KR100350003B1 - A system for determining a word from a speech signal - Google Patents

Abstract

Speech recognition generates a test signal from a speech signal that is compared to a predetermined reference signal to form a score. Each subsequent test signal is compared with a reference value located within a predetermined proximity of the reference value that is optimally determined for the preceding test signal. Depending on this proximity, a transition value that matches the transition probability is added to the score. It has been proposed to adapt the transition value according to the speed of the speaker in order to remarkably improve the result when the instantaneous speaking speed of the speaker is different. Another enhancement can be obtained by applying the reference value to the pronunciation of the associated speaker. This adaptation can also be repeatedly performed in multiple steps.

Description

A system for determining a word from a speech signal

The present invention relates to a system for determining a word of a given vocabulary from a speech signal,

First means for taking a voice signal and supplying a sequence of digital test signals,

Second means for storing sequences of reference signals corresponding to words of the lexicon;

A first and second means for comparing the test signals with the first reference signals to form a score for each first reference signal depending on the difference between the test signal and the first reference signal The third reference signal is a signal neighboring or the same as the second reference signal in the related sequence in which the comparison is successfully performed with respect to the preceding test signal according to a predetermined method, Said third means being arranged to increase said score by a transition value according to a transition probability according to a distance from said first means,

Summing the incremented scores for each sequence of reference signals compared to successive test signals, determining an optimal sequence with a minimum sum of the incremented scores, and outputting a word or words associated with the optimal sequence 4 means.

This type of system is known from DE 32 15 868 C2. The known system acts to determine a word sequence in which the corresponding sequence of reference signals is stored for an individual word and a specific step is carried out and taken to determine the word transition. The comparison or the result of the successive test signal and the reference signal can be represented by a two-dimensional grit, and for each sequence of the reference signal originating from the starting point that has been compared in successive test signals from the starting point, It is determined whether the reference signal generates a minimum score sum in relation to the next test signal to find the path in the grit from the given starting point in the word to the end of the word. Thus, within the word, the next test signal is compared to a reference value wave located at a given neighbor around the end of the path just reached. In this way, a non-linear temporal adaptation between the actually pronounced word and the reference value sequence of this word . Within the word, the proximity of a plurality of transitions, i. E., A reference value determined to be optimal for a test signal and an optimum reference value for that preceding test signal, is handled in the same manner.

DE 37 10 507 A1 discloses a similar system for recognizing pronounced words in which another proximity of the optimal reference signal for a continuous test signal is taken into account. Thus, the transition probability is clearly modeled. In particular, a fixed transition value is added to the score depending on the proximity. It is assumed that in this position the score is formed by the negative logarithm of the probability that the actually pronounced word corresponds to the relevant reference signal.

Because the velocity of the pronounced word corresponds to the reference signal sequence, the diagonal path of the path may be preferred by the selection of the appropriate transition value, since the diagonal path is the most appropriate. Therefore, even in the case of different pronunciation methods, words can be recognized with different scores. The pronunciation rate is modeled by the choice of the transition value.

A reference value is determined based on a test sentence to be pronounced prior to actual use of the system. When the system is for a given user, these test statements are recorded for that user only. Thus, the pronunciation speed is modeled simultaneously. However, when the system is for a plurality of users or is completely independent of the speaker as an ideal case, the reference value may be derived from test sentences pronounced by a number of different speakers. The average value is determined for the reference value as well as the transition value, and the average value is assumed to be the same for all positions in all words. However, this means that when the speaker speaks very fast through the word the optimal path of the reference signal is tilted more than the diagonal, the overall transition value forms a less attractive score, and the pronounced word is misidentified as a similarly sounding reference word Because it is easy, it decreases recognition, reliability.

It is an object of the present invention to provide a highly reliable system to be used by various speakers.

This object is achieved by a method for changing the transition values to new transition values for subsequent comparisons, depending on the difference between the length of the speech signal from which the test signals compared to the optimal sequence of reference signals are derived and the optimal sequence length of the reference signals 5 means are provided.

Thus, by adaptation of the transition value, the reference signal sequence in the system according to the invention is adapted to the rate at which the instantaneous user speaks. As soon as a word is recognized and known, the transition value can be adapted to make subsequent words more recognizable.

The possibility of adaptation of the transition value according to the invention is due to the fifth means being arranged to change the transition value a to a new transition value a 'according to n = T / N as follows:

The index i, i means that the first reference signal and the second reference signal are the same, and the index i, i + 1 is the length of the reference signal, The indexes i and i + 2 mean that the first reference signal and the second reference signal are separated by another reference signal, and the indexes i and i + 2 indicate that the first reference signal and the second reference signal are directly adjacent to each other, Is a predetermined proportional factor.

The ratio of the length of the actual spoken word to the length of the reference signal sequence is used to change the transition value so that the overall transition probability is advantageously used as the degree to which another deviation is suppressed .

