JPH04362700A - Voice recognition device - Google Patents

Abstract

PURPOSE:To realize a logarithmic converting method which can handle even a small input without increasing an error and eliminates an abrupt change in the characteristic of an output value and to perform fast logarithmic conversion without using a large memory. CONSTITUTION:The voice recognition device consists of an acoustoelectric transducer 10, an acoustic analyzing means 20, a registered voice storage means 30, and a recognizing process means 40 which performs a recognizing process by using the feature quantity of an input voice obtained by the acoustic analyzing means 20 and the feature quantity of a registered voice stored in the registered voice storage means 30. This acoustic analyzing means 20 has a logarithmic converting means 24 and converts the input signal according to a logarithmic curve when the input signal is larger than a predetermined threshold value and according to a straight line which connects with the logarithmic curve at the threshold value and passes an origin when the input signal is less than the threshold value.

Description

[Detailed description of the invention]

0001

[Technical Field] The present invention relates to signal processing technology, more specifically,
Regarding logarithmic conversion technology for audio signals, for example, audio processing equipment used in environments with large differences in audio levels or in noisy environments, such as audio processing used in offices, cars, factories, and homes. This is suitable for application to devices.

0002

2. Description of the Related Art In audio processing technology, logarithmic transformation is often used to handle feature quantities such as power and power spectrum. This may be to accommodate a wide dynamic range of audio, to match the human hearing characteristics, or to treat feature amounts of audio with different levels equivalently. In this transformation, the input (≧
0), by simply applying logarithmic transformation, when the input is 0, we get
This is not practical because the output becomes negative infinity. Therefore, μlog conversion is generally used. That is, for example, y=A·log(x+1) (A: constant). However, in addition to the logarithmic transformation itself giving a large error for small inputs, this transformation further increases the error for small inputs by adding 1. As a conventional technique for solving these problems, there is Japanese Patent Publication No. 63-34477, but the invention described in this publication

0003

[Math 1]

[0004] The same output is given to inputs below a certain value, eliminating the error caused by taking the difference (or differentiation) of the outputs. Another disadvantage is that because the curve in the above equation is not smooth, the difference (or differential) of the output becomes discontinuous for an input around a constant value, which causes a new disadvantage in that it adversely affects subsequent processing. Further, although logarithmic conversion generally refers to a ROM table, this method requires a large memory area and is not suitable for processing in a DSP (digital signal processor) or the like with a small memory.

[0005]

[Purpose] The present invention has been made in view of the above-mentioned circumstances, and is capable of handling even small inputs without increasing errors in audio signal processing, and in which the output value suddenly changes Achieving a logarithmic conversion method that does not change the process, executing logarithmic conversion at high speed without using a large amount of memory, and accurately detecting speech intervals even for small inputs. However, this method was developed for the purpose of accurately extracting voice features.

[0006]

[Structure] In order to achieve the above objects, the present invention provides (1)
an acousto-electric conversion means for inputting a voice; an acoustic analysis means for acoustically analyzing the input voice obtained by the acousto-electric conversion means; a registered voice storage means for storing feature quantities of registered voices registered in advance; A speech recognition device comprising a recognition processing means for performing recognition processing using the feature quantity of the input speech obtained by the acoustic analysis means and the feature quantity of the registered speech stored in the registered speech storage means, The acoustic analysis means has a logarithmic conversion means, and the logarithmic conversion means converts the input signal based on a logarithmic curve when the input signal is equal to or higher than a predetermined threshold value. is below the threshold, connecting to the logarithmic curve at the threshold and converting the input signal based on a straight line passing through the origin, or (2) converting the input signal according to (1) above. In the speech recognition device, the logarithmic conversion means has a comparison means for performing a comparison operation of digital signals, a shift means for performing a shift operation of the digital signals, and an addition means for performing addition of the digital signals, and the logarithmic conversion means has the above-mentioned ( 1) Approximate calculation of the logarithmic transformation as described above is performed, or in the speech recognition device of the above (1) or (2), (3) the input signal input to the logarithmic transformation means is based on the voice input. (4) It is a power spectrum or an amplitude spectrum of the audio input. Hereinafter, the present invention will be explained based on examples.

