JPS61175700A - 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 speech recognition device, and more particularly to a speech recognition device that reduces the speech matching processing time of a speech registration type recognition device.

[Prior art] This type of conventional speech recognition device sequentially superimposes input speech and all standard speech patterns registered in advance, calculates the distance between them each time, and selects the standard pattern with the smallest distance. was extracted and used as the speech recognition result. Therefore, as the total number of registered words increased, the matching processing time increased, and the phenomenon of a decrease in the recognition rate became noticeable.

In order to avoid this, for example, standard phonetic words can be classified and registered according to concepts such as words, numbers, monosyllables, etc., and when recognition is performed, it is possible to select the memory section in which the various quantity groups are stored. A method that performs strict matching processing is considered effective.

There are two ways to select and change the various quantity collective storage units, such as by key operations and by voice. Although the selection and change of various quantity collective storage parts can be carried out reliably by key operation, the operation is complicated and the burden on the user is heavy because key input and voice input must be performed at the same time.

In addition to the original registered voice pattern group, the voice method requires a command to select and change the memory of the registered voice pattern group, and for this purpose, a voice with a name representing each preword collective memory is required. A separate storage unit for the entire group of changed words consisting of patterns must be provided; that is, the original speech pattern is divided into several memory units for the entire group of words depending on the characteristics of the word, and each storage unit has a command for changing the word. As, for example, the selection of the 0-word orchid collective memory that incorporates "rhencora"
When changing, first input "Henkora" by voice, and it will be recognized in the various quantity group memory that was selected at that time, and this will select the memory of the entire group of changed terms. Selection/change can be made by vocally inputting a name representative of the entire group memory of the word desired to be selected/changed. However, this method
It is time-consuming because it requires voice input twice to make changes. Furthermore, since the voice pattern of the above-mentioned change command is registered for each word group storage unit, voice patterns of the same word but with different utterance volume levels and utterance durations are registered. Even if you make a change to the selection contest, the way it is perceived will be different, so there is a worst case scenario lol
In some cases, it became impossible to select the entire group of changed terms from the 1-millet group memory.

[Objective] The present invention has been made in view of the above-mentioned drawbacks of the prior art, and its purpose is to improve speech recognition by incorporating utterance time information of input speech into a speech recognition method. An object of the present invention is to provide a speech recognition device capable of shortening matching processing time and obtaining a high recognition rate.

[Embodiments] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

Fig. 1 is a block diagram of a speech recognition device according to an embodiment of the present invention. In Fig. 0, 1 is a microphone that converts speech into an electrical signal, and 2 is a band pass that divides the frequency range of 200 to 6000 Hz into 8 to 30 channels. The feature extraction unit consists of a group of filters and extracts features such as power signals and formant signals. 3 is an A/D converter that samples and quantizes the extracted feature cover every 5 to 10 m5, and 4 and 14 are registration/recognition switches that switch the signal path between standard voice registration and input voice recognition. Switch means 5 and 12 are buffer memories for storing input voice features until the utterance time length of the input voice is calculated during registration or recognition; 6 is a buffer memory for storing input voice features from the power signal of the input voice; A start/end detection circuit detects a point corresponding to the end; 7 is a utterance length measurement circuit that measures the time from the start to the end; 8 is a utterance length measuring circuit that detects and measures the utterance duration; The utterance length selection circuit that generates the selection signal, lO, is the collective memory unit 10□~
ion memory, 9 is a word variation group storage unit 10. - 10n, 11 is a switch that switches various quantity group storage units 101 to 10n during speech recognition, and 13 is a switch that switches and reads input speech patterns and vocabulary group storage units 101 to 10n during speech recognition. 15 is a general-purpose central processing unit (CPU) that processes recognition results; 16 is an operation keyboard; and 17 is a display unit that displays recognition results and the like.

The operation of the embodiment will be described in detail below.

First, the utterance time length of the input voice is determined by the time difference between the start and end of the input voice. Various methods can be considered for detecting the start and end of audio, but in this embodiment, power information P after A/D conversion is used.

FIG. 2 is a diagram showing power information P of the input voice output every 5 to 10 m5 on a time axis. First, in order to remove background noise mixed into the input voice, the average value of the noise power in the laboratory is calculated in advance, and this is set as the threshold value pN. Further, a threshold value for the level of a word-initial consonant that is easily devoiced or a word-initial consonant with low power is set as Pc, and the average value of the two thresholds PPJ and pc is set as PM. Also, let TP be the minimum pause time from when there is an input voice until the next input voice, and let TW be the minimum utterance time recognized as input voice.

[Detection of starting point SO] Find the first point where the power signal P output every 5 to 10 mS satisfies P≧PM. If the state where P≧PM continues for Tw or more after this point, the first point where P≧PM becomes the starting point SO. If it ends up being less than Tw, it is regarded as noise, the next point where P≧PM is found, and the same operation as above is performed.

[Detection of end point Eo] After detecting the start point SO, find the first point where the power signal P satisfies P < PM. If the state of P<PM continues for 12 hours or more after this point, the first point at which P becomes PM is set as the end point Eo.

In this way, the start end φ end of the input voice is detected, and the utterance length measuring circuit 7 starts the timer when the start end So is detected by the start end/end end detection circuit 6, and starts the timer when the end end Eo is detected and output. Stop and calculate the duration of vocalization,
The value is sent to the utterance length selection circuit 8.

Incidentally, the above-mentioned operation can be realized by a microprocessor having a built-in control program shown in FIG. Step S
1 initializes the timer t to 0. In step S2, the process waits until the power signal P becomes equal to or higher than PM. If P≧PM is satisfied as determined in step S2, the process advances to step S3 and the contents of timer t at that time are saved in the start register So. In steps S4 and S5, the process waits until the state of P≧PM continues for a time Tw or more. If P≧PM is not satisfied during the process, the process returns to step S1 and the portion up to that point is treated as noise. If the process continues for 7W or more, the process advances to step S6 and the contents of the start end register S0 are finalized. In step S6, P<
If it is determined in step S6 that P<PM is satisfied, the process advances to step S7, and the contents of timer t at that time are saved in the rear end register EO. Steps S8 and S
In O, wait until the state of P < PM continues for 12 or more times, and if P < P y is not satisfied on the way, proceed to step S.
6, and the power signal P up to that point is treated as valid. If the process continues for 12 or more times, the process advances to step SlO and the contents of the rear end register E0 are finalized. In step SIO, the interval from time So to Eo is set as utterance length v1.

Next, the storage structure of the memory 10 will be described. Table 1 shows the storage structure adopted in this embodiment. The apparatus has various quantity group storage sections 101 to 101 classified on the utterance time length side. In this case, quantities having utterance time lengths of C145 to 3S are used as words to be recognized, and the storage units for each quantity group are arranged in parallel at time intervals of 0.2S.

Margin table below 1 When registering the standard voice, each contact C of the switch 4.14 is connected to the contact 41 side and the contact 141 side, respectively, as shown in FIG. When the utterance length Vl of the input voice is sent to the utterance length selection circuit 8, the words IJ# group storage unit 1 are classified according to the utterance length.
It is converted into a selection signal of 01 to Ion. The selection signal is sent via the contact 141 of the switch 14 to the word-real-word-actual-collection-storage unit changeover switch 9 for registration, where the corresponding quantity group storage unit is selected. The characteristic pattern of the voice stored in the buffer memory 5 (for example, the portion from SO to Eo) is stored as a standard pattern in the word search group storage unit selected in this way. In this way, speech patterns of various utterance lengths are stored in word cluster storage units determined for each utterance length.

Further, during speech recognition, each contact C of the switch 4.14 in FIG. 1 is connected to the contact 42 side and the contact 142 side, respectively. Therefore, the selection signal outputted from the utterance length selection circuit 8 is sent to the recognition word group storage unit changeover switch 11 via the contact 142 of the switch 14, and the word group memory unit corresponding to the detected utterance length Vl is selected. Next, the selected standard patterns in the word collection storage section are sent one by one to the pattern matching section 13, where they are pattern matched with the characteristic patterns of the input speech in the buffer memory 12, and the two are matched. The standard pattern with the highest degree of similarity is extracted, and its corresponding code is output to the CPU 15 as a recognition result.

Next, the above-mentioned operation will be explained according to a specific example. First, at the time of registration, for example, a certain word A is input, its voice feature amount is stored in the buffer memory 5, and the occurrence time length is calculated as 0.855 by the start/end detection circuit 6 and the utterance length measurement circuit 7. In the zero utterance length selection cycle 8, the word fact group storage unit 103 is selected from this time information 0.853 according to Table 1, and the feature pattern of word A in the buffer memory 5 is registered in the storage unit 103. During voice recognition, the storage section 103 is selected by the switch 11 in a similar operation, and the storage section 103 is selected by the switch 11.
Matching is sequentially performed between the plurality of standard patterns in the buffer memory 12 and the characteristic patterns of the word A stored in the buffer memory 12.

By the way, even if the word is the same, if the utterance time at the time of standard pattern registration and the utterance time at the time of voice recognition are different, the desired expense group storage unit may not be selected at the time of voice recognition.For example, in Table 1, the word B's utterance length at registration is 0.795
5, and the utterance length at the time of recognition is 0.85, word B is registered in the storage unit 102, and recognition matching is performed between it and the standard pattern in the storage unit 10a. will no longer be recognized. Therefore, in order to avoid problems caused by variations in utterance length, the utterance group memory section is selected using utterance time information that takes into account the true utterance of the input word during recognition and a predetermined variation range for the word length. If we consider the fluctuation range of ±0.015 and add it to the true utterance length of 0.85, the utterance length of word B becomes 0.799 to 0.8.
015. These units are storage unit 102 and storage unit 103.
Therefore, matching is first performed using the standard pattern in the storage unit 102, and then matching is performed using the standard pattern in the storage unit 103. On the other hand, if the utterance length at the time of registration is 1.059 for word C and the true utterance length at the time of recognition is 1.1O3, the utterance at the time of recognition takes into account the fluctuation values of ±o and ois. The length is also the same storage unit 10 as at the time of registration.
Since this is within the word group No. 4, matching can be performed within that group. In this way, it is possible to provide a speech recognition device that is resistant to variations in utterance length. Comparing the results of recognizing 500 words using the device of this embodiment with the conventional method, the recognition processing time was shortened by 100 to 500 m5, and the recognition rate was 20 m5.
% or more, with an average recognition processing time of 280 m5 and 98
．． A recognition rate of 5% was obtained.

In this embodiment, the PR was determined using background noise in a laboratory, but the value of PM can be freely varied under any noise environment depending on the purpose of the speech recognition device.

In addition, the number of classifications in the collective storage section, the capacity of each storage section, the duration of utterances, the range of variation in the length of utterances taken into account during recognition, etc. can be freely varied depending on the application so as to always obtain the best recognition results. . In addition, by applying the present invention, high speed,
A highly reliable voice typewriter can be constructed.

[Effects] As described above, according to the present invention, it is possible to shorten the time for speech recognition processing by adding utterance time information to the speech feature amount. In other words, pattern matching candidates during speech recognition are narrowed down to small groups based on utterance time information, so even if the overall registration costs are large, the recognition processing time is short, and one utterance time information is also important for speech recognition. Since this is nothing but information, using this for recognition processing is also effective in improving the recognition rate.

[Brief explanation of the drawing]

FIG. 1 is a block configuration diagram of a speech recognition device according to an embodiment of the present invention. FIG. 2 is a diagram showing input speech power information P on a time axis. FIG. 3 is a flowchart showing utterance length measurement processing according to an embodiment. It is. Here, 1...microphone, 2...feature extraction unit,
3... A/D converter, 4... Registration/recognition changeover switch means, 5... Registration buffer memory, 6... Start/end detection circuit, 7... Utterance length measurement circuit, 8 . . . Utterance length selection circuit, 9 . . . Miscellaneous expense group changeover switch means for registration, 10 . ,
DESCRIPTION OF SYMBOLS 12... Recognition buffer memory, 13... Pattern matching section, 14... Registration/recognition changeover switch means, 15... CPU, 16... Keyboard, 17...
・It is a display section. Patent applicant: Canon Co., Ltd. Figure 2 - Ml'1 Figure 3

Claims (2)

[Claims] (1) utterance length detection means for detecting the utterance length of input speech;
Registered voice storage means for storing voice standard patterns grouped according to utterance length, and registered voice reading means for reading out voice standard patterns of corresponding groups from the registered voice storage means based on the utterance length output from the utterance length detecting means. A speech recognition device comprising: a speech recognition device which sequentially compares the speech standard pattern read by the registered speech reading device and the input speech pattern to perform speech recognition. (2) The registered speech reading means determines the corresponding group based on the utterance length detected by the utterance length detecting means and the value obtained by adding a predetermined fluctuation range to the utterance length.
Speech recognition device described in section.