice section due to the influence of. In addition, the first to third thresholds necessary for executing each process are all set in advance taking into consideration the general characteristics of voice, so the 9!
U-end detection can be performed through simple processing, and an increase in processing time can be suppressed, making it very effective as a step in speech recognition.

[Brief explanation of the drawing]

Fig. 1 is a flowchart of an embodiment of the present invention, Fig. 2 is a block diagram thereof, and Figs. 3 (a) and (b) show the results of performing each process on the voice pronounced "ga". , Figures 4(a) and (b) each have 51 syllables for the sound pronounced ``°°ba''.
A diagram showing the results of executing the process a, Figure 5 <a> and (b)
) is a diagram that does not show the results of performing each process on the voice pronounced "ba". In the figure, 1 is a microphone, 2 is an amplifier, 3 is an A/D converter, 4 is a CPU, 5 is a ROM, and 6 is a RAM.

Claims (1)

[Scope of Claims] 1. Sampling means for sampling amplitude changes with respect to temporal changes in audio and converting the samples into a string of a plurality of values; and square conversion means for converting the plurality of values into a string of values obtained by squaring each of the plurality of values. and representative value converting means for dividing the row of squared values into a plurality of groups and converting the row of representative values into a row of representative values for each group; a high-energy interval detection means for detecting a high-energy interval that exceeds the first threshold; means for detecting, means for setting a start detection interval from the point in time when the row of representative values falls below the second threshold value to the point in time a predetermined time back; and when the row of representative values exceeds the second threshold value again. Detection of the point in time when the value falls below the third threshold set to be less than or equal to the second threshold is performed in the backward direction in time within the start edge detection section, and the data signal at the detected point is detected in the voice section. A voice start edge detection device comprising: a start edge output means for outputting the start edge of the voice.