[0007] The present invention will be described in detail below. FIG. 1 is a block diagram for explaining an embodiment of a speech recognition device according to the present invention. In the figure, 10 is a microphone that converts input speech into an electrical signal. 20 is microphone 10
An acoustic analysis section performs an acoustic analysis on the signal obtained in the above and extracts the feature amount of the input voice, and 21 is an A/D conversion section that converts an analog signal into a digital signal. 22
23 is a power calculation unit that consists of a squarer and a smoother (not shown) and calculates the power of the input audio. 23 is a power calculation unit that consists of a bandpass filter bank, a squarer, and a smoother (not shown), and calculates the power of the input audio. This is a power spectrum calculation unit that calculates a power spectrum. 24 is the power obtained by the power calculation unit 22;
and a logarithmic transformation unit 25 that logarithmically transforms the power spectrum obtained by the power spectrum calculation unit 23;
a voice section detection unit that detects the voice section of the input voice using the well-known two-threshold method from the power of the input voice obtained in the step;
26 is an LSFL correction calculation unit that performs correction (LSFL correction) using a known least squares error approximation straight line to the power spectrum of the input voice obtained by the logarithmic conversion unit 24;
7 is the voice interval information obtained by the voice interval detection unit 25;
This BTSP calculation unit calculates a known binary time spectrum pattern (BTSP) from the power spectrum of the LSFL-corrected input audio obtained by the LSFL correction calculation unit 26 and uses it as a feature quantity of the input audio. 30 is a pre-registered audio binary time spectrum pattern (BTS
The registered voice storage unit 40 that stores the sound analysis unit 2
0 and the registered voice storage unit 30
This is a recognition processing unit that performs recognition processing based on the feature quantities of the registered speech stored in the , and uses the result as the recognition result of the input speech.

The operation of the logarithmic conversion section 24 in the acoustic analysis section 20 will be explained in detail below. The natural logarithm takes the input as x
(≧0), and the output is y (≧0), then y=logx

It is expressed as (1) and touches the straight line, y=x/e (2) at x=e. Therefore, consider the following function.

[0009]

[Math 2]

A function of this equation is shown in FIG. this function is,
Since it is continuous and differentiable at all points, the converted output also changes smoothly in response to smooth changes in the input. FIG. 3 shows an embodiment in which this conversion method is implemented using a ROM table 50. Another embodiment of the invention is:
An approximate expression is calculated without using a ROM table. Formula (3
) is replaced with an example in which a 12-bit input digital value is logarithmically converted into an output 8-bit digital value. For ease of handling, using 2 instead of the base e of the natural logarithm, formula (3) can be written as

[0011]

[Math 3]

[0012] This function is shown in Figure 4a. Furthermore, the logarithm approximated by a broken line is

[0013]

[Math 4]

[0014] This function is shown in FIG. 4b. Although this function uses a polygonal line, the error from the logarithmic function is small, and the difference in slope between the two polygonal lines at each contact point is small, so there is no problem in practice. FIG. 5 shows a flowchart of this process, and FIG. 6 shows a program written in C language. This program is very simple and runs fast. Furthermore, even if the number of input/output bits is different from this example, it can be implemented in the same way.

In this example, logarithmic transformation is performed only on the power and the power spectrum, but it can be similarly performed on other analyzed quantities related to acoustic analysis that require logarithmic transformation. can do. Also, the section detection section 25
The interval detection method calculates the rise and fall of audio from the difference on the time axis of the power after logarithmic transformation, etc.
Other methods are also possible. Further, the feature amount of the input voice determined by the voice analysis section 20, the feature amount of the registered voice stored in the registered voice storage section 30, the recognition method performed by the recognition processing section 40, etc. may be other methods. Furthermore, as is clear from the above explanation, the basic technical idea of the present invention is to logarithmically convert an electrical signal, and the speech recognition device described above is an example of an application of this logarithmic conversion technology. It is.

