JPS5849996A - Average phonemic pattern preparation system - 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] For each phoneme section corresponding to the phoneme of the phoneme pattern series,
By creating an average phoneme vowel turn sequence by averaging the types and lengths of phonemes, we can identify local features common to multiple phoneme vowel turn sequences (elimination, insertion, and substitution of consonants).
This makes it possible to average the overall length while preserving vowel substitutions (including articulatory combinations), vowel substitutions (including articulatory combinations), length, etc., and is useful for creating word phonological dictionaries for speech recognition devices, etc. By applying this method, the aim is to automatically create a word phonological dictionary that absorbs phonological fluctuations specific to words, and to improve the word recognition rate.

An example of creating a word phoneme dictionary for a speech recognition device based on conventional phoneme recognition will be described using FIGS. 1 and 2. FIG.

Figure 1 (,) shows a word phonology dictionary based on simple phonological notation (romanization) that does not take into account differences in the length of word phonemes, and Figure 1 (b) shows consonants and vowels (silence). ))
This shows a word phoneme dictionary based on phoneme notation (romanization) that takes into account the difference in phoneme length of 1:3 and 12 silences. However, the sequence of phonetic sounds in actual words is
As shown in Figure 2(b), the length of the phoneme varies from word to word.
), the voiced consonant (B) at the beginning of the word is dropped), and insertion and substitution (in the second (8) (,), the voiced consonant in the middle of the word (RI is replaced with a voiceless consonant (i); in Figure 2 (b), the The middle semi-vowel (G) is replaced with a voiced consonant (G)) and vowel replacement (in Figure 2 (b), the vowel (0) is replaced with a vowel (,)). When performing turn matching with a word dictionary based on phonological notation such as Matching, etc.) resulted in a decrease in the matching score.This resulted in word recognition errors, causing a decrease in the recognition rate of speech recognition devices based on phoneme recognition.

Based on the above, the present invention creates a word phoneme dictionary that expresses the differences in phoneme length and phoneme deformation specific to doodles by automatically equalizing them from the phoneme turns of word utterances by multiple speakers, and This improves the recognition rate of speech recognition devices and the like based on recognition.

An example of the overall configuration of the present invention is shown in FIG. 3, and will be described below.

In FIG. 3, 1 is a phoneme recognition unit, and phoneme recognition unit 3
, a phoneme sequence merging section 4, and a phoneme standard and turn storage section 5. 2 consists of a phoneme averaging section, a phoneme sequence averaging section 6, and a word phoneme dictionary 7.

In the phoneme recognition unit 1, it is assumed that the input speech is expressed as a series of feature vectors as shown in the following equation (1).

X1X2...XN...(1) Each Xi, i=1. ..., N are m-dimensional vectors, and are expressed as Xi=(Xi1,..., xim). Here, for example, the feature vector is m
Channel bandpass filter output x1(t),・
..., xj(t), ..., xm(t) can be considered to be time-sampled. Furthermore, a section of audio represented by a feature vector is sometimes referred to as a frame. (1
) subscripts 1, 2. ..., N is a parameter representing time.

In equation (1), each feature vector can be considered to correspond to a specific phoneme of Zuka speech.

For example, when the word /nara/ is input, in the series of equations (1), The following correspondence can be made: ■/r/ Xi3+1,...,XN■/a/. Here, 1<i1<i2<i
3<N.

In the following, associating feature vectors with phonemes is referred to as phoneme recognition in a narrow sense. It can simply be called phoneme recognition as long as it does not cause confusion. In addition, the process of determining phoneme intervals from sounds and vowel turns is called segmentation. In the above example, 1,...,i1|i1+1,...,i2|i2+1
,...,i3|i3+1,...,N. Here, 1 is a symbol representing a boundary between phoneme sections. Furthermore, in addition to phoneme recognition and segmentation in a narrow sense, phoneme recognition in a broad sense refers to converting an input speech line into a sequence of phonemes. In other words, the speech pattern expressed by equation (1) is determined by the phoneme recognition unit 3.
The degree of similarity with all the phoneme standards and mother turns imported into the phoneme standard and mother turn storage unit 5 is calculated, and the phoneme with the highest similarity is set as the recognition result of its feature vector, and the necessary If so, calculate the second candidate, third candidate, etc. When the phoneme sequence merging unit 4 receives the output from the phoneme recognition unit 3, if the same phoneme continues, it performs processing such as combining these into one phoneme and outputs a phoneme sequence.

