JPS6052440B2 - voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to improvements in speech recognition or speaker recognition;
In particular, the process is performed in two stages to ensure the recognition results.

With the use of computers and the development of voice response devices (devices that output the results of computer processing in the form of voice), information service systems using telephones as terminals began to be put into practical use.

In that case, speech recognition and speaker recognition are required for the following reasons. 1) As a means of inputting information, use the PB (push button) 1--re, L'nT)
It is still low (in Japan), and no alternative keyboard acoustic coupler has been approved.

For this reason, in most terminals (in general households) that use ordinary rotary dial telephones, the only means of inputting information is by voice. 1. Need for voice recognition (2) In information services that involve the exchange of money or that are confidential to individuals (membership-based product sales, pari-mutuel voting, checking deposit balances, etc.), the terminal user must be a member (or It is necessary to confirm whether the person providing the information is the person providing the information. Necessity of single-speaker recognition Judging from the current level of speech recognition technology, only a single number can be recognized almost reliably (recognition rate of 99% or more) in real time: digit sound 1 of a specific speaker.
There are about 0 types, and it must be considered that the recognition rate will decrease considerably for other types. Also, in speaker recognition, although voices with the same utterance content as the registered voice of a registered speaker can be identified fairly reliably, it must be considered that errors will increase under other conditions. Furthermore, considering that the target is audio transmitted over a telephone line, it must be considered that correct recognition will become even more difficult due to the effects and changes in line characteristics. The present invention provides improved means for correctly recognizing voices or speakers under such difficult circumstances.

In order to make the content of the invention easier to understand, a specific example will be given and explained.

As a specific example, consider a bank deposit balance inquiry (inquiry).

Users wishing to receive this service are connected to the bank's balance inquiry center by dialing a specific telephone number preset by the bank. Once the line connection is established, the computer at the center will instruct you to enter your bankbook number (a certain number of digits) from the voice response. speak the number sounds). The center recognizes the numbers from this voice, determines the number, matches it with the center's file, and notifies you of the balance of the deposit.
In the current system, if the number is correctly recognized, the balance will be reported without confirming the user of the terminal, resulting in a Freeaccess system where the number is the only key. Naturally, due to personal or corporate confidentiality, there may be a request that no one other than a specific individual (for example, the owner of the deposit or the person in charge of management) should know the contents, and depending on the user's wishes, this may not be possible. This system cannot be put to practical use unless it has such a secret-keeping function. To do this, it is necessary to recognize the speaker from the numeric sound itself, or to request input for personal identification, such as a name, and to confirm the speaker based on this. At this time, the target of the second recognition (identification of the speaker) can be limited by using the recognition result of the deposit number, which is the key of the first recognition, to efficiently perform effective confirmation.

The first problem with this system is how accurately the first number recognition is performed.

When it comes to this recognition, it must be considered that the accuracy of the recognition will be low because the speaker cannot be known in advance. It is certainly possible to assume that the speaker is one of the registered speakers, so in principle it is possible to compare all of them and produce a recognition result. However, in reality, the number of registrants is large and it is difficult to do so in terms of time. Therefore, we must assume that at least the first recognition can be quite error-prone. When the accuracy of the recognition result is low, all this is done by R.
If you decide not to accept it as eject,
The t rate becomes too high, making the system impractical. On the other hand, if the answer is accepted as correct even though it contains an error, the balance of other deposits will be revealed, which is undesirable.

Therefore, in the present invention, the certainty of recognition (left degree) is calculated for each digit, and all numbers whose k degree (likelihood) is greater than a certain value (preset) are pre-registered as candidates. , and then performs a reconfirmation process on it. Now, an N-digit number is specified as the bankbook number, and if each digit has Ml9m29゜゜9mi9value゜゜゜mN numbers, then (Mj, ku10, j=1,2,...,N) The number of candidate digit strings is M=. 〃M, which is sufficiently smaller than 1D.

1=1 The reconfirmation method is as follows: (1) Assuming that the i-th digit is I (one of the candidate digits), reconfirm by searching for its feature in the input pattern.

If it cannot be reconfirmed, it is not I. (2) Assuming that the i-th digit is I, recognition is performed using a feature by another feature extraction means, the result is compared with I, and if they are not equal, it is determined that I is not true.

(3) Read the registered digit sound pattern of the virtual number (one out of M) from the individual's registered digit sound file, compare it with the input, collate it, and reconfirm it. Various methods such as

However, in order to use the method (3), it is necessary to have a registered file (voice pattern) of the audio deposit number for each individual, but it can be said to be a reliable means of confirmation. After these two steps of confirmation, a final number is determined and a voice response is sent to the user for confirmation.

The user listens to the result and says “Hi/No” or “R”.
Enter yes/noj and proceed to the next process. The number of digit strings to be confirmed is M for each digit, so the total number is N-M, but since each digit is successively narrowed down to one correct result (a most probable number), the total number is N-M. Check number is Mlm2lllmN+M2m
3l●●●●MN+e●●●●●＋●●●●◆●＋MN
=M(1+1/m1+1/Mlm2+...+1
/Mlnl2IOIOlmN-1) Ku? becomes.