[0016]

[Effect] According to the invention set forth in claim 1, since the transformation is performed using a linear function passing through the origin for small inputs, it is possible to handle small inputs without increasing errors. , and for large inputs, a normal logarithmic function is used, and since the linear function and the logarithmic function are connected at a continuous and differentiable point, the output value changes properties rapidly. It is possible to realize a logarithmic conversion method that does not cause such problems. According to the invention described in claim 2, since the conversion is performed using a simple processing method using comparison operations, shift operations, and addition operations on digital signals, logarithmic conversion can be performed at high speed without using a large amount of memory. can be executed. According to the invention set forth in claim 3, voice section detection can be performed using the logarithmically transformed power as set forth in claim 1 or 2, so that it is possible to accurately detect voice sections with low levels. can. In particular, if speech interval detection processing is performed using, for example, the power difference after logarithmic transformation, low-level speech can be treated in the same way as high-level speech with less influence of errors and noise. . According to the invention set forth in claim 4, since the voice feature amount of the input voice can be determined using the power spectrum subjected to the logarithmic transformation according to claim 1 or 2, the error is small in a small band of the spectrum. Become.

[Brief explanation of drawings]

FIG. 1 is a configuration diagram for explaining an embodiment of a speech recognition device according to the present invention.

FIG. 2 is a diagram illustrating an example of logarithmic transformation used in implementing the present invention.

FIG. 3 is a diagram showing an example of a circuit that performs the logarithmic transformation shown in FIG. 2;

FIG. 4 is a diagram illustrating an example of approximate calculation of logarithmic transformation.

5 is a diagram showing an example of a flowchart for executing the approximate calculation shown in FIG. 4. FIG.

6 is a diagram showing an example of a program in which the flowchart shown in FIG. 5 is written in C language.

[Explanation of symbols]

10...Microphone, 20...Acoustic analysis section, 21...A/D
Conversion unit, 22... Power calculation unit, 23... Power spectrum calculation unit, 24... Logarithmic conversion unit, 25... Voice section detection unit, 26...
LSFL correction calculation section, 27... BTSP calculation section, 30... Registered voice storage section, 40... Recognition processing section, 50... ROM.

Claims (4)

[Claims] 1. Acoustoelectric conversion means for inputting voice, acoustic analysis means for acoustically analyzing input voice obtained by the acoustoelectric conversion means, and registration device for storing feature quantities of registered voices registered in advance. It comprises a voice storage means, and a recognition processing means for performing recognition processing using the feature amount of the input voice obtained by the acoustic analysis means and the feature amount of the registered voice stored in the registered voice storage means. In a speech recognition device,
The acoustic analysis means has a logarithmic conversion means, and the logarithmic conversion means converts the input signal based on a logarithmic curve when the input signal is equal to or higher than a predetermined threshold value. is less than or equal to the threshold value, the speech recognition device converts the input signal based on a straight line that connects to the logarithmic curve at the threshold value and passes through the origin. 2. The speech recognition device according to claim 1, wherein the logarithmic conversion means comprises a comparison means for performing a comparison operation on digital signals, a shift means for performing a shift operation on the digital signals, and a shift means for performing an addition operation on the digital signals. 2. A speech recognition device comprising: addition means for performing the approximate calculation of the logarithmic transformation according to claim 1. 3. The speech recognition device according to claim 1, wherein the input signal input to the logarithmic conversion means is the power or amplitude of the speech input. 4. The speech recognition device according to claim 1, wherein the input signal input to the logarithmic conversion means is a power spectrum or an amplitude spectrum of the speech input. Device.