KR940005043B1 - Recognizing method of numeric sound - Google Patents

Abstract

The method improves the effect of numeric voice recognition by frequency spectrum variation. The method employs a microphone (1) which converts voice signal into electric signal, an interface (3) which interfaces eletric signal, a clock analog unit (4) which filters, samples, and outputs electric signal, a clock logic unit (5) which converts filtering and sampling voice signal into digital signal, a digital signal processor (6) which controls total action, an address decoder (7) which decodes data selection signal (10S), input-output selection signal (1IS) and program selection signal (1PS), a buffer (8), reference data RAMs (9)(10), a program ROM (11), an input-output decoding logic unit (12), and a voice recognition controller (13).

Description

Recognition method of digits by spectral change

1 is a block diagram of a numerical tone recognition device according to the spectrum change of the present invention.

2 is a signal flow diagram of a digit recognition algorithm made in the digit recognition apparatus of FIG.

3 is a signal flow diagram for calculating the spectral difference of FIG.

4 is a signal flow diagram for detecting the spectral peak of FIG.

5 is a signal flow diagram for calculating the spectral threshold of FIG.

FIG. 6 is a signal flow diagram of dividing speech into a frame bordering the peak of FIG.

7 is a signal flow diagram for explaining the matching process by the DTW of FIG.

8 is a signal flow diagram for explaining the backtracking process of FIG.

* Explanation of symbols for main parts of the drawings

1: microphone 2: speaker

3: interface portion 4: clock analog portion

6: digital signal processor 7: address decoder

8: Buffer 9, 10: Reference data

11: Program ROM 12: Input / Output Decoding Logic

13: voice recognition controller

The present invention relates to a speech recognition device for recognizing a speech signal by comparing a reference pattern of a database-based speech signal with a section of an input speech signal. Particularly, the speech section is calculated by calculating an amount of change in the frequency spectrum that occurs when the speech region is changed. By dividing, the present invention relates to a method for recognizing a number sound by changing a spectrum so that the number sound can be recognized regardless of the number of spoken number sounds.

The conventional speech recognition apparatus detects the start point and end point of the input pattern of the input voice signal, that is, the first frame and the last frame, and DP matching, i.e., the start point and the end point of the detected input pattern, from the start point to the end point of the reference pattern. In the process of comparing the reference pattern and the input pattern, the difference between the two patterns is obtained by stretching the pattern of the voice signal unevenly distributed on the time axis on the time axis and making the distance between the two patterns smallest. Voice signal is recognized.

However, such a conventional speech recognition apparatus has a defect in that the difference of the unnecessary speech signal becomes large when the endpoint is erroneously detected, thereby misrecognizing the input speech signal.

Accordingly, an object of the present invention is to obtain the central coefficient from the change of the frequency spectrum that occurs when the voice region is changed in view of the conventional defects, and calculate the boundary to find the boundary of the voice interval, regardless of the number of digits spoken. The present invention provides a method for recognizing a number sound by a spectrum change to recognize a signal.

The method for achieving the object of the present invention comprises the steps of filtering and sampling the voice signal to be recognized as a digital signal, and collecting the voice signal of the sampled data in a predetermined number to separate the frame unit; Extracting a feature using a capstrum coefficient for each divided frame, calculating a spectral difference for all the extracted frames, and a frame in which the difference in the spectrum of the calculated frames becomes a peak Detecting a peak value of a peak value of the detected frame below a threshold value, dividing a voice region into segments at a boundary of a frame found to be a peak value above the threshold value; The DTW calculates the distance by comparing the pattern of the segment with the database-based reference pattern. It is described as being made of an accomplished by matching step and the step of the DTW distance result is output to the smallest reference pattern as a recognition result, less detail with reference to the accompanying drawings, the present invention.

에 의해 음성인식장치의 전체 동작을 제어하는 디지탈 신호프로세서(6)와, 상기 디지탈 신호프로세서(6)에서 출력되는 데이타선택신호

, 입출력선택신호

및 프로그램선택신호

를 디코딩하는 어드레스디코더(7)와, 상기 디지탈신호프로세서(6)에서 출력되는 어드레스신호(AD)를 전송하는 버퍼(8)와, 상기 어드레스디코더(7)의 출력신호에 의해 칩선택되고 버퍼(8)에 의해 어드레스가 지정되어 해당 번지내에 저장된 기준데이타(DATA)를 디지탈신호프로세서(6)에 입력하는 기준데이타램(9)(10)과, 상기 디지탈 신호프로세서(6)가 수행할 프로그램을 저장하는 프로그램롬(11)과, 상기 어드레스디코더(7)를 통해 디코딩된 디지탈 신호프로세서(6)의 입출력선택신호

에 의해 상기 기준데이타램(9)(10)으로부터 출력된 데이타(DATA)를 디코딩하는 입출력디코딩로직부(12)와, 상기 입출력디코딩로직부(12)에서 디코딩된 음성신호에 대한 데이타와 디지탈신호프로세서(6)가 인식한 음성신호에 따라 음성을 인식하여 인터페이스부(3)를 통해 스피커(2)로 출력함과 아울러 인터럽트 및 아날로그/디지탈변환 클럭로직부(5)를 제어하는 음성인식제어기기(13)로 구성한다.FIG. 1 is a block diagram of an apparatus for recognizing a number sound according to a spectrum change of the present invention. As shown in FIG. 1, a microphone 1 for converting a voice signal inputted into an electrical signal and a signal input from the microphone 1 are input. An interface unit 3 for processing and interfacing electrical signals for voice signals, a clock analog unit 4 for filtering, sampling, and outputting electrical signals for voice signals input from the interface unit 3, and the clock; An interrupt and analog / digital conversion clock logic unit 5 for converting the voice signal filtered and sampled by the analog unit 4 into a digital signal and generating an interrupt signal, and the interrupt and analog / digital conversion clock logic unit 5. Interrupt signal generated from

A digital signal processor 6 for controlling the overall operation of the voice recognition device and a data selection signal output from the digital signal processor 6

, I / O selection signal

And program selection signal

The chip is selected by the address decoder 7 for decoding the signal, the buffer 8 for transmitting the address signal AD output from the digital signal processor 6, and the output signal of the address decoder 7. 8) reference data 9 and 10 for inputting the reference data DATA stored in the address to the digital signal processor 6 and a program to be executed by the digital signal processor 6; An input / output selection signal of the digital signal processor 6 decoded through the program ROM 11 to be stored and the address decoder 7.

I / O decoding logic section 12, which decodes data DATA output from the reference data 9 and 10, and data and digital signals for voice signals decoded by the I / O decoding logic section 12. Speech recognition control device that recognizes the voice according to the voice signal recognized by the processor 6 and outputs it to the speaker 2 through the interface unit 3 and controls the interrupt and the analog / digital conversion clock logic unit 5. It consists of (13).

FIG. 2 is a signal flow diagram of a number sound recognition algorithm of the number sound recognition apparatus of FIG. 1, as shown in FIG. 1, filtering and sampling a voice signal to be recognized and inputting it as a digital signal (ST100). Collecting a predetermined number of speech signals of the collected data into frame units (ST101), extracting a feature of speech using a cap stratum coefficient for each of the divided frames (ST102), and extracting all of the extracted frames. Calculating a spectral difference for the frame (ST103), detecting a frame in which the difference in the spectrum of the calculated frame becomes a peak (ST104), and excluding the peak value of the detected frame having a spectral difference less than or equal to a threshold value Step ST105 and if the peak value of the frame is found to be a peak value higher than or equal to the threshold value through the step ST105, Dividing the speech region into segments (ST106), comparing the pattern of the divided speech segment with a database-based reference pattern, and matching by DTW to obtain a distance (ST107), and the DTW result. In step ST108, the reference pattern having the smallest distance is output as a recognition result.

Thus, when described in detail with reference to Figures 1 to 8 the effect of the present invention configured as follows.

As shown in FIG. 1, when the connection number voice, which is first uttered when the reference pattern of the voice signal is stored, is input through the microphone 1, the input voice signal is processed through the interface unit 3, thereby providing a clock analog unit. It is entered as (4).

The clock analog unit 4 performs low pass filtering and sampling of the input audio signal and inputs the interrupted and analog / digital conversion clock logic unit 5.

를 출력하게 된다.The interrupt and analog / digital conversion clock logic unit 5 converts the sampled and input voice signal into a digital signal to the digital signal processor 6 for an interrupt signal.

Will print

에 응답하여 인터럽트인식신호

와 클럭신호(XF)를 출력하면 인터럽트 및 아날로그/디지탈변환 클럭로직부(5)는 클럭신호(XF)에 따라 디지탈신호로 변환한 음성신호를 데이타신호(DATA)로 하여 디지탈 신호프로세서(6)에 입력하게 된다.In this state, the digital signal processor 6 provides an interface signal.

Interrupt recognition signal in response to

And outputting the clock signal XF, the interrupt and the analog / digital conversion clock logic unit 5 converts the audio signal converted into the digital signal according to the clock signal XF as the data signal DATA. Will be entered.

The digital signal processor 6 obtains a frequency spectrum by performing Fast Fourier Transform (FFT) on the input digital data and divides the received digital data into frames having a predetermined length to form a reference pattern.

, 입출력선택신호

및 프로그램선택신호

를 출력하여 기준데이타램(9)(10)과 프로그램롬(11)의 칩을 선택하고, 이어서 버퍼(8)를 통해 어드레스를 선택한 후 상기 기준패턴을 데이타라인을 통해 기준데이타램(9)(10)에 저장하며, 이와 같은 동작을 인식하고자 하는 음성신호에 대하여 반복수행하여 음성신호의 기준패턴을 저장하게 된다.At this time, the digital signal processor 6 transmits a data selection signal through the address decoder 7.

, I / O selection signal

And program selection signal

Outputs the chips of the reference data (9) 10 and the program ROM (11), selects an address through the buffer (8), and then applies the reference pattern through the data line to the reference data (9) ( 10), and repeats the voice signal to recognize the operation to store the reference pattern of the voice signal.

On the other hand, in the case of performing a recognition process for recognizing the spoken voice signal, first, when the spoken connection number voice is input through the microphone 1, the input voice signal is processed through the interface unit 3 to clock. The analog unit 4 is input.

The clock analog unit 4 performs low pass filtering and sampling of the input audio signal and inputs the interrupted and analog / digital conversion clock logic unit 5.

를 출력하게 된다.The interrupt and analog / digital conversion clock logic unit 5 converts the sampled and input voice signal into a digital signal to the digital signal processor 6 for an interrupt signal.

Will print

에 응답하여 인터럽트인식신호

와 클럭신호(XF)를 출력하면 인터럽트 및 아날로그/디지탈변환 클럭로직부(5)는 클럭신호(XF)에 따라 디지탈 신호로 변환한 음성신호를 데이타신호(DATA)로 하여 디지탈신호프로세서(6)에 입력하게 된다.In this state, the digital signal processor 6 interrupts the signal.

Interrupt recognition signal in response to

And the clock signal XF are outputted, the interrupt and the analog / digital conversion clock logic unit 5 converts the audio signal converted into a digital signal according to the clock signal XF as the data signal DATA. Will be entered.

The digital signal processor 6 obtains a frequency spectrum by using a capstrand coefficient, and then divides the input speech signal into frames having a predetermined length to create an input pattern, and stores the input pattern and the reference data 9 and 10 in the input signal. By comparing the reference patterns one by one, the voice signal is recognized and the voice recognition control device 13 is controlled according to the recognition result.

That is, referring to FIG. 2 to FIG. 8 specifically, first, as shown in FIG. 2, the digital signal processor 6 includes a clock analog 4 and an interrupt and analog / digital conversion clock logic unit (see FIG. 5) When a voice signal filtered and sampled sequentially and converted into a digital signal is input (ST100), the voice digital data is first divided into frame units (ST101), and features are extracted for each frame (ST102). The difference between the spectra is calculated to check whether the partition boundary conditions are satisfied with the defined characteristics (ST103).

A peak value detection step (ST104) for finding a local peak value of the data thus obtained, a threshold comparison step (ST105) for removing when the spectral difference between the local peak values is smaller than a constant value, and a local peak value finally obtained as described above. Is divided into speech segments using the boundary of the speech section (ST106), and matching step by DTW comparing the speech segment with a reference pattern stored in the reference data 9 (10) to obtain a distance (ST107). And outputting the reference pattern having the smallest DTW distance as the recognition result (ST108).