Furthermore, if we first apply the method (1) or {2) above to further narrow down the number of candidates (for example, to 2 or less for each digit), and then apply the method (3), all patterns (M
(can be made smaller).

In addition, by processing combinations of items with high uniformity for each digit, it is possible to further reduce the probability of the procedure to arrive at the correct answer. The above concept is shown in Figure 1 in the form of a flowchart.

The contents of the present invention will be explained below based on the block diagram of FIG. Terminal users 1 to U are telephones 11 and 12.
... is connected to the information guidance center 100 via 1U.

When the terminal user 1 calls the center using the telephone number, the telephone 11 is connected to the center via the exchange 2 and becomes ready to send and receive information. When the line is connected, a voice response device 200 consisting of a control device 3 and a voice file 4 to 1 This is the 00 bank balance inquiry center. A command to enter your bankbook number is output to telephone 11. After hearing this, the user utters the bankbook number, for example, Rl2345J (ichi, niichi, san●yon◆go). This audio is A
/D converted and buffer memory 5 as digital information.
is memorized. This content is first (number) voice recognition device 6
At that time, a numeric string 7 representing the recognition result and a left degree sequence 8 indicating the certainty of the recognition are output and stored in buffer memories 9 and 10. At this time, the output of 7 is not limited to just one for each digit, but all those whose left degree is greater than a preset threshold value are output together with their left degree values. The results of this preliminary recognition are sequentially read out for each digit, starting from the one with the highest frequency, and are stored in the buffer 11 in sequence.

Correspondingly, input information roughly corresponding to that part is led to the feature extraction device 13 via the switch 12, and the extracted feature parameters are temporarily stored in the buffer 14.
A confirmation check is performed on the assumption that this feature is the number of the corresponding primary recognition result, and the result is output to the buffer 15. If the result is negative, a check is performed as the next left degree number. Confirmation is 0.
If the result is K, the recognition result is stored in the buffer 11 to buffer 16. After this reconfirmation process is completed for all digits, the number string is limited to a relatively small number.

If there are few recognition candidates, it is possible to limit the recognition to a small number of numeric strings without having to reconfirm all digits. (At this time, if there is no confirmed digit with sufficient accuracy, it will be rejected at this stage and a voice response will output 0, please enter the number again.)
For this small number of candidate number strings, the voice number of the individual having the hypothetical number is read out from the voice number file 17 through the control circuit 18, and the actual input voice is input through the switch 12. If the result is equal to or higher than a preset degree of similarity, it is output as a correct answer from the threshold circuit 20 and sent to the voice response control section 3.

As a result, a voice message is sent to the user of the terminal to confirm that your bankbook number is 12345.

The terminal user listens to this and answers yes if it is correct, and this response is recognized by the (number) recognition device 6 and proceeds to the next process, for example, confirmation of the speaker. In this explanation, it is assumed that the assumption numbers (candidate numbers) are compared in illuminance in order from the one with the largest overall left degree, so if there is one whose degree of similarity is higher than the threshold, the comparison is stopped there and it is assumed that it is accurate.

However, it is also possible to calculate the similarity for all candidates and select the one that gives the maximum value as accurate. Note that we will not specifically explain number sound recognition itself, feature extraction itself, or similarity calculation itself, but
Many of them are already known, and any of them may be used. For example, one way to recognize digit sounds is to use an autocorrelation coefficient sequence as a feature sequence and match it with a standard pattern using a DP (dynamic programming) method, or use a partial autocorrelation coefficient sequence as a feature sequence to identify each digit sound. There is a method to recognize it as a standard pattern of numeric sounds that gives the maximum left degree from the left degree matrix.

Further, as a feature extraction method for comparison, a spectrum or a Vormand pattern based on spectrum analysis of a speech waveform can be considered. The inner product (normal correlation coefficient) between the input and the standard is taken as the degree of similarity. As explained above, according to the present invention, speech recognition under difficult real situations is narrowed down to a few candidates with high certainty through standard processing under general conditions. By performing a two-step process of performing limited comparison and matching recognition under special conditions assumed by hypothetical results, the accuracy of recognition can be increased and the entire process can be made faster (effectively) in terms of time. This can be said to have a great effect on the practical implementation of various information guidance services.

[Brief explanation of the drawing]

FIG. 1 is a flow chart for explaining the concept of the present invention, and FIG. 2 is a block diagram of an embodiment of the present invention.

Claims (1)

[Claims] 1. Speech recognition means that selects recognition candidates with a predetermined likelihood or higher for each speech pattern by sequentially comparing a plurality of input speech patterns with a standard pattern, and recognition for each predetermined speech pattern from the speech recognition means. means for outputting candidates in descending order of likelihood; feature extraction means for extracting features of the plurality of speech patterns; and output of the feature extraction means corresponding to the speech patterns to determine whether the recognition candidates of the speech patterns are correct or not. and a means for confirming based on the predetermined pattern, the speech recognition device recognizing the plurality of input speech patterns by limiting the recognition candidates with respect to the predetermined pattern.