Now, for convenience of explanation, the processing of the phoneme recognition unit 1 will be described in detail. The pattern stored in the phoneme standard pattern storage unit 5 is
,,y,...,yM. It is assumed that each pattern is an m-dimensional vector and is expressed as Yj=(yj1,...,yjm) j=1,...,M.

Now, if the speech pattern shown by equation (1) is generated by an m-channel bandpass filter, and the standard pattern is also created based on the output of the bandpass filter, then the i-th feature vector Xi and the phoneme standard? The degree of similarity with turn Yj can be expressed by the Euclidean distance defined, for example. Assume that the speech pattern and the phoneme standard pattern are the output of a bandpass filter, and let us evaluate the degree of similarity using Euclidean distance.

MinS(i,j) 1≦j≦M Now, assume that the recognized phoneme of the i-th feature vector is the minimum phoneme J1i that satisfies , and if necessary, set the second candidate phoneme to the phoneme j2 with the next smallest similarity. , the third and subsequent candidates are defined in the same way.

This process is applied to all feature vectors X1.

..., XN, and the result is output from the phoneme recognition unit 1.

The phonemes are vowels (/a/, /i/, ..., /o/), long vowels (/ai/, /iu/, ...), and semi-vowels (/y/).

/w/) voiced consonants (/m/, /n/, /g/.

/z/,...), voiceless consonants, plosives, consonants, mujins, etc. Among these phonemes, vowels have a relatively long utterance interval and are stable as they are not affected by the phonemes before and after them, whereas consonants are the vowels before and after them (in Japanese, except for the beginning of a word, consonants are similar to vowels). This is an extremely unstable turn that is affected by (which appears in a surrounded form).

Next, the phoneme average part will be explained.

In the phoneme averaging section 2, the phoneme sequence averaging section 6 calculates the average of a plurality of input phonemes and turn sequences corresponding to one word.

The word phoneme dictionary 7 stores the average phoneme/cuttern sequence calculated by the phoneme sequence averaging unit 6 for each word.

The function of the phoneme sequence averaging section 6 will be explained below.

As shown in FIG. 4, the phoneme sequence averaging unit 6 includes a written phoneme storage unit 8. Input phoneme sequence phoneme segment separation unit 9. Input phoneme sequence storage unit 10. Phonological interval average part 11. It consists of an average phoneme sequence storage section 12.

The written phoneme pattern storage unit 8 stores, for each word, a word written phoneme-turn sequence for which an average phoneme/mother turn sequence is to be determined. For example, as shown in Figure 1 (,), if the word is "ranch" or "1 sea bathing" using the Roman alphabet, /BO・KuZiYO/ /KAiSuiYO・Ku/ is stored. (/ is a word break mark) The input phoneme series phoneme segment separation unit 9 first divides the written phoneme turn storage unit 8 and the input phoneme sequence storage unit 10 into certain characteristic phoneme intervals (for example, CVC interval or VCV interval: V vowel interval, C consonant). After dividing the phoneme into two sections (intervals) and associating the characteristic phoneme intervals of both, the associated characteristic phoneme intervals are associated with phoneme intervals (C interval, ■ interval, etc.), and finally the correspondence is established. The input phoneme pattern sequence is separated into phoneme sections corresponding to one phoneme of the written phoneme pattern sequence by associating one phoneme of the turn with the written phoneme and a plurality of phonemes of the input phoneme/turn for the phoneme interval.

The phoneme segment averaging unit 11 calculates the type of phoneme (the phoneme identified by the phoneme recognition unit) for each phoneme segment for the phoneme vowel turn series in which the phoneme segments in the input phoneme sequence storage unit 10 and the average phoneme sequence storage unit 12 are separated. After calculating the average of the phoneme duration (type of phoneme) and phoneme duration (number of phoneme frames), the average phoneme sequence storage unit 12
Output to.

Regarding the -th input phoneme pattern sequence, immediately after the phoneme sections are separated without performing average calculation, it is transferred to the average phoneme sequence storage section 12.

Here, a detailed explanation of the function of the input phoneme sequence phoneme segment separation section 9 will be given below.

The input phoneme sequence phoneme segment separation unit 9 converts the characteristic phoneme intervals into CV
In the case of creating a C interval or a VCV interval, first, as shown in Table 1, the written phoneme/guttern series and the input phoneme pattern series are separated into consonant intervals and vowel intervals.

Regarding the written phonological pattern series, due to the nature of Japanese syllables, regarding consonant intervals, if the beginning of a word starts with a vowel and silence (・) is considered a consonant, Table 1 (
, ), Ci=CRiCRi+1 1≦i≦−NCRj(−
(CU0U.) It can be expressed as 1≦j≦N+1. In addition, considering diphthongs, the vowel interval is generally as shown in Table 1 (b), Vi=VRiVRi+1 1≦i≦NVRi=(-
(VU0) It can be expressed as 1≦i≦N+1. Regarding the input phoneme pattern series, since the phoneme changes and the duration of the phoneme is not constant, the consonant interval is generally as shown in Table 1 (b), CPi=CPRiCPRi+1...CPRn 1≦i
≦MCPRi(-(CU0U・) 1
≦j≦M+n↓ Similarly, the vowel interval is generally expressed as shown in Table 1 (b), VPi=■PRiVPRi+1...VPRm 1≦i
≦MVPRj(-(VU0)
It can be expressed as 1≦j≦M+n.

In the input phoneme pattern series, as explained in the phoneme recognition unit 1, vowels have a relatively long duration and are stable, and consonants have a short duration and are unstable. However, consonants contain a lot of information as points of phonological change.

1.By characteristic phonetic intervals such as BCVC and VCV,
Correlation of phoneme intervals allows accurate mapping of consonant omissions, insertions, and substitutions, as well as vowel substitutions.

Tables 2 and 3 show the correspondence between characteristic phoneme intervals using CVC.
The explanation will be given below using a table.

CVC of written phonological pattern sequence and input phonological turn sequence
Assuming that the terminal consonant intervals to which the intervals are associated are C1 and C1t, respectively, as shown in Table 2 (,), the written phoneme or turn sequence is C1■kCk+.

section (Rk section) or CkvkCk+, vk+, C
k+2 interval (Rk+1 interval) and CPtvPtCPt+1 interval (I, interval) or C of the input phoneme/mother turn sequence
PtvPtCPt+1vPt+, CPt+2 section ('t
+1 interval). If consonant section Ck+1 is dropped, use Rk+11 as the corresponding section and consonant section CP.
When t+1 is inserted, correspondence between Rk and ■t+4 sections occurs. Table 2 (
As shown in b), since consonants have a lot of information as phonological change points, (Ck+4.CPt+,),
(Ck+1°CPt+2), (Ck+□#CPt+1)
The similarity of the consonant intervals of is calculated, and the maximum value is (Ck
+1. C1t+1), immediate response in Rk, It section,
In case of (Ck+,・CPt+2), correspondence of R,k”t+1 interval, in case of (Ck+2.CPt+1), Rk+
11' is used to associate sections. To make these correspondences directly, as shown in Table 3, (Rk#
The similarity of Iz)#(Rk'eJt+1)#(Rk+1#'t) is calculated, and the combination that takes the maximum value is defined as the corresponding interval.

When the association of CVC sections is completed, in the case of Rk and sIz sections, Ckvk and CPtvPt are removed from the chain, and R
k.

10kvk and ptvpt for t+1 section
+1vpt+1 is removed from the chain and Rk+1. In the case of the It section, CkvkCk+, vk+, and CPtvP't are removed from the chain, and the same correspondence is made again for subsequent sections.

Tables 4 and 5 show the correspondence between characteristic phoneme intervals using VCV.
The explanation will be given below using a table.

VC of written phoneme/mother turn sequence and input phoneme pattern sequence
The terminal vowel interval to which the V interval is associated is shown in Table 4 (,)
As shown in , if V, v, and t are respectively VkCk+1vk+1 section (R2' section) or ""k+1"k+1 section ("
k+1 interval) and the vpt interval of the input phoneme/mother turn sequence (
Consonant interval Ck+
, if R' is dropped, +1"p' interval correspondence to R',
If consonant interval CPt+1 is inserted, Rk/”P
+1 interval correspondence occurs. To easily make these correspondences, as shown in Table 4 (b), since the vowels are stable, (Vk, V, t).

(vk, vplvpl+1), (vkvk+1, vpl
) The similarity of the vowel interval is calculated, and its maximum value is (vk
, vPtVpl+1), then the correspondence between RIk and the interval is (V,).

In the case of VPtvPt+, ), there is a correspondence of +"j+1 interval to R', and in the case of (vkvk+1.vPt), there is a correspondence of 2R(+1.
(En interval) Make a correspondence. To make these correspondences directly, as shown in Table 5, (RIk, I
'l), (RIk, ■It+,), (+1.equal to R')
The similarity is calculated, and the combination that takes the maximum value is defined as the corresponding interval.

When the association of the VCV sections is completed, in the case of the R', Cl, section, vkCk+1 and vPtCPt+1 are removed from the chain, and in the case of the Il+1 section, ■kCk+, and V
Remove PtCPt+1vPt+1CPt+2 from the chain and R
”For k+1l1 section, %vkCk+1vk+ICk
+2+vptCpt+1? Remove it from the chain and perform the same mapping again for subsequent sections.

To easily and accurately match characteristic phoneme intervals,
It is possible to consider a combination of correspondence between CvC intervals and correspondence between VCV intervals. Assume that the correspondence between the CVC sections described above is determined based on the similarity of only the consonant sections, and the correspondence between the VCV sections is determined based on the similarity of only the vowel sections. Initially, the mapping of CVC sections is applied, and when significant consonant sections are dropped or inserted, and when it becomes impossible to map using the above method, the mapping of vC ■ sections is applied, and the above method is applied. When the matching by method is completed, C
A method of associating VC sections can be considered.

C between the written phonological pattern sequence and the input phonological turn sequence
Once the correspondence between VC or VCV characteristic phoneme intervals is completed, dropouts and insertions of consonant intervals within the characteristic phoneme interval can be immediately detected, so that the correspondence between consonant intervals and vowel intervals of both series within the characteristic phoneme interval can be easily detected. I can do it. Based on the definitions of the consonant and vowel intervals of the written phoneme Noreen series and the definition of the input phoneme turn, the similarity between the phonemes shown in Figure 5 is calculated as shown in Table 6 (a) and (b). This is accomplished by associating one phoneme of the written phoneme vowel turn sequence with multiple phonemes of the input phoneme Noreen sequence.

As described above, in the present invention, the characteristic rhyme interval (e
ve interval, VCV interval, etc.), then phonological intervals (C interval, The input phoneme i-turn sequence is divided by matching. Then, the phoneme classification and phoneme duration are averaged for each of the six phoneme sections to determine the average of multiple input phonemes and turn sequences. and creation can be done automatically. Furthermore, by applying the present invention to a word phoneme dictionary in a speech recognition device based on phoneme recognition, it is possible to improve the recognition rate.

[Brief explanation of the drawing]

Figure 1 shows a word phonological dictionary based on the conventional Roman alphabet phonetic notation. Figure 1 (a) shows the case in which the order of one sound is represented by one character, and Figure 1 (b) shows the case where the order of one sound is represented by one character. Figure 2 is a diagram showing the difference in the phoneme length of vowels expressed by the number of characters. Figure 2 is a diagram showing the phoneme sequence of the phoneme recognition results for words. FIG.
4 is a diagram showing a configuration example of the average phoneme pattern sequence creation method according to the present invention, FIG. 4 is a diagram showing a detailed configuration example of the phoneme sequence averaging section of the configuration example according to the present invention shown in FIG. 3,
FIG. 5 is a diagram showing the degree of similarity between phonemes. 1... Phoneme recognition section, 2... Phoneme averaging section, 3... Speech rate recognition section, 4... Merging section, 5... Phoneme standard pattern storage section, 6... Phoneme sequence averaging section , 7... Word phoneme dictionary, 8... Written phoneme/mother turn storage section, 9... Input sound ml series phoneme section separation section, 10... Input phoneme sequence storage section, 11... Phoneme section Average part, 12... Average phoneme sequence storage part. Patent applicant Matsushita Electric Industrial Co., Ltd.

Claims (1)

[Claims] A phoneme recognition means for recognizing phonemes (vowels, voiced consonants, voiceless consonants, silence, etc.) from an input speech turn sequence, a phoneme sequence averaging means for determining the average or turn of a plurality of phoneme/mother turn sequences, The phoneme series averaging means calculates the written phoneme (romanization, etc.) for each phoneme/lean series.
The phoneme interval corresponding to the phoneme is divided into the phoneme Noreen series and the written phoneme/ya turn series as a constant characteristic phoneme interval (CVC).
After dividing into sections (c: consonant, v: vowel), VCV section, etc.) and making a correspondence between the characteristic phoneme sections, the characteristic phoneme section is detected by making a correspondence between the phonemes. An average phoneme method that is characterized by creating an average phoneme-turn sequence of a plurality of phoneme/mother-turn sequences by calculating the average of the phoneme type and phoneme length for each phoneme interval corresponding to the phoneme of the written phoneme. Turn creation method.