JPH0464076B2 - - Google Patents

Description

[Detailed description of the invention]

[Technical Field of the Invention] The present invention relates to a highly practical speech recognition device that can accurately recognize continuously uttered input speech. [Technical background of the invention and its problems] Automatic speech recognition is extremely important as an interface technology that enables direct information input from humans to machines. Therefore, this automatic recognition of speech converts a speech pattern made by continuously uttering a series of linguistic symbols such as phonemes, syllables, and words into discrete linguistic symbols, and then converts each of these linguistic symbols into discrete linguistic symbols. Each can be thought of as a process of recognition. By the way, the human vocal organ is a physical thing with a certain mass, and therefore the voice produced does not change discretely. As a result, it is undeniable that the speech pattern corresponding to the linguistic symbol of continuously uttered speech is affected by the surrounding environment.
For example, if we consider phonemes as linguistic symbols, it is possible to cut out the sound pattern (speech wave) of the "O" part of the word "aoi" and recognize it as a clear "O" even if we listen to it. is difficult. Also, if we consider words as linguistic symbols, the number string “83
While the ``chi'' part in ``(Hachi-san)'' is often devoiced, it is almost never devoiced in the case of ``81 (Hachi-ichi)''. In other words, there is a phenomenon in which the same number "8" has slightly different voice patterns. In this way, the linguistic symbols of the sounds that humans continuously utter often change due to the strong influence of the surroundings, and for this reason, one linguistic symbol is regarded as one speech pattern, and the surrounding environment It was impossible to recognize and process them independently. Conventionally, it has been considered to categorize various transformations and prepare a plurality of sound patterns as standard patterns for one linguistic symbol. However, when trying to recognize all sounds and syllables with many types of surrounding environments, or continuous utterances of words with a large vocabulary, it would be impractical to prepare standard patterns such as the one described above because the number would be enormous. There was a problem with it missing. On the other hand, as mentioned above, since the speech patterns for linguistic symbols are deformed due to the influence of the surrounding environment, it has been considered to perform recognition processing by removing this influence. However, in order to eliminate the influence of the environment around a certain symbol in this way, the symbols before and after it must already be determined. Moreover, when determining the symbols before and after the symbol, it is necessary to determine the symbols before and after the symbol, and there is a contradiction in proceeding with this process.
In other words, because each linguistic symbol influences each other, it is not possible to easily eliminate the influence of the context. [Object of the Invention] The present invention has been made in consideration of the above circumstances, and its purpose is to eliminate the influence of the context between consecutive linguistic symbols from continuously uttered input speech. It is an object of the present invention to provide a highly practical speech recognition device that can recognize with high accuracy. [Summary of the Invention] The present invention analyzes input speech, separates its feature parameter series into units of discrete linguistic symbols, and divides each unit into one or more linguistic symbol candidates and each of these symbol candidates. The likelihoods for the above units are calculated, each of these likelihoods is updated as a function of the matching coefficient between each symbol candidate between temporally adjacent units, and these updated likelihoods exceed a predetermined threshold. When this happens, the symbol candidate that obtains that likelihood is output as the recognition result for that unit. [Effects of the Invention] Thus, according to the present invention, the likelihood of a linguistic symbol candidate obtained for each unit of input speech separated into discrete units of linguistic symbols is calculated based on the likelihood of a linguistic symbol candidate obtained for each unit of temporally adjacent units. Since the compatibility coefficient between symbol candidates in between is updated as a function, even if the previous and subsequent symbol candidates are not determined, it is possible to determine the most likely symbol candidate while sequentially removing their influence. becomes. In other words, for multiple symbol candidates, the likelihood is sequentially updated in the form of a compatibility coefficient between units based on their context, so the