This method improves recognition reliability by considering the actual pronunciation speed of the actual user of the system.

A further improvement is achieved in an embodiment of the present invention in which a sixth means is provided for changing the reference value r _i to a new reference value r _i 'as follows.

That is, y _t , is the test signal compared with the reference value r _i at the optimum value of the reference value, and c is a predetermined value. Thus, not only is the speed of speech taken into account, but also the intonation, that is, the vocal tract of the speaker, is also taken into account. Although the theory of adapting the reference values to the instantaneous speaker is known, it is not related to the adaptation to the pronunciation rate.

The adaptation to the pronunciation rate and the instantaneous speaker's pronunciation must be performed in a deep cycle since it must not be adapted to accidental extreme values in the word (s) pronounced by the instantaneous speaker in a particular way, This is because his pronunciation method can be changed later. The adaptation degree can be realized by the proportional factor b for the pronunciation speed and by the predetermined value c for changing the reference value,

The size of both arguments should not be too large. However, in order to make an appropriate adaptation to the instantaneous pronunciation method of the speaker, in another embodiment of the present invention, at least the change of the reference value in the same voice signal is performed several times. A gradual adaptation of the instantaneous speaker's pronunciation will be achieved after several recognized words.

Embodiments of the present invention will be described below in detail with reference to the drawings.

The block diagram of FIG. 1 shows a microphone 2 for converting a voice signal pronounced by a speaker into an electrical signal. This signal is processed and digitized in the block 10, for example, the individual frequency components of the speech signal are determined as a segment. The length of these segments has a constant value between 10 ms and 20 ms. Block 10 outputs a test signal to block 30.

At block 30, the test signal is compared to a reference signal supplied from the memory 20, which is controlled and addressed by the block 30. These reference signals were predetermined by interpretation of test sentences pronounced by some other speaker. This comparison generates an increased score by the transition value stored in the block 30. [ The incremented score at block 40 is summed over several paths through different words. However, such summation may occur concurrently with a comparison for determining the score. At the end of the word or at the end of the speech signal consisting of several words, the optimal overall path is determined at block 40 and the corresponding word sequence is output to block 70. [ This block 70 may be, for example, a display screen, but is preferably formed by a device controlled by voice commands.

The comparison of the reference signal of the other words with the successive test signal and the determination of the optimal sequence will be described later with reference to FIG. 2,

The time axis t represents a sequence of test signals determined from the received speech signal, while the R axis represents a sequence of reference signals for a plurality of words. FIG. 2 shows that the first test signal most corresponds to the time magnitude R 1 of the reference signal associated with the word W 1. A comparison with the starting points of the other sequences R2 and R3 related to the other words also starts every hour. However, it is assumed that the similarity is small so that this series of comparisons ends quickly. Generally, a new comparison with the sequence Rl of the test signal is started again using the subsequent test signal. However, this comparison also ends quickly because later portions of the spoken voice signal deviate too far from the starting point of the sequence Rl of the reference signal.

After the end of the path through the sequence Rl of the reference signal (this path corresponds to the word Wl as mentioned above), the comparison with the starting point of the sequence of reference signals R1 to R3 is continued and, in this example, It is assumed that the path starting from the sequence R3 becomes the optimum path, and the word W3 is subsequently recognized and output. Because multiple words are pronounced, the comparison continues in a similar way when the speech signal lasts for a long time.

The events that occur during the execution of a comparison in one word are shown in detail in Figure 2 for two consecutive test signals only at time t and t + 1 with some reference signals r _i , r _{i + 1} , r _{i +} Will be described in detail with reference to FIG. It is assumed that the optimum path P is terminated at the reference value ri for the t-moment test signal. The next test signal is applied to the reference signal r _i , r _{i + 1} at time t + 1 in accordance with the allowed transition _{i, i} , a _{i, i + 1} , a _i, , r _{i + 2} . In the present embodiment, these transition values correspond to a negative log of transition probability. The comparison of the instantaneous test signal with the reference signal r _i at time t + 1 depends on the difference between the two signals and generates an increased score by the transition value a _{i, i} . In a similar manner, the comparison of the reference signal r _{i + 1} and this test signal results in an increased score by the transition values a _{;, i + 1} . Similarly, the comparison of this test signal with the reference signal r _{i + 2} results in an increased score by the transition value a _{i, i + 2} . For example, the transition values a _{i, i} and a _{i, i + 2} have the same or almost no difference, while the transition values a _{i, i + 1} are substantially smaller. When the t + 1 instantaneous test signal exhibits approximately the same difference compared to all of the three reference signals shown (which is quite possible because the neighboring reference values are often similar), the minimum score increased by the transition value is the reference value r _{i +1,} and the path P ended at the reference value r _i continues in the diagonal direction. Thus, this diagonal direction has the privilege.

In the case of a fast-pronounced speaker, the test signal at the instant t + 1 becomes more similar to the reference value r _{i + 2} , but if the transition values a _i , _{i + 2} are too large for the transition values a _{i, i +} , The diagonal direction may be excessively given. If this occurs repeatedly in the word, i. E., Within the reference signal sequence, ultimately the sum of the less interesting scores is obtained between the test signal sequence and the reference signal sequence, except for excessively fast pronunciation will be. Eventually, this results in poor perceptions. Thus, if it is established that the speaker is pronunciating too quickly or too slowly to a predetermined range (i.e., if it is normalized), the transition value will be changed to give a privilege to the direction deviating somewhat from the diagonal.

This adaptation is performed in block 50 of FIG.

It is known how many test signals are required for this purpose when a short sequence or word of a word is recognized, i. E. When a path leading to the end of the relevant sequence is determined in the reference signals of at least one sequence. Since the number of reference signals in this sequence is given, a ratio n can be calculated.

That is, n = T / N

Where T is the number of test signals for which the word is recognized and N is the number of reference signals in the associated sequence. Using this n, a new transition value a 'is determined from the current transition value a.

Here, the proportional factor b determines the degree to which the transition value is adapted to the speaker's pronunciation. To ensure that this adaptation does not depend too much on accidental maxima in the speaker's pronunciation, the b value should not be too large. When the transition value a is expressed by the negative logarithm of the transition probability as described above, it has been found that a value of b = 180 is a suitable trade-off. Thus, the transition value for the diagonal line is kept constant regardless of the value, because in the case of fast pronunciation, the transition value a _{i, i + 2, which} is more sloppy _, increases the transition value for the less sloped transition a _i, As shown in FIG. As a result, an overall, more inclined transition between the reference signals for successive test signals has the privilege. Again, the same result occurs for slow pronunciations. A new transition value is sent to block 30, where a subsequent comparison .

The reliability of recognition can be further improved by adapting the reference signals to the pronunciation of the speaker. This is realized in block 60 of FIG. 1 as follows.

After the word is recognized in this way, the test signal is again compared with the reference signal sequence for which the best path was found before, and each reference signal r _i is changed to an adapted reference signal r ' _i as follows.

Here, y _t is the test signal compared with the reference signal r _i at time t, and the argument c indicates the range in which the current reference signal is changed. A value of c = 0.13 is particularly effective when the reference signal is adapted to a plurality of steps with successive words.

So far, it has been assumed that the sequence of reference signals represent each word. However, there is also a speech recognition system in which the individual sequences of reference signals represent the same phoneme in a plurality of words. At this time, the word is formed from the recognized phoneme. The method of adapting the transition value and possibly the reference signal is also suitable for use in such a system.

Moreover, contrary to the above embodiment, the method can also be used, although the transition values are not the same for all reference signals and are separated from each other according to the position of the reference signal in the sequence. The adaptation factor b should then be determined in a position dependent manner if necessary.

Figure 1 is a block diagram of a system according to the present invention;

FIG. 2 illustrates path formation through a continuous word; FIG.

FIG. 3 illustrates a transition for two successive test signals; FIG.

DESCRIPTION OF REFERENCE NUMERALS 2: microphone 20: memory

Claims (4)

A system for determining a word of a given vocabulary from a speech signal, comprising: first means for taking the speech signal and supplying a sequence of digital test signals; Second means for storing sequences of reference signals corresponding to words of said lexicon; A first and second means for comparing the test signals with the first reference signals to form a score for each first reference signal depending on the difference between the test signal and the first reference signal Wherein the first reference signal is a signal neighboring or the same as the second reference signal in the related sequence in which the comparison is successfully performed with respect to the preceding test signal according to the predetermined anticorruption, The third means being arranged to increase the score by a transition value according to a transition probability according to a distance from the signal, Summing the incremented scores for each sequence of reference signals compared to successive test signals, determining an optimal sequence with a minimum sum of the incremented scores, and outputting a word or words associated with the optimal sequence 4. A system for determining words of a given vocabulary from a speech signal, Which changes the transition values to new transition values for subsequent comparisons, depending on the difference between the length of the speech signal from which the test signals compared to the optimal sequence of reference signals are derived and the length of the optimal sequence of reference signals, 5 < / RTI > means are provided. The method according to claim 1, Wherein the fifth means sets the transition values a according to u = T / N to new transition values a ' Where T is the length of the sequence of test signals, N is the length of the sequence of reference signals, index i, i means that the first reference signal and the second reference signal are identical, and the index i, i + 1 means that the first reference signal and the second reference signal are directly adjacent to each other, and indices i and i + 2 indicate that the first reference signal and the second reference signal are separated by another reference signal , And b is a predetermined proportional factor. 3. The method according to claim 1 or 2, The reference values r _i are calculated by the new reference values r _i ' Wherein y _t is a test signal compared to a reference value r _i in the optimal sequence of reference values, and c is a predetermined value. The method of claim 3, Wherein at least a change of the reference values in the same voice signal is performed several times.