JP2862306B2 - Voice recognition device - Google Patents

Description

Description: FIELD OF THE INVENTION The present invention relates to a speech recognition method, and more particularly to a speech recognition method capable of recognizing speech generated by an arbitrary speaker.

[Conventional technology]

Several recognition methods related to unspecified speaker recognition have been devised, but a conventional example of an unspecified speaker recognition system having the most common configuration at present and relatively close to the present proposal will be described.

Conventionally, a recognition system aiming at an unspecified large vocabulary has a configuration as shown in FIG. The voice input from the voice input unit 1 is used by the voice analysis unit 2 to obtain a filter bank output including a power term of the voice and characteristic parameters such as an LPC cepstrum. Here, parameter compression and the like (in the case of a filter bank output, K -Dimensional conversion by -L conversion or the like). (Because the analysis is performed on a frame basis, the feature parameter after compression is hereinafter referred to as a feature vector.) Next, a process for determining a phoneme boundary from continuous speech is performed by the phoneme boundary detection unit 3. In the phoneme identification unit 4,
Phonemes are determined by a statistical method. A phoneme standard pattern storage unit 5 stores phoneme standard patterns created from a large number of phoneme samples. Reference numeral 6 denotes a word dictionary 7 based on the output result of the phoneme identification 4 or a correction restricting unit 8 from the output candidate phonemes.
Is a word identification unit that outputs the final recognition result after the correction, and 9 is a recognition result display unit that displays the recognition result.

Usually, the phoneme boundary detection unit 3 uses a discriminant function or the like, and the phoneme discrimination unit 4 performs the same determination. Generally, as the output of each of these components, a candidate satisfying a certain threshold is output. Although a plurality of candidates are further output for each candidate, top-down information such as 7 or 8 is used to narrow down the word to the final point.

[Problems to be solved by the invention]

However, since the basic configuration of the above-described conventional recognition apparatus is a bottom-up type, if an error occurs at a certain point in the recognition / process, it is likely to have a bad influence on a subsequent process. (For example, in the phoneme boundary detection unit 3,
If the phoneme boundary is incorrect, the effect on the phoneme discriminator 4 and the word discriminator 6 is large depending on the way of the error. That is, the final speech recognition rate decreases in proportion to the product of the error rates in each process. , High recognition rate could not be obtained.

In particular, when configuring a recognition device for an unspecified speaker, it is very difficult to set a threshold value for determination in each process. If the threshold is set so that at least the target exists in the candidates, the number of candidate groups in each process increases, and it becomes very difficult to accurately narrow down the target word from a plurality of candidate words. I was In addition, when an attempt is made to use a recognition device in a real environment, non-stationary noise and the like are considerably large, and even a recognition device for a small number of words has a low recognition rate and is actually difficult to use.

Means for Solving the Problems In order to solve the above conventional problems, the present invention provides a speech recognition method using a word dictionary storing dictionary word data and a phoneme dictionary storing dictionary phoneme data. And
By inputting voice data and matching the dictionary word data stored in the word dictionary with the input voice data using a spotting method, the voice section of the input voice data corresponds to the voice section. Selecting a candidate word, selectively extracting phoneme data for a dictionary of phonemes that can constitute the selected candidate word from the phoneme dictionary, matching the input phonemic data with the input voice data of the selected voice section, and A voice recognition method for determining a recognition result of the input voice data based on a result of matching between phoneme data for use and input voice data.

Embodiment 1 FIG. 1 is a basic configuration diagram of a speech recognition system according to the present invention. 100 is a speech input unit, 101 is a speech analysis unit that analyzes and compresses inputted speech and converts it into a time series of feature vectors. ,Ten
3 is a word standard pattern storage unit for storing standard patterns obtained from word data uttered by many speakers, 102 is a feature vector sequence of the speech analysis unit 101 and each standard stored in the word standard pattern storage unit 103 A word distance calculation unit using a continuous Mahalanobis DP that calculates a pattern using a continuous Mahalanobis DP for each frame of input data, and a word standard calculation unit 104 uses a word standard value based on a distance value between each frame obtained from the continuous Mahalanobis DP and a word standard pattern. A candidate word discriminating unit for discriminating candidate words from patterns, 105 is a candidate word
A parameter storage unit that stores parameters of feature vectors of one or more word sections, a phoneme standard pattern storage unit that stores a standard pattern created in phoneme units from speech uttered by many speakers, and 107 is a candidate A continuous Mahalanobis DP phoneme distance calculation unit that calculates a distance between input data and a phoneme standard pattern using a continuous Mahalanobis DP for each phoneme unit for a feature vector sequence of a word that has become a word, and 108 is a phoneme distance calculation unit for each of one or more candidate words. An identification unit that identifies and outputs the most appropriate word from a phoneme sequence based on a recognition result in phoneme units. Reference numeral 109 denotes a result output unit that outputs a voice recognition result by means such as a voice response. In the figure, the first
The section narrows down word candidates along with voice segmentation,
The second part shows a recognition unit for each phoneme in the candidate word.