influence between units reflected in the compatibility coefficient is gradually removed, and here continuous utterances are It becomes possible to recognize each linguistic symbol of the input speech with high accuracy. Therefore, it is possible to recognize continuously uttered speech composed of linguistic symbol strings that influence each other throughout the whole, very effectively and with the above influences removed, and it is possible for humans to recognize It has great effects as an interface technology for inputting information to machines. [Embodiment of the Invention] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. Although phonemes will be explained here as examples of linguistic symbols, it is of course possible to treat syllables, words, etc. as linguistic symbols. FIG. 1 is a schematic configuration diagram of an embodiment device. Speech that is continuously uttered and input is led to the acoustic analysis section 1, where it is acoustically analyzed at predetermined analysis time intervals and converted into characteristic parameters. The acoustic analysis section 1 is constituted by a filter bank consisting of a plurality of band pass filters, which is known as a method of spectrum analysis, for example. The characteristic parameter time series of the input speech that has been acoustically analyzed and converted is input to the segmentation/symbolization processing unit 2, where it is segmented and segmented into phonemes, which are units of discrete linguistic symbols, and each unit is It is symbolized as . This segmentation/symbolization processing unit 2 detects phoneme boundaries from characteristic parameters such as voice power and spectral changes, performs segmentation/segmentation using phonemes as units, and then uses a statistical discriminant function for each unit. One or more phoneme candidates, that is, linguistic symbol candidates, are obtained for each unit by performing a matching process with a standard pattern. At this time, the likelihood expressed by distance, similarity, etc. is also calculated for each phoneme candidate. In other words, this likelihood can be said to indicate the likelihood of the phoneme candidate when compared with the feature parameters of the target unit. FIG. 2 shows the relationship between the phoneme candidates obtained in this way and their likelihoods in relation to the voice power of the input voice "Onsay". In other words, an example is shown in which the phoneme candidates for each unit obtained by dividing and symbolizing the word "oNsei" are shown in alpha alphabets, and the likelihood is shown as a perfect score of (100). As is clear from this example, the series of phoneme candidates with high likelihood found for each unit is "omsai", and the feature parameters of each unit change due to the influence of the previous and subsequent units. It can be seen that Therefore, in conventional recognition processing methods using only such encoding results, errors occur due to mutual influence between units. Thus, information on one or more phoneme candidates and their likelihoods for each unit, as shown in FIG. . The information stored in the storage section 3 is evaluated by the symbol evaluation section 5 under the control of the control section 4 . Next, this evaluation process will be explained. As described above, each separated unit of phoneme is influenced by the phonemes before and after it. Therefore, in order to determine whether or not a phoneme candidate found for a certain unit of phoneme is correct, the likelihood alone is not sufficient. Therefore, the likelihood of the phoneme candidate α found for the i-th unit phoneme, that is, the correct probability, is defined as P _i (α), and for the units (i-1) and (i+1) before and after it, Let the obtained phoneme candidates be β and γ, respectively. In addition, the mutual influence of phoneme candidates β and α, or α and γ is calculated using a compatibility coefficient r(β,
α), r(α, γ). Note that this compatibility coefficient r (β, α) takes a large value for the correct phoneme sequence (β, α), and for the phoneme sequence (β, α) with other phonemes δ.
For δ) and (δ, α), a scale that gives a small value is used. For example, the probability that a phoneme string (β, α) is correctly recognized may be determined as the matching coefficient r(β, α). This compatibility coefficient r(β, α) is the probability that a phoneme string (β, α) will be correctly encoded, and the probability that another phoneme string (β, δ) will be erroneously encoded as a phoneme string (β, α). This method takes advantage of the property that the probability is usually larger than the probability of occurrence, and can be easily determined statistically from, for example, the processing results of the division/symbolization processing section 2, such as the recognition rate or the frequency of appearance of characters. The following table shows only typical adaptation coefficients obtained in this way.

