JPS5995600A - Voice recognition equipment - 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] [Technical Field] The present invention relates to an improvement in a speech recognition device, and more particularly, to an improvement in power normalization when a parameter including power information, such as the output of a filter bank, is used as a feature parameter. is added.

<Prior Art> In a speech recognition device, when a signal containing power information such as a filter bank output is used as a feature parameter, power normalization processing is usually performed to remove the influence of gain.

Conventional power normalization processing methods include a method using an arithmetic mean and a method using a geometric mean, and each can be expressed by the following equations.

However, k: the order of the feature parameter Pi: the i-th parameter before normalization Wi: the i-th parameter after normalization When attempting to perform power normalization according to such a formula, the equation (1) In the method, the right side k
It is possible to expand it to i, eliminating the need for division, and to calculate the logarithmic value nogP i at high speed using a read-only memory (ROM), but it is not easy to handle it with a ROM because it becomes large. .

In addition, in the method of equation (2) above, the logarithm value ogPi is R
Although it can be easily determined using OM, there is a problem that the calculation speed is slow because it includes one division.

If normalization is not performed further, there is a problem in that the influence of the gain is too large to provide a practical device.

Purpose of the present invention The present invention has been made with the purpose of providing a speech recognition device that eliminates the above conventional problems.According to the present invention, the power normalization according to the above formula (2) is performed using division. Instead, it is performed approximately using a shift register, and the power information is retained (thereby improving the recognition rate and increasing the calculation speed.) be done.

<Example> Hereinafter, one embodiment of the present invention will be described in detail.

According to an embodiment of the present invention, there is provided an extraction section for extracting feature parameters, a calculation means for calculating the sum of each channel of the logarithmic output from the extraction section, and a shift for introducing the summation calculated by the calculation means. It is equipped with a register and an arithmetic means for subtracting the output of this shift register from the logarithmic power of each channel described above, calculates the sum of each channel of the logarithm output from the feature parameter extraction section, and shifts this calculated sum. By register, 2'(+1=1.2°・
), the approximate average logarithmic power is obtained, and this average logarithmic power is subtracted from the logarithmic power of each channel to approximately perform power normalization. With such a configuration, it is possible to improve the recognition rate and increase the calculation speed.

Next, one embodiment of the present invention will be described in detail using the drawings.

FIG. 1 is a block diagram showing an example of a speech recognition device embodying the present invention.

The input voice uttered in FIG. 1 is converted into an electrical signal by a detector such as a microphone.

This detector/ is a human audible frequency. 2θ several Hz~
It is preferable to use a device that can detect a frequency of about 7.5 KHz, and in particular can detect frequencies in the human conversation voice range without distortion. An amplifier is connected to the output side of this detector. This amplifier is preferably one that can amplify the audio frequency without distortion. A feature parameter extraction section 3 is connected to the output side of this amplifier 2.
The characteristic parameters of the input voice are extracted.

The feature parameter extraction unit 3 includes a plurality of band filters having different passbands, a sample hold circuit that holds the input and output of each band filter, and sequentially samples the output of the sample hold circuit at intervals of about 10 milliseconds. It consists of an analog switch, a % converter that converts the output of the analog switch into a 7.2-bit digital signal, and outputs a characteristic parameter Pi, and a logarithmization ROMJ/ that logarithmically converts the characteristic parameter Pi.

The feature parameter Pi output from the 7% converter in the feature parameter extraction section 3 is logarithmically transformed by the logarithmization ROM 3/, and this logarithmically transformed value is used in the power normalization section and pattern compression section 5 in the next stage. The normalized feature parameters and the standard feature parameters stored in the standard pattern memory B are compared by the matching section Z to recognize the input speech, and the result is output to the output section 2. It is configured as follows.

As shown in Fig. 1, the power normalization section calculates the logarithmic transformation amount of the conversion output Pi in the feature parameter extraction section 3.
Calculate the sum of each channel of gPi ±1 The output (sum) of the adder g/ and this adder 41/ m The input shift register of 70 gPi 1 The output of iit62 and this shift register g is the logarithmic power log of each channel ogPi and a subtracter Z3 that subtracts from.

In the above configuration, the parameter Pi output from the converter in the feature parameter extraction unit 3 is stored in the logarithmization ROM.
3/, the sum of all channels of the output is calculated by the adder, and the sum is input to the shift register 4t, 2.
Then, the sum is shifted to the lower n bits and subtracted from the output fflogPi of each channel by the subtracter G3, thereby creating a power-normalized feature parameter Wi.

That is, according to the present invention, the following equation (
Normalization according to 3) is performed.

However, 1. , -, n (n=1, 2, -) where, i,
Since is an operation of the factorial of -, the above shift register lt
This shift operation allows calculations to be performed quickly and easily.

Also, convert the /(/ term in equation (3) above to -(-±□) kl' to '2 to'3. However, k'i =, 2'' (rr=1, 2, -) , it is also possible to increase the accuracy.

For example, '1=, 2, 2=, 2, '3=
, 2', the output of the adder/is input in parallel to the 3-bit and Z-bit shift registers, the output is added to the adder/subtractor, and the calculation result is input to the 2-bit shift register. It is possible to calculate the correction term for .

Effect> As described above, according to the present invention, the sum of each channel of the logarithmic output from the feature parameter extraction unit is calculated.1. By shift register, 2″″n (n=1゜2,...)
By multiplying the approximate average logarithmic power, and subtracting it from the logarithmic power of each channel, power normalization can be approximately performed at high speed without involving division processing. Recognition considering power information can be performed at high speed and relatively easily.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of a speech recognition device embodying the present invention, and FIG. 1 is a block diagram showing the configuration of its main parts. 3... Feature parameter extraction unit, G... Power normalization unit, 3/... Logarithmic conversion ROM, G/... Adder, G... Shift register, G3... Subtractor, Pi...
- i-th feature parameter before normalization, Wi... i-th feature parameter after normalization. Agent Patent attorney Aihiko Fukushi (and 2 others) 575 1 Patent application for indication of the case 1987-207178 2 Name of the invention Speech recognition device 3 Person making the amendment Relationship to the case Patent applicant address 108 Tokyo Port Address: 6, 22-22 Nagaike-cho, Abeno-ku, Osaka-shi, Osaka 545 Subject of amendment 7, contents of amendment (1) The description of "Kara" in line 20 of page 1 of the specification has been corrected to "improvement"1. (2) Page 7 of the same book! Change the description of “factorial of 2” on line 18 to “
I have corrected it to ``a power of 2.''that's all

Claims (1)

[Claims] 1. An extraction unit that extracts feature parameters, an arithmetic unit that calculates the sum of each channel of logarithmic output from the extraction unit, and a shift register into which the sum calculated by the arithmetic unit is introduced. and arithmetic means for calculating the output of the shift register from the logarithmic power of each of the channels, the speech recognition device being configured to normalize the power of the feature parameter.