Next, the operation flow will be described. First, the phoneme input unit 100
Receives an audio signal from a microphone and transfers the input waveform to the audio analysis unit 101. The voice input unit 100 always captures voice or a surrounding noise signal during the reception time of voice input, and transfers the voice input waveform to the voice analysis unit 101 as a digital value converted waveform. The voice analysis unit 101 analyzes a waveform that is always input with a window width of about 10 msec to 30 msec,
For each frame having a length of 2 msec to 10 msec, the types of feature parameters for obtaining feature parameters are LPC cepstrum and LPC mel cepstrum, which can be analyzed at relatively high speed. Cepstrum and the like are common, and there is also a filter bank output value.

The normalized power information is used, or a weighting coefficient is multiplied for each dimension of the parameter, and the parameter is analyzed for each frame with a parameter most suitable for the usage state of the system. Next, compression is performed on the dimensions of the analyzed feature parameters. The cepstrum parameter is usually the required number of dimensions (eg, 6
(Dimension)), and this is used as a feature vector. When the filter bank output is used as a feature parameter, for example, K
-Dimensionally compressed by orthogonal transformation such as L transformation or Fourier transformation, and use lower order terms. These compressed parameters for one frame are called a feature vector, and the time series of the dimensionally compressed feature vectors are called a series of feature vectors (or simply, parameters).

In this embodiment, the analysis window length is analyzed at 25.6 msec, the mel cepstrum coefficient is obtained from the spectrum envelope passing through the peak of the FFT spectrum at a frame period of 10 msec, and then the coefficient 2
The next to sixth orders are used as feature vectors for one frame. Here, the 0th order term of the mel cepstrum represents power.

Next, a method of creating a standard pattern stored in the word standard pattern storage unit 103 will be described. In this system, as an example, 10 digits including vocal transformations “Zero, San, Ni, Ray, Nana, Yon, Go, Maru, Shi, Roku, Ku, Hachi, Shichi, Kiyu, Ichi” and “High”, “No” The standard pattern is created from word voices uttered by many speakers, and in this embodiment, voice samples of 50 people are used to create a standard pattern of one word (voice sample). FIG. 2 (a) shows a flowchart illustrating a procedure for creating a standard pattern.

First, a core pattern (core pattern) that is a temporary comparison target when a standard pattern is created from a voice sample is selected (S200). The selection method uses the word with the longest utterance time and utterance pattern among 50 words. Next, a sample word is input (S201), time-axis expansion / contraction matching between the input word and the core pattern is performed, and an average vector and a variance share are set for each frame along a matching path that minimizes the time-normalized distance. A variance matrix is created (S202). Here, DP matching is used as a method of time axis expansion matching. Next, change the speaker number of the input word one after another (S204)
For 50 words Si (i = 1 to 50), the average value of the feature vectors and the variance-covariance matrix are obtained for each frame (S203, S205). In this way, a word standard pattern is created for each of the 17 words in the same manner as described above, and stored in the word standard pattern storage unit 103.

In the continuous Mahalanobis DP word distance calculation unit 102,
The continuous Mahalanobis with the standard patterns of all the words stored in the word standard pattern storage unit 103 for the time series of the feature vectors successively input by the continuous Mahalanobis DP
Perform matching by DP and calculate the distance.

Here, the continuous Mahalanobis DP will be described. Continuous
The DP method is a general method in which a target word or a unit such as a syllable or the like is searched for in a sentence continuously uttered by a specific speaker. This is called word spotting, and is an epoch-making method in which recognition is performed simultaneously with extraction of a target voice section. In this embodiment, the unspecifiedness is absorbed by using the Mahalanobis distance as the distance in each frame of the continuous DP method.

FIG. 2 (b) shows a continuous Mahalanobis DP of a standard pattern of the word “zero” and a time series of feature vectors including an unvoiced section of the input voice when the word “zero” is uttered. It shows the result of matching. In the figure, the places where the black pattern appears dark are where the distance between the standard pattern and the input pattern is large, and the places where the black pattern is light and close to white are where the distance between the standard pattern and the input pattern is small. The time change of the cumulative distance is shown below the result of the matching. This accumulated distance indicates the distance of the DP path ending at that time. The DP path is obtained and the value is stored in the memory. The DP path stored in this memory is used to find the beginning of the voice section. For example, in this figure, the DP path when the distance is minimum is shown, but if the standard pattern and the input pattern are similar, the cumulative distance will be smaller than the arbitrarily determined threshold, and the words of the standard pattern will be candidates. Admit it as a word. Then, in order to cut out a voice section from the input pattern, the DP path is called from the memory and backtracked from the time when the accumulated distance is smaller than the threshold value and is further minimum, and
The beginning of the voice section is determined. The time series of the feature vector of the voice section thus obtained is stored in the parameter storage unit 105.

