JPS60140300A - Nasals identifier for 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 a nasal identification device in a speech recognition device. The nasal sound identification device according to the present invention is used, for example, in a speech recognition device in an information processing device such as a word processor.

Prior Art Conventionally, in a speech recognition device, the speech waveform 'f' is sampled at intervals of about 10kHz and converted from analog to digital.
Data within a window (referred to as a frame) with a width of approximately 10 m from the starting point of the sampled data is subjected to frequency analysis while moving the window, and features are extracted. In this way, when frequency analysis is performed within a narrow window with a window width of about t o m see, the time resolution Δt is high, but the product of the time resolution Δt and the frequency resolution Δf (Δt
Since there is a mathematical principle that XΔf) cannot be made smaller than a certain value, the frequency resolution is low.

Therefore, when performing speech recognition using such a narrow window, it is possible to identify that the target speech waveform is a nasal sound, but the consonant part "m",
Discriminating whether η is either "n" or "n" results in a poor discrimination rate because a sufficiently high frequency resolution necessary for the discrimination cannot be obtained.

OBJECTS OF THE INVENTION The purpose of the present invention is to perform frequency analysis on the entire speech section (nasal section) that is identified as a nasal sound, obtain a spectrum with high frequency resolution, and analyze the nasal sounds 1m'', "n'", ′
Based on the concept of clarifying the difference in the frequency structure of ``1'', the objective is to improve the identification rate for identifying whether the nasal sound to be recognized is m'', ``n'', or 'u゛.

Structure of the Invention In the present invention, there is provided an audio data buffer that holds an input audio signal, a frame unit recognition unit that recognizes audio data in the audio data buffer frame by frame, and a nasal section that is determined from the recognition result of the frame unit recognition unit. a nasal interval window operation for performing a window operation to remove the influence of vowel intervals on both sides of the nasal interval on the audio data in the audio data buffer using the nasal interval information detected by the nasal interval detection means; A nasal sound identification device in a speech recognition device is provided, which includes a nasal sound recognition means that performs frequency analysis, feature extraction, and matching on speech data subjected to window calculation by the nasal sound interval window calculation means.

Embodiment of the Invention A nasal sound identification device as an embodiment of the present invention is shown in FIG. In FIG. 1, microphone l is connected to audio input section 2.
Connected to the 0 input terminal. The audio input section 2 samples the audio signal S input from the microphone/lophone 1 at regular intervals, performs analog-to-digital conversion, and outputs the resulting audio digital data to the audio data buffer 3.

The audio data buffer 3 is a memory that stores the audio digital data, and its output is led to the input terminal of the frame unit recognition section 4 and the multiplication circuit input terminal of the nasal interval window calculation section 6, respectively.

The frame unit recognition unit 4 is a frame unit recognition device of a known type that recognizes an input speech waveform in units of frames (windows) of, for example, 10 m sec, and includes a frequency analysis unit, a feature extraction unit, a dictionary unit, and a matching unit. It is composed of such things as

The output as the recognition result of the frame unit recognition unit 4 is sent to one input terminal of the identification result integration unit 5 and to the nasal segment detection unit 6.
0 input terminal, respectively.

The nasal segment detecting unit 6 detects the nasal segment by detecting the start and end points of the frame sequence determined to be nasal in the frame unit recognition unit 4, and its detection output is sent to the window of the nasal segment window calculating unit 7. It is guided to the input terminal for the function calculation circuit. The nasal interval window calculation unit 7 converts a window function W using the nasal interval detected by the nasal interval detection unit 6 into the audio data buffer 37.
This circuit multiplies the audio data read out from 7) and extracts audio data in the nasal section from the audio data.

A frequency analyzer 8 and a feature extractor 9 are sequentially connected to the output side of the nasal interval window calculation unit 7 = the output of the feature extractor 9 is connected to the matching unit 11 and the nasal dictionary 1 via a switch IO.
2 respectively. Switch IO is recognition side terminal 1
0a and a registration side terminal 10b, and a recognition side terminal 1.
A matching unit 11 is connected to 0a, and a nasal dictionary 12 is connected to the registration side terminal 10b.

By switching the switch 10 to the registration side terminal 10b and deriving the feature pattern from the feature extractor 9, the nasal dictionary 12 extracts the features corresponding to the nasal sounds "m" and "n-ji'.
The pattern is registered in advance. Matching section 11
When the switch 1o is switched to the recognition side terminal 10a, the recognition pattern of the feature extraction unit 9r and others is compared with the registered pattern of the body sound dictionary 12, the distance between the two is calculated, and the distance between the two is calculated. The category name of the smaller dictionary (
That is, either "m" or n-゛η'' is output.

The output as the identification result of the matching section 11 is led to the other input terminal of the identification result integrating section 5.