Each step of the flow will be described in detail with reference to FIGS. 3 to 8 as follows.

First, the step (ST101) and the feature extraction step (ST102) for dividing the input voice signal into frame units are performed by 256 samples (voice intervals) of the voice digital input data filtered, sampled, and converted into digital data (25.6 msec). Collect them into 1 frame.

Thereafter, the feature is extracted for each frame. The feature used at this time is a linear predictive encoding coefficient calculated using a capstrum coefficient, and the linear predictive encoding analysis dimension is set to 12th order.

From the obtained feature vector linear predictive encoding capstrum coefficient, a spectrum difference is obtained through the above step ST103, which is the same as in FIG.

In other words, if the j th frame capstrum vector is C _j , C _j = C _j ¹ , C _j ² , C _j ³ ,. C _j ¹² ), the spectral difference di in the i-th frame is as follows.

로 구한다.

Obtain as

Wherein I = 1, 2,... M is the last frame, CK is the capstem feature vector in the Kth frame, and L is the maximum recognizable number of strings.

As described above, after obtaining the spectral difference di for all input frames, the peak value detection step ST104 is performed.

The peak value detecting step ST104 in which the spectral difference is a local peak is shown in FIG.

That is, if the j th frame satisfies both d _j > d _j-1 and d _j > d _{j + 1} , the frame is determined as the local maximum value and stored in the P array. If j frame is not the last input frame, Repeat by incrementing j frames.

After the above process is performed up to the M-1 < th > frame, the process goes to the threshold comparison step ST105.

The threshold comparison step ST105 is shown in FIG.

In this step, if the spectral difference dk is smaller than the predetermined value Tspec among the frames determined as the local peak value in the step ST104, the P array is removed from the P array. Save it.

After repeating the above process by the determined number of local peak values np, the voice segmentation step ST106 is performed.

The speech segmentation step ST106 determines the final local peak values passed through the threshold value comparison step ST105 as the boundary of the speech and then divides the speech region based on the speech boundary.

Referring to FIG. 6, the input voice signal is divided into voice segments based on the remaining local peak values except for the local peak value having the ∞ value which is a frame, not the local peak value.

After all local peaks, the next step, matching step ST107 and voice recognition step ST108, is performed.

In the matching step ST107, as shown in FIG. 7, the value having the smallest difference is compared by comparing with the reference pattern for the digits from 0 to 9 having the voice segment obtained in the voice segment dividing step ST106 in advance. The process of printing.

In FIG. 7, T _k is a difference value when the first reference value is matched to the k-th segment, B _k ¹ is the last reference value matched when the first reference value is matched to the k-th segment, and C _k ¹ is 1-. The first reference point indicates a matched segment, DTW (I, K) is the difference between DTW of I and k, Iij is the interval from input frame i to j, Rm is the mth reference value, M Is the number of reference values, N is the maximum number of segments that can be matched with the reference value, K is the number of segments that have passed the process up to step ST106, L is the maximum recognizable number of sequences, and Q1... Qn is a recognized string of numbers.

FIG. 8 is a flowchart of a process of finding a recognized sequence of numbers by trace back from the minimum distance obtained through the matching process of FIG.

Since the recognized number is detected and output, the number recognition sound ends.

As described in detail above, the present invention can calculate the change in the frequency spectrum that occurs when the voice region is changed to calculate the center coefficient to recognize the numeric sound without being limited by the number of the number sound generated by finding the boundary of the voice interval. It can be effective.

Claims (1)

Filtering, sampling, and inputting a voice signal to be recognized as a digital signal, collecting a predetermined number of voice signals of the sampled data into frame units, and capturing coefficients for each of the divided frames. Extracting a feature using; and calculating a spectral difference for all the extracted frames; detecting a frame in which the difference in the spectrum of the calculated frames becomes a peak; and among the peak values of the detected frames. Subtracting the spectral difference below the threshold from the peak value, dividing the speech region into segments around a frame found to be a peak value above the threshold, comparing the pattern of the divided speech segment with a reference patterned database; Matching step by DTW to obtain a reference pattern and a reference pattern with the smallest DTW result distance The method for recognizing a number sound by changing the spectrum, characterized in that it comprises the step of outputting as a recognition result.

Images