With the processing system described so far, first, candidate words,
A series of feature vectors obtained by analyzing the voice section and a result of the cumulative distance by the continuous Mahalanobis DP are obtained. Here, when a plurality of candidate words having overlapping voice sections, such as “Shichi” and “Shii”, are selected, in this case, “Shichi” is selected and “Shishi” is discarded. Similarly, when “Roku” and “G” are included in “Roku” when most of the voice section of “G” is included (here, 80% or more),
"K" is rounded down, and only "Roku" is verified.

In the present embodiment, the vowels (a,
i, u, e, o) and consonants (z, s, n, r, g, m, sh)
A phoneme standard pattern is created for i, k, h, and ci). The creation method is created in advance in the same manner as the word standard pattern storage unit 103. The phoneme distance calculation unit 107 based on the continuous Mahalanobis DP performs matching with each phoneme in a speech section cut out as a candidate word stored in the parameter storage unit 105.

Similar to the word distance calculation unit 102 using the continuous Mahalanobis DP, the section of the phoneme is calculated from the position where the cumulative distance is minimized. (Similar to the candidate word discriminating unit 104, the time point at which the cumulative distance becomes the minimum is taken as the end of the phoneme, and the starting point is obtained by the backtracking of the continuous DP path.) In this embodiment, for example, “zero” → “zero” In this case, four “z”, “e”, “r”, “o”
Matching is performed only for each type of phoneme. As a result of matching of the four phonemes and the above-mentioned "zero" and matching of candidate voice sections, the positional relationship and the average value of the minimum distances at the point where the cumulative distance of each phoneme is the minimum are shown. This is shown in FIG.

The minimum value of the distance as a result of matching for each phoneme and its position are represented by a frame and sent to the recognition unit 108 based on the recognition result for each phoneme. In this example, the minimum value of “z” is “j” and the frame position is “z _f ”. The recognition unit 108 based on the recognition result for each phoneme performs final word recognition based on the data transmitted from the phoneme distance calculation unit 107 using the continuous Mahalanobis DP. First, the order of the phoneme sequence of the candidate words (position of the frame) is examined whether or not _{z f <e f <r f} <o f. If this order, the recognized word is "zero" (zer
o) “Average And if the value of X is smaller than the threshold value H, "zero" is output as the recognition result.

FIG. 2D shows the output result of the word candidate (candidate word discriminating unit).
104 output result). Is the word "bee", the word "shi", and the word "shi" are output as candidates. However, as described above, since 80% or more is included in the section and the same "shi" is present in the section, discrimination at the phoneme level is performed.

Case word S ¹ of the phoneme string "| h | a | c | i |" and the phoneme string of words ^{S 2 "| sh | i |} c | i |" As a result of Matsuchingu for, both of phoneme the order is, the candidate word And the distance between the individual phonemes is smaller than H (threshold) ⇒ the smaller of the average cumulative distance X is output.

Case Both orders are different, but the distance between individual phonemes is smaller than the threshold value (H) ⇒ Perform DP matching with words and phoneme strings. The distance is determined by the threshold value (I).

Case The order is correct or the threshold of each phoneme does not clear (H) ⇒ reject case The order is different and the threshold of phoneme is not clear ⇒ reject Recognition method of words based on recognition result of phoneme unit Is not limited to the above method. As will be described later in other examples, how to define the unit of phoneme and create a standard pattern,
Alternatively, by preparing a plurality of the same phonemes, the value of the threshold value H used for phoneme determination or the identification algorithm is different. Therefore, an identification algorithm for determining which of the average cumulative distance and the phoneme rank is prioritized is not uniquely determined.

For example, a speech (word) output as a final result by the recognition unit 108 based on the recognition result for each phoneme is output by the result output unit 109. In the case of performing recognition using only voice information such as telephones, the recognition result is “zero”. ], For example, using voice synthesis means. As a result of the word identification, if the distance is sufficiently small, the process proceeds to the next process corresponding to the recognition result without checking the recognition result.

[Second Embodiment] In the first embodiment, a standard pattern is created for all phonemes included in a word to be recognized based on the recognition result of the latter phoneme unit. However, depending on the types of phonemes, phonemes undergo severe deformation due to differences in the surrounding phonemic environment, speakers, and the like. Therefore, if a plurality of phonemes with different patterns are prepared for the same phoneme in accordance with the pattern, a more accurate recognition result can be obtