JPH03228100A - Voice recognition device - Google Patents

Abstract

PURPOSE:To select a candidate word and segment a voice section at the same time by performing word spotting by continuous Maharanobis' DP, word by word, in a 1st stage. CONSTITUTION:A voice input part 100 inputs a voice signal from a microphone and a voice analysis part 101 transfers the input waveform. A word distance calculation part 102 by the continuous Maharanobis' DP matches a time series of feature vectors which are inputted one after another by the continuous Maharanobis' DP with the standard patterns of all words stored in a word standard pattern storage part 103 by the continuous Maharanobis' DP to calculate distances. A candidate word decision part 104 decides candidate words among word standard patterns according to the distances between respective frames and the word standard patterns which are found by the continuous Maharanobis' DP. The parameters of feature vectors in >=1 candidate section are stored in a parameter storage part 105.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention is a speech recognition device, and in particular, a speech recognition device that recognizes speech such as words continuously uttered by any speaker with a high recognition rate. The present invention relates to a speech recognition device.

[Conventional technology]

Although several recognition methods for speaker-independent recognition have been devised, we will discuss a conventional example of a speaker-independent recognition system that is currently the most common and has a configuration relatively similar to the present proposal.

Conventionally, a recognition system aiming at an unspecified large amount of words has a configuration as shown in FIG. The voice input from the voice input unit 1 is processed by the voice analysis unit 2 to obtain characteristic parameters such as a filter bank output including the voice power term, LPC cepstrum, etc. -dimensional compression by L transformation etc.) is also performed.

(Since the analysis is performed frame by frame, the compressed feature parameters are hereinafter referred to as feature vectors.) Next, processing for determining phoneme boundaries from continuous speech is carried out by the phoneme boundary detection unit 3.
This is done by The phoneme identification unit 4 determines phonemes using a statistical method. 5 is a phoneme standard pattern storage unit that stores phoneme standard patterns created from a large number of phoneme samples.

Reference numeral 6 denotes a word identification unit that corrects the output result of the phoneme identification unit 4 using a word dictionary 7 or the correction rule unit 8 from among the output candidate phonemes and outputs the final recognition result. 9 displays the recognition result. This is the recognition result display section.

Usually, the phoneme boundary detection section 3 uses a discriminant function or the like, and the phoneme identification section 4 also performs the same discrimination.

Each of these components generally outputs candidates that satisfy a certain threshold. Multiple candidates are output for each candidate, but Top d like 7.8
Own information is used to narrow down the words to the final word.

[Problem that the invention is trying to solve]

However, since the basic configuration of the conventional recognition apparatus described above is of a bottom-up type, if an error occurs at a certain point in the recognition process, it is likely to have an adverse effect on subsequent processes. (For example, if the phoneme boundary detection unit 3 makes a mistake in identifying a phoneme boundary, the effect on the phoneme identification unit 4 and word identification unit 6 will be large depending on how the error occurs.) In other words, the final speech recognition rate depends on each process. A high recognition rate could not be obtained because it decreased in proportion to the product of error rates.

In addition, especially when constructing a recognition device targeted at unspecified speakers, it is extremely difficult to set threshold values for determination in each process.

If you set a threshold so that the target word exists among the candidates, the number of candidate groups in each process will increase, making it extremely difficult to accurately narrow down the target word from among multiple candidate words. In addition, when trying to use a recognition device in a real environment, there was a considerable amount of non-stationary noise, and the recognition rate was low even with a recognition device for a small number of words (,
In fact, it had become a useful item.

[Means to solve the problem]

According to the present invention, in order to solve the above-mentioned conventional problems, the speech section of each word is cut out by the spotting method,
By providing two steps: selecting candidate words and then performing matching on a phoneme basis, candidate words can be selected and speech sections can be extracted all at once, and candidate words can be narrowed down easily.

Furthermore, according to the present invention, by preparing standard patterns of phonemes under a plurality of environments, it is possible to recognize input speech in a large number of situations with a smaller amount of information than when standard patterns of words are prepared for a plurality of environments. becomes possible.

[Embodiment 1] FIG. 1 is a basic configuration diagram of a speech recognition system according to the present invention, in which 100 is a speech input section, and 101 is a speech analysis section that analyzes and compresses input speech and converts it into a time series of feature vectors. , 103 is a word standard pattern storage unit that stores standard patterns obtained from word data uttered by a large number of speakers, and 102 is a word standard pattern storage unit that stores the standard patterns obtained from word data uttered by many speakers. A word distance calculation unit using continuous Mahalanobis DP calculates the distance of a standard pattern using continuous Mahalanobis DP for each frame of input data; 104 is a word distance calculation unit using continuous Mahalanobis DP, which calculates a distance between each frame and the word standard pattern calculated by continuous Mahalanobis DP; 105 is a candidate word discrimination unit that discriminates candidate words from standard patterns; 105 is a parameter storage unit that stores parameters of feature vectors of one or more word sections that have become candidates; 106;
107 is a phoneme standard pattern storage unit that stores standard patterns created in units of phonemes from voices uttered by multiple speakers;
108 is a phoneme distance calculation unit using continuous Mahalanobis DP that calculates the distance between the input data and the phoneme standard pattern using continuous Mahalanobis DP on a phoneme basis for the feature vector series of candidate words, and 108 performs matching for each of one or more candidate words. An identification unit based on recognition results of phoneme units that identifies and outputs the most appropriate word from each phoneme string.

Reference numeral 109 denotes a result output unit that outputs a voice recognition result by means such as voice response. In the figure, the first part shows the extraction of speech sections and the narrowing down of word candidates, and the second part shows the recognition part of phoneme units within the candidate words.

Next, the flow of operation will be explained. First, the audio input section 100 inputs an audio signal from a microphone and transfers the input waveform to the audio analysis section 101 . The audio input unit 100 always takes in audio or ambient noise signals, etc. during the audio input reception time, and
The audio input waveform is converted into a digital value and transferred to the audio analysis unit 101 as a waveform. The speech analysis unit 101 analyzes the constantly input waveform with a window width of about 10 m5 ec to 30 m5 ec, and calculates the feature parameter for each frame with a length of 2 m5 ec to 10 m5 ec, which is relatively fast considering the type of feature parameter. LPC that can be analyzed in
Cepstrum, LPC melkebstrum, FFT cepstrum, FFT if you want to extract parameters with high precision
Melkebstrum etc. are common, and there are also filter bank output values.

In addition, each frame is analyzed using the parameters most suitable for the usage status of the system by using normalized power information or by multiplying each parameter dimension by a weighting coefficient. Next, compression is performed on the dimensions of the analyzed feature parameters. Cepstral parameters are usually determined by the required number of dimensions (for example, 6
dimension) and use this as a feature vector.

When the filter bank output is used as a feature parameter, the dimension is compressed by orthogonal transformation such as -L transformation or Fourier transformation, and low-order terms are used. These compressed parameters for 17 rems will be referred to as feature vectors, and the time series of feature vectors after dimension compression will be referred to as feature vector series (or simply parameters).

In this example, the analysis window length was 25.6 m5 ec, the frame period was 10 m5 ec, and the mel cepstral coefficients were obtained from the spectrum envelope passing through the peak of the FFT spectrum. Let it be a feature vector for one frame. Here, the zero-order term of Melkebstrum represents power.

Next, a method for creating standard patterns to be stored in the word standard pattern storage section 103 will be described. In this system, for example, the 10 numbers "Zero, San, Two, Rei, Nana, Yon, Go, Maru, Shi, Roku, Yu, Hachi, Shichi, Kyu, Ichi" and "Hai, Iie" including vocalization variations are used. We will describe the recognition of a total of 17 words. Standard patterns are created from word sounds uttered by multiple speakers. In this embodiment, voice samples from 50 people are used to create a standard pattern for one word. (The larger the number of audio samples, the better.) FIG. 2(a) shows a flowchart showing the procedure for creating a standard pattern.

First, select a core pattern that will be a temporary comparison target when creating a standard pattern from a voice sample (S
200). The selection method uses the word with the most average utterance duration and utterance pattern among the 50 words. Next, input a sample word (5201), perform time axis expansion/contraction matching between the input word and the core pattern, and calculate the average vector and variance/covariance for each frame along the matching path that minimizes the time normalized distance. Create a matrix (S202)
. Here, DP matching is used as a time axis expansion/contraction matching method. Next, change the speaker numbers of the input words one after another (S204) Words Si for 50 people (i=1 to 50)
For each frame, the average value of the feature vector and the variance-covariance matrix are determined (S203.5205). In this way, word standard patterns are created for a total of 17 words in the same manner as above, and the word standard pattern storage unit 1
Store it in 03.

The word distance calculation unit 102 using continuous Mahalanobis DP matches the time series of feature vectors input one after another using continuous Mahalanobis DP with the standard patterns of all words stored in the word standard pattern storage unit 103. and calculate the distance.

Here, continuous Mahalanobis DP will be explained. The continuous DP method is common and is a method of searching for a target word, syllable, or other unit from a sentence continuously uttered by a specific speaker. This is called word spotting.
This is an innovative method that performs recognition at the same time as cutting out the desired audio section. In this example, continuous DP
Unspecificity is absorbed by using the Mahalanobis distance for the distance within each frame of the method.

Figure 2 (b) shows the standard pattern of the word “zero” and “
This figure shows the results of matching the input speech when the word "zero" was uttered, including silent sections, analyzed into a time series of feature vectors using continuous Mahalanobis DP. The areas where the distance between the standard pattern and the input pattern is large, and the areas where black is thin and close to white are where the distance between the standard pattern and the input pattern is small. Below the matching result is the cumulative distance time. This cumulative distance indicates the distance of the DP path that ends at that point in time.The DP path is calculated and its value is stored in memory.The DP path stored in this memory is For example, this figure shows the DP path when the distance is the minimum, but if the standard pattern and the input pattern are similar, the cumulative distance is smaller than an arbitrarily determined threshold. Recognize words in the standard pattern as candidate words.Then, in order to extract speech sections from the input pattern,
From the point where the cumulative distance is smaller than the darkness value and is also the minimum, D
By recalling the P path from memory and backtracking, the start of the voice section is determined. The time series of feature vectors of the voice section thus obtained is stored in the parameter storage unit 105.

The processing system described so far first obtains a candidate word, a series of feature vectors obtained by analyzing its speech interval, and cumulative distance results using continuous Mahalanobis DP. Here, when multiple candidate words with overlapping phonetic intervals, such as "nchi" and "shi", are selected, in this case "shichi"
Select the one and discard “shi”. “Roku” and “Ri”
Similarly, most of the vocal section of “ri” (here, 80
% or more) When it is included in "Roku", "Ri" is discarded and only "Roku" is verified.

In this embodiment, the vowel (a;
i, u, e, o) and consonants (z, s, n.

rSg, m) sh i, ni, h, c i) A standard phoneme pattern is created in advance using the same method as the word standard pattern storage section 103. Phoneme distance calculation unit 107 using continuous Mahalanobis DP
Then, matching of each phoneme is performed for the voice section cut out as a candidate word stored in the parameter storage unit 105.

Similar to the word distance calculation unit 102 using continuous Mahalanobis DP, the section of the phoneme is calculated from the position where the cumulative distance is the minimum. (Similar to the candidate word discriminator 104, the point where the cumulative distance is the minimum is the end of the phoneme, and the start is the continuous D
In this embodiment, for example, if "zero"#"zero" is a candidate word, "Ze"r"0" 4 for that speech section.
Matching is performed only for different types of phonemes. As a result of matching the four types of phonemes with the above-mentioned "zero" and the candidate speech interval, we will calculate the positional relationship of the point where the cumulative distance of each phoneme is the minimum and the average value of the minimum distance. It is shown in FIG. 2(C).

The minimum value of the distance as a result of matching for each phoneme and its position are expressed in a frame and sent to the recognition unit 108 based on the recognition results for each phoneme. In this example, the minimum value for "Z" is "J" and the 1st frame position is "Z,".The recognition unit 108 based on the recognition result of each phoneme receives the information sent from the phoneme distance calculation unit 107 using continuous Mahalanobis DP. The final word is identified based on the data.First, the order (frame position) of the phoneme string of the candidate word is z, <6.<r+<
Check whether it is O+. In this order, the recognized word is "zero". If it is smaller than the average H, "zero" is output as the recognition result.

FIG. 2(d) shows the output result of word candidates (candidate word discriminator 1
04) is shown.

■ is the word “Hachi”, ■ is the word “Shichi”, ■ is the word “N”
is output as a candidate. However, as mentioned above, 80% or more of ``■'' is included in the interval ``■'', and the same ``'' exists in ``■'', so identification at the phoneme level is performed for ``■■''.

Case ■ Phoneme sequence “1hlalc” of word S1 and word S
As a result of matching for the phoneme sequence "1shli clil" of 2, if the order of the phonemes is the same as that of the candidate word, and the distance of each phoneme is smaller than H (threshold), the one with the smaller average cumulative distance X, Output.

Case ■ If the order is different in both cases, but the distance between individual phonemes is smaller than the threshold (H), DP matching is performed using the character string of the middle word and the phoneme string. It is determined based on the distance threshold (1).

Case ■ Is the order correct or the threshold of each phoneme (
If H) is not cleared, medium reject case ■ If the order is different and the phoneme threshold is not cleared, the word identification method based on the recognition result of the phoneme unit is not limited to the above method. As will be described later in other embodiments, the value of the threshold H used for phoneme discrimination can be determined by defining the unit of phoneme and creating a standard pattern, or by preparing a plurality of the same phoneme. The identification algorithm is different. Therefore, the identification algorithm, such as which one to give priority to, the average cumulative distance or the phoneme order, cannot be uniquely determined.

) The result output unit 109 outputs, for example, speech (word) output as the final result by the recognition unit 108 based on the recognition result of the intermediate level. When recognizing only voice information from a telephone or the like, the recognition result is confirmed using, for example, voice synthesis means, such as, "Isn't it 'zero'?" As a result of word identification, if the distance is sufficiently small, the process moves to the next corresponding process without checking the recognition result.

[Example 2] In the above-mentioned Example 1, standard patterns were created for all the phonemes included in the word to be recognized as the recognition result of the second half of the phoneme unit. However, depending on the type of phoneme, it is subject to severe deformation due to differences in the surrounding phonetic environment, speakers, etc. Therefore, if you prepare multiple phonemes with different patterns for the same phoneme, you can obtain more accurate recognition results.For example, for vowel 11, as seen in "ichi"hachi", it is possible to obtain a more accurate recognition result. There are many cases where the voice is muted. Recognition at the phoneme level uses candidate words and
Since it is necessary to strictly test and produce results in that speech interval, several types of standard patterns are created for the vowel fil, 111 for voiced and 11 for unvoiced.The same is true for other phonemes, for example. Some, such as 1g, have a bus part and others do not.

However, when creating standard patterns for these phonemes,
In order to create at least one standard pattern, approximately n2+α audio data are required, assuming that the number of dimensions of the feature vector of each frame is n.

[Example 3] Further, as another example of identifying by phoneme unit, better results can be obtained by changing the unit of phoneme. In Example 1, la
l lit... (As shown in mnl lrl, vowels and consonants, which are quite small units of speech, were treated separately.

In fact, apart from announcers and the like, in daily life, continuous word sounds uttered by humans are rarely uttered clearly with the characteristics of individual phonemes.

Looking at the data, the part where it can be determined that this is lal and this is 1ml is quite short in terms of time, and most of it is the articulatory junction. (The articulatory junction is the part where, for example, when you say "ia", the steady part of "i" transitions to the steady part of "a" (halfway).) Therefore, the unit of phoneme is the articulatory junction. A VCV type including
When CV is used for word beginnings, even when multiple candidate words described in the first embodiment appear, the proportion of cases in which they appear in different orders is reduced, making it easier to determine the final output word. (■...Vowel SC...Consonant Con5
In onant, vCv is a vowel-consonant-vowel chain)
Of course, the standard vCv pattern is created from samples extracted from continuous audio.

[Embodiment 4] In the above embodiment, the multiplication of phoneme patterns stored in the phoneme standard pattern storage unit 106, the definition of phoneme units, and the method were described.

The same can be said about the word standard pattern storage section 103. However, when it comes to word standard patterns, attempting to strictly categorize them may result in too many patterns. Furthermore, since it is not easy to collect and analyze utterance samples of many speakers for each word, here, the categorization is performed based on the length of utterance time of each word. In the first stage of this recognition system, it is a prerequisite that 100% of the candidate words contain the target word. This is because this method basically performs time expansion/contraction matching, so there is a high possibility that the word will be rejected if the length of the utterance is extremely different from the standard pattern.

Therefore, at least for the recognition device, the time length of the voice uttered by the cooperative speaker is investigated, and the standard pattern is multiplied so as to cover the entire time length.

When performing multiplexing, it is difficult to obtain samples of extremely long utterances (because it is difficult to obtain samples of extremely long utterances), the number of frames of the average feature vector may be doubled or tripled as shown in FIG. 2(e).

FIG. 2(e) shows an example in which the utterance time length of the standard pattern in which the unit is the phoneme la1mlul "am" is doubled.

When expanding Hatsufujiyamaga, care must be taken when, for example, plosive consonants such as lpl, Ml, and lkl are included. As shown in this example, even if the utterance time length of some consonants becomes longer, the utterance time length of the consonant part does not change much. Therefore, having a means to individually change the method of expansion depending on the consonant using a table or the like would make it easier and more accurate.
Standard patterns with different time lengths can be created.

A better standard pattern can be created by actually collecting voice samples with a long utterance time and creating a standard pattern from this data.

FIG. 2(f) is a table showing the overlapping magnification of consonant frames when the standard pattern length is expanded by overlapping one frame of a vowel twice, three times, and four times. Figure 2 (g)
Figure 3 shows the situation when the log (vowel) magnification is set to 3x.

Furthermore, the word standard pattern storage section 103 shown in FIG. 1 is not limited to word units. It can be a phrase unit or a chain of meaningless syllables. In this case, the unit of word standard pattern storage 103 is (VCV, VCVCV, cv, vv, cvcv, -
・etc.), and the unit (CV
, VC, V, etc.).

[Example 5] In Example 1, the second example of the basic configuration of the processing system shown in FIG.
In the first part, the candidate words obtained as the output of the first part are further divided into finer phoneme units (for example, C1V, CV, cvcSvc
We have described a method for performing spotting processing such as continuous DP using a software (e.g., v) and outputting the results. However, in this embodiment, a method other than spotting will be described as a method for recognizing candidate words output in the first part in units of phonemes. This method involves matching a word created by connecting standard phoneme patterns obtained from multiple speech samples according to the phoneme sequence of a candidate word with a feature vector of the input speech extracted as a speech interval. This method also provides a high recognition rate.

FIG. 4 shows the basic configuration of the phoneme-based recognition processing system in this embodiment.

The candidate words discriminated by the candidate word discriminator 10.4 in FIG. 1 and the feature vectors of the input speech cut out as speech sections are thereafter processed in the configuration shown in FIG. 4. First, the feature vector of the input speech is sent to the parameter storage section 105, and the candidate words are sent to the standard pattern generation rule section 110.

The standard pattern generation rule unit 110 connects the phoneme standard patterns in the phoneme standard pattern storage unit 106 according to the phoneme sequence of the candidate word, and performs pattern matching between this and the feature vector of the input speech stored in the parameter storage unit 105. This is performed in the pattern matching section 111.

Result output unit 1 outputs the speech recognition results obtained by pattern matching.
Output from 09.

A detailed configuration diagram of the standard pattern generation rule section 110 is shown in FIG. First, the phoneme sequence of the candidate word output as a result of the first part and the feature vector of the input speech cut out as a speech section are output.

Here, for example, when you input "tokus imasi (Tokushima cloth)", "tOkusim" is selected as a candidate word.
asi f ukus imas 1 (Fukushima cloth)
We will describe the process when the three words ``hiroshimasi (Hiroshima cloth)'' are selected. First, these candidate words are divided into phonemes that are optimal for continuous speech recognition in the standard pattern generation rule unit 110. In the example,
The phoneme at the beginning of a word and the CV (vowel in a consonant), and the phoneme at the middle and end of the word are VCV (vowel semi-consonant + vowel).

Next, the length of the characteristic parameter of the input speech is divided by the number of phonemes, and the average duration length per mora is determined by the average duration detection unit 152. Used when selecting a suitable phoneme standard pattern from among them.

FIG. 6(a) is an example in which a word output as a candidate word is divided into phoneme symbol strings by the phoneme division processing unit 150. FIG. 6(C) is a correspondence table between each phoneme and the address of the memory where the standard pattern is stored. The phoneme position label addition unit 151 selects from among a plurality of phoneme standard patterns in correspondence with the phoneme position of the candidate word. The type of phoneme, D2, is the duration of the standard phoneme pattern, D.

are types of phoneme standard patterns in a plurality of situations; for example, the standard pattern for the phoneme lal is entered from address 001-1. Further, the address 001-1.1 contains a standard pattern of devoiced lal. VCV-type phonemes such as 1aSa1 include the standard ones included in address 931-1, as well as ■VCV-type phonemes with the entire CV devoiced (VC
V) is in 931-1.1, in VCV, the sound (VCV) where CV sound is devoiced is in 931-1.2, in VCV, V
The sound (VCV) in which the C sound is devoiced is included in 931-1.3. In addition to this, for each phoneme unit,
It has multiple standard patterns.

FIG. 6(b) shows the phoneme when the time length (D2) of the phoneme standard notation of three candidate words is 1, and its address is associated with the selected phoneme. Here, based on the rule that ``the beginning and end of a word should be considered including the pattern in which the vowel becomes devoiced,'' the word ``tokushimasi'' is chosen as shown in Figure 6 (b).
4 shown in Figure 6(d) using the address of the phoneme shown in Figure 6(d).
You can create combinations of street patterns.

You can't connect if you don't. Type of standard phoneme pattern, D
FIG. 6(e) shows combinations that can be connected by . FIG. 6(e) shows only the time length D2 and type of a standard pattern of a certain phoneme. For example, b/b in the top row is a certain time length (b
=to(), which is voiced, indicates the connection between b and comrades. The next row, b/b, 2, is a voiced standard pattern of a certain phoneme of a certain length of time (set as b), and a standard pattern of a certain phoneme of a certain length of time (set as b-). Since the first half of is a voiced sound and the second half is an unvoiced sound, and the first half of the phoneme b and 2 are equal, there is no need to show the time length D2 of the standard pattern of phonemes. Since the average duration per mora is determined by the mora utterance time length detection unit 152, this becomes b, which is constant within the word.

However, what is shown in FIG. 6(e) is only a part of the phoneme combination rules, and there are many other acoustic phoneme combination rules used when producing speech. In FIG. 6(d), “toku s i
Although only the combination of "ma s i" is shown, combinations are created for other candidate words in the same way.Once the combination of phoneme standard patterns is created, the phoneme standard patterns are connected in the phoneme standard pattern connection section 153, and the word A standard pattern is created.Connection methods include direct connection, linear interpolation, etc., and an example of connecting phonemes 0.P, Q, and R is shown in FIG. 5 and will be described below.

(a) of FIG. 7 is an example of direct connection to generate the word 0PQR, and (b) is an example of the phoneme 0. The vowel parts are cut out from P, Q, and R by a few frames as interpolated parts, and are then used as Q' and P'.
, Q', and R', and the blank parts are filled in while linearly interpolating the parameter elements of each dimension to generate a continuous word standard pattern. Phoneme interpolation methods are suitable or unsuitable depending on the nature of the parameters, so
Here, we will use the interpolation method that is most suitable for the parameters.

Finally, the plurality of word standard patterns output from the phoneme standard pattern connection section 153 and the input pattern are matched in the pattern matching section 111, and the word with the minimum distance is outputted from the result output section 109 as, for example, speech.

Although there are many pattern matching methods such as linear expansion/contraction and DP matching, good results can be obtained with DPP matching.

Here, as the distance measure, a statistical distance measure such as Mahalanobis distance is used.

〔Effect of the invention〕

As explained above, by performing word spotting by continuous Mahalanobis DP on a word-by-word basis in the first step, it is possible to simultaneously select candidate words and cut out speech sections.

By performing matching on a phoneme basis in the second stage, a high recognition rate can be obtained because recognition is performed in two stages.

In addition, since the standard patterns under multiple environments are made in units of phonemes (not words), the amount of information can be reduced.

Furthermore, in the second stage, only the phonemes corresponding to the candidate words are matched, so there is an advantage that it does not take much time.

[Brief explanation of drawings]

Figure 1 is a basic configuration diagram of the processing system of the first embodiment of the present invention. Figure 2 (a) is a flowchart showing the flow of standard pattern creation operations. Figure 2 (b) is a continuous Mahalanobis DP. Figure 2 (C) is a diagram showing the state of phoneme matching, Figure 2 (d) is a diagram showing the relationship between multiple candidate words and input signals, and Figure 2 (e) is a diagram showing the utterance time length. Figure 2 (f) is a diagram showing the magnification corresponding to the phoneme due to the change in the magnification of the utterance duration, and Figure 2 (g) is the standard pattern that is doubled. A diagram showing what happens when the utterance time length is tripled according to the multiplication factor. Figure 3 is a configuration diagram of a conventional speaker-independent speech recognition system. Figure 4 is a configuration diagram of the second phoneme recognition process of the present invention. , Figure 5 is a configuration diagram of the standard pattern generation rule section, Figure 6 (a)
Figure 6(b) is a diagram showing the phoneme decomposition of a candidate word, Figure 6(b) is a diagram showing the addresses of standard patterns for each phoneme of a candidate word, and Figure 6(C) is an example of addresses depending on the type of phoneme standard pattern. Figure 6(d) is a diagram showing a combination of generated standard patterns, Figure 6(e) is a diagram showing an example of a combination of standard patterns that can be connected, and Figure 7 is a diagram showing an interpolation method. . In the figure, 1 is a speech input device, 2 is a speech analysis section, 3 is a phoneme boundary detection section, 4 is a phoneme identification section, 5 is a phoneme standard pattern storage section, 6 is a word identification section, 7 is a word dictionary, and 8 is a correction section. Rules section, 9
100 is a recognition result display section, 100 is a speech input section, 101 is a speech analysis section, 102 is a distance calculation section using continuous Mahalanobis DP, 103 is a word standard pattern storage section, 104 is a candidate word discrimination section, 105 is a parameter storage section, 106 is a phoneme standard pattern storage unit, 107 is a distance calculation unit using continuous Mahalanobis DP, 108 is an identification unit based on the recognition result of each phoneme,
109 is a result output section, 110 is a standard pattern generation rule section,
111 is a pattern matching section, 150 is a phoneme division processing section, 151 is a phoneme label adding section, 152 is a monomolar utterance time length detection section, and 153 is a phoneme standard pattern connection section. Figure 1 Basic configuration of the processing system ``]Eco]+09 Figure 2 (a) Standard pattern creation flow Figure 2 (c) Phoneme matching Figure 2 (d) Relationship between multiple candidate words and input signals Figure 2 (e) Standard pattern with twice the utterance time length Figure 2 (f) Magnification corresponding to phoneme due to change in utterance time length Figure 2 (9) utterance time length tripled Fig. 4 Configuration diagram of the second phoneme recognition process of the present invention Fig. 6 (a) Situation of phoneme decomposition of a candidate word Fig. 6 (b) Address of the standard pattern of each phoneme of the candidate word Fig. 6 ( d) Combination of generated standard patterns Figure 6 (e) Example of combination of standard patterns that can be connected Good ---Hehe-- (a) Interpolation force -P -Q -R

Claims (6)

[Claims] (1) Speech analysis means for analyzing input speech to obtain a time series of feature vectors, word standard pattern storage means for storing word standard patterns obtained from a plurality of speech samples, and applying a spotting method to the input speech feature vector time series. candidate word identification means for detecting a speech interval and selecting a candidate word from the word standard patterns by using a phoneme standard pattern storage means for storing a phoneme standard pattern obtained from a plurality of speech samples; A recognition means for recognizing the input speech by matching a time series of feature vectors of the input speech with the phoneme standard pattern of the candidate word, and an output means for outputting a result recognized by the recognition means. speech recognition device. (2) The candidate word identification means further includes a statistical distance measure;
Distance calculation means that performs continuous DP using Mahalanobis distance and calculates the cumulative distance of DP paths; Storage means that stores the DP path; A point in time when the cumulative distance is smaller than a preset threshold and is minimum is determined as the end point. 2. The speech recognition apparatus according to claim 1, further comprising speech segment recognition means for reading the DP path from the storage means, determining the starting end of the DP path, and recognizing the speech segment. (3) The matching with the input speech is performed using a spotting method of the phoneme standard pattern corresponding to the candidate word in the speech section (
1) The speech recognition device according to item 1). (4) The matching with the input speech is performed with a standard pattern generated by connecting the phoneme standard pattern according to the phoneme string of the candidate word by a standard pattern generation rule means. speech recognition device. (5) The phoneme units stored in the phoneme standard pattern storage means are CV (consonant-vowel), VCV (vowel-consonant-vowel), and VV (vowel-vowel). ).The speech recognition device described in ). (6) The speech recognition device according to claim 1, wherein the phoneme standard pattern includes a plurality of standard patterns depending on factors such as a speaker, a speaking time, and a speaking environment.