[Table] Therefore, the symbol evaluation unit 5 uses such a compatibility coefficient to update the likelihood P _i (α) of the phoneme candidate α according to the relationship with the phoneme candidates β and γ before and after it. It turns out. Now, if the likelihood when update processing is performed k times (k=1, 2...) is P ^k _i (α), then
The updating process is performed by performing the calculation shown in the following equation. S ^k _i (α)=P ^k-1 _i (α) ×{1+ 〓 〓P ^k _i-1 (β)r(β, α)+ 〓 〓P ^k-1 _i+1 (γ)r(α , γ)} ...(1) P ^k _i (α)=S ^k _i (α)/ 〓 〓S ^k _i (δ) ...(2) However, P ^k _p (β)=P ^k _I+1 (γ)=0 That is, in the above equation (1), processing is performed using the above-mentioned compatibility coefficient as a function in order to remove the mutual influence between word candidates in adjacent units. After that, S ^k _i (α) obtained through this process is normalized using the above equation (2) so that the sum of the probabilities in that unit becomes 1, and this is updated. It is required as a new likelihood. Thereafter, this process is repeatedly executed k times over the entire section (all units) of the input speech, and the likelihood of each phoneme candidate is updated. Note that k in the second term in parentheses in the above equation (1) is for speeding up the convergence of the likelihood in the iterative update process, and may be k-1. Furthermore, depending on how the adaptation coefficient r is set, the above convergence may not be guaranteed. Therefore, taking this into consideration, the upper limit of the value of k may be determined in advance. Also k
It is sufficient to determine the initial value of the likelihood at =1 as the ratio of the similarity values obtained by the above-described division/symbolization processing section 2. FIG. 3 is a control flow showing the above-mentioned likelihood updating process, and the symbol evaluation section 5 operates according to this control flow. Then, each time a new likelihood for a phoneme candidate is determined, the likelihood for each phoneme candidate stored in the storage unit 3 is updated. FIG. 4 shows a transition process when the information shown in FIG. 2 is updated using the adaptation coefficients shown in the table mentioned above. As is clear from the updating process shown in FIG. 4, the likelihood of each phoneme candidate changes due to the updating process, and the ranking of the candidates changes. This means that the influence between units is gradually removed, and in this example (k
= 5), the correct phoneme symbol (oNsei) is obtained, and the candidate rankings are not changed thereafter. In other words, it can be said that it has converged. Therefore, when the phoneme candidates for each classification unit converge, or when the likelihood of the first-order phoneme candidate exceeds a predetermined threshold, if that phoneme candidate is output as a processing result, a correct symbol string will be created. A recognition result will be obtained. In this way, according to the present device, mutual influence between linguistic symbols of continuously uttered speech can be removed, and the symbol strings can be easily and correctly recognized. Therefore, its practical advantages are extremely high, and great effects can be achieved. Note that the present invention is not limited to the above embodiments. For example, if the unit phoneme of interest is influenced by multiple phonemes before and after it, weighting is performed according to the distance between the phonemes, and the result is expressed as shown in equations (1) and (2) above. All you have to do is process it. That is, the weight ω _ij is set to ω _ij =1/|i−j| (i≠j), and S ^k _i (α)=P ^k−1 _i (α)×{1+ 〓 ^j < ⁱ ω _ij 〓 〓P ^k _j (β)r(β,α)+ 〓 ^j > ⁱ ω _ij 〓 〓P ^k-1 _j (γ)r(α, γ)} (3) The calculation may be performed as follows. At this time, it is preferable to give weights such as ω _ij =1 to adjacent phonemes and weights such as ω _ij =0.5 to phonemes outside of the adjacent phonemes. Further, although the above-described updating process is linear, a non-linear relational expression expressed as a product may be introduced. Furthermore, it goes without saying that the adaptation coefficient may also be set using the values of other feature parameters. Also, as mentioned earlier, even though the word strings "83 (Hachisan)" and "81 (Hachiichi)" have the same word "8",
The sound pattern differs due to the influence of the preceding and following words (in this case, the subsequent word). As a result, the magnitude of the similarity value for "8" is also different, and there is a large risk of misrecognition. It is clear that even in such a case, if the above embodiment is applied as is, the influence of the preceding and succeeding words can be removed. Furthermore, the linguistic symbol units may be syllables, words, etc., and in short, the present invention can be implemented with various modifications without departing from the gist thereof.

[Brief explanation of the drawing]

FIG. 1 is a schematic diagram of an apparatus according to an embodiment of the present invention;
FIG. 2 is a diagram showing input speech and its phoneme candidates, FIG. 3 is a diagram showing the flow of likelihood updating processing, and FIG. 4 is a diagram showing an example of the updating processing. 1...Acoustic analysis section, 2...Division/symbolization processing section, 3
...Storage section, 4...Control section, 5...Symbol evaluation section.

Claims (1)

[Claims] 1 Analyze the input speech to separate the feature parameter series of the input speech into discrete linguistic symbol units, and for each of these separated units, one or more linguistic symbol candidates and these linguistic symbols means for determining the likelihood of a symbol candidate for that unit; and means for updating each of the likelihoods of these linguistic symbol candidates as a function of a compatibility coefficient between each linguistic symbol candidate between temporally adjacent units; Speech recognition characterized by comprising means for outputting a linguistic symbol candidate as a recognition result for that unit when the updated likelihood exceeds a predetermined threshold or when the update processing has been performed a predetermined number of times. Device.