The identification result integrating section replaces the nasal part of the recognition result of the frame unit recognition section 4 with the output from the matching section 11 and outputs it as the identification result.

One specific side of the nasal segment detection unit 6- and the nasal segment window calculation unit 7 in the device shown in FIG. 1 is shown in FIG.

In FIG. 2, the recognition result from the frame unit recognition section 4 and the clock CLK (1) synchronized with the frame period are led to the nasal section detection section 6, where the recognition result is clocked. The signal is guided to the shift register 61 in synchronization with CLK(i).

The shift register 61 has registers 6 connected in cascade in two stages.
Consisting of 1a and 61b. The register 61a holds the recognition result of the current frame (the frame of interest at present), and the register 61b holds the recognition result of the previous frame.

Each output of registers 61a and 61b is sent to comparator circuit 6.
2.63, respectively. The other input terminals of the comparison circuits 62 and 63 receive the nasal sound M resulting from the recognition.
A character code corresponding to N II is derived.

The output of the comparison circuit 62 is led to one input terminal of the AND circuit 67 via the NOT circuit 65 and to one input terminal of the AND circuit 66. It is led to the other input terminal of AND circuit 66 and to the other input terminal of AND circuit 67.

These NOT circuits 64, 65, AND circuits 66, 67'C
The logic circuit consisting of ``performs a logical operation to determine the starting point and ending point of the nasal interval.

The outputs of the AND circuits 66 and 67 are then led to the set (S) input terminal and reset (R) input terminal of the nasal interval detection section 70 and the solve flop 71, respectively, and the output Q of the flip-flop 71 is a window function. It is led to a data input terminal of memory 72. In this window function memory, a counter 73 is connected to its pledge address input terminal, a counter 74 is connected to its read address input terminal, and AND circuit groups 76a, 76b, . . . , are connected to its data output terminal. .

The counter 73 has a clock CLK(1) introduced to its input terminal, and provides a count value obtained by counting the clock CLK(1) to the window function memory 72 as a write-in address.

Further, the clock CLK (2) is introduced to the input terminal of the counter 74 via an M frequency divider circuit 75. Clock CL
K (2) is synchronized with the sampling period of the audio input,
There are M clock signals per frame. Therefore, the counter 74 receives a signal C obtained by dividing CLK (2) by the data θM per l frame by the M frequency dividing circuit 75.
LK (3) is input and counted, and the counted value is given to the window function memory 72 as a read address. In addition, these flip-flops 71° window function memory 7
2. Counters 73, 74, and M frequency divider circuit 75 constitute a window function calculation circuit.

AND circuit groups 76a, 76b...76n constitute a multiplication circuit, each bit of the digital audio data from the audio data buffer 3 is led to each input terminal, and each output is the frequency analysis section 80
led to the input terminal.

The operation of the apparatus of FIGS. 1 and 2 will now be described with reference to FIG. FIG. 3 is a diagram of signal waveforms at various parts in the apparatus shown in FIGS. As a result of recognition, (3) is the waveform of the window function W in the nasal interval window calculation circuit 7, and (4) is the clock CL.
The waveform of K(1), (5) is the waveform of clock CLK(2), and (6) is the waveform of signal CLK(3)+7J, which is the clock CLK(2) divided by M.

Audio signal S input from microphone l (Figure 3A)
is sampled by the audio input section 2 at the sampling timing of the clock CLK (2), converted into digital data, and stored in the audio data buffer 3. The frame unit recognition unit 4 refers to the audio data buffer 3 and obtains recognition results in frame units determined by the clock CLK(1). Figure 3 (2) shows the rAN of the audio signal.
It shows how OJ is recognized frame by frame.

The nasal segment detecting unit 6 detects the start and end points of the frame sequence of the nasal sound "N" from among the recognition results shown in FIG. 3(2), and determines the nasal segment. The nasal segment window calculation unit 7 calculates the now determined nasal segment 1! of the audio data in the audio data buffer 3! I
Install a window as shown in Figure 3 (3). The detailed operation of the nasal interval window calculating section 7 will be explained below with reference to FIG.

In FIG. 2, the recognition results of the frame unit recognition unit 4 are as follows:
The signals are sequentially footed into the shift register 61 in synchronization with the clock CLK(1). Register 6.1a contains the recognition result of the current frame of interest, and register 61b also contains the recognition result of the current frame of interest.
holds the recognition result of the previous frame. The recognition results of the current frame and previous frame held in these registers 61a and 61bK are then sent to comparison circuits 62 and 63, where the nasal sound 'N' is output.

The starting point and ending point of the nasal sound are determined by the following logic.

The starting point of the middle nasal sound is determined by the fact that the current frame is the nasal sound "N" and the previous frame is not the nasal sound "N".

(11) The end point of the nasal sound is determined by the fact that the current frame is not the nasal sound 'N'' and the previous frame is the nasal sound II N11.

The logical operations in (1) and (11) above are performed by the comparator circuit 62.6.
3. It is realized by NOT circuits 64.65 and AND circuits 66.67, and depending on the calculation result, the AND circuit 66 outputs the nasal starting point detection signal S(a), and the AND circuit 67
From there, the nasal end point detection signal S (e) is output to the nasal interval window calculation unit 7.

In the nasal interval window calculation unit 7, a flip-flop 71
is set by the start point detection signal 5 (8) and reset by the end point detection signal S (e) to generate the window function W shown in FIG. 3 (3). The data is stored in the memory 72. The processing address of the window function memory 72 is based on the clock CLK (1) by the counter 73.
Obtain it by counting.

On the other hand, the audio waveform data D from the audio data buffer 3
is the AND circuit group 76a in synchronization with the clock CLK(2).
, 76b...76TL, and the window function memory 7
The window function read from 2 and this AND circuit group 76a,
76b...76n and output to the frequency analysis section 8. The read address of the window function memory at this time is
The clock CLK (2) is divided by the number of data M per frame by the M frequency dividing circuit 75, and the signal CLK obtained by dividing the clock CLK(2) by the M frequency is obtained.
(3) counted by the counter 74 is used. Therefore, the form shown in FIG. 3(3) is applied to the portion of the audio data that corresponds to the nasal section, and the influence of the vowel sections on both sides of the nasal section is removed.

The frequency analysis section 7 in FIG. 1 performs frequency analysis of the nasal interval windowed data, and the feature extraction section 8 performs band power calculation based on the spectrum obtained by the frequency analysis section 7.
Performs feature extraction such as moment calculation. In this case, frequency analysis is not performed for each unit frame, but is performed using the entire interval determined as a nasal interval as a window, so that the nasal sounds n” and ′m−゛η゛ can be identified with a high discrimination rate. Sufficient frequency resolution can be obtained.

The feature parameters of the feature extraction section 9 are sent to the matching section 11 via the switch 10 which is switched to the recognition side terminal 10a. In the matching sound tstt, m゛°, ″'n−“°, which are registered in advance in the ascending sound dictionary 12, are used.
The distance is calculated between the pattern of η゛′ and the feature parameter from the feature extraction unit 9 to be recognized, and the category name of the dictionary with the smallest distance, that is, m゛,′″ln
″, °′η”. Identification result integration unit 5
Now, the nasal part of the recognition result from the frame unit recognition unit 4 is replaced with the g recognition result from the matching unit 11, and the recognition results are integrated and output to a subsequent processing step.

Effects of the Invention According to the present invention, the difference in the frequency structure of the nasal sounds "m", n", and ゛)J" can be clearly distinguished, and whether the nasal sound to be recognized is m", n--, °" The identification rate for identifying whether the speech is true or not is improved, thereby improving the performance of the speech recognition device.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing a nasal identification device in a speech recognition device as an embodiment of the present invention, and FIG. 2 is a block diagram showing the configuration of a nasal segment detection unit and a nasal segment window calculation unit in the device shown in FIG. , FIG. 3 is a waveform diagram showing signal waveforms in each section of the first section 1 and the device shown in FIG. 2... Audio input section, 3... Audio data buffer, 4
...Frame unit recognition unit, ...Identification result integration unit, 6...Nasal section detection unit, 7
... Nasal interval window calculation section, 8... Frequency analysis section, 9
...Feature extraction section, ll...Matching section, 12.
...Nasal dictionary, 61...Shift register, 62.63
... Comparison circuit, 71 ... Flip-flop, 72
...Window function memory, 73.74...Counter, 7
5...M frequency dividing circuit. Patent applicant Fujitsu Ltd. Patent application representative Patent attorney Akira Aoki Patent attorney Oro Nishi Patent attorney 1) Yukio Patent attorney Aki Yamaguchi] Simplified

Claims (1)

[Claims] 1. An audio data buffer that holds an input audio signal, a frame-by-frame recognition unit that recognizes the audio in the audio data buffer frame by frame, and a nasal interval detection unit that detects a nasal interval from the recognition result of the frame-by-frame recognition unit. , nasal interval window calculation means for performing a window operation on the audio data in the audio data buffer to remove the influence of vowel intervals on both sides of the nasal interval using the nasal interval information detected by the nasal interval detection means, and the nasal interval window calculation means; Frequency analysis is performed on the audio data subjected to window calculation by the interval window calculation means,
1. A nasal recognition device in a speech recognition device, characterized by comprising nasal recognition means for extracting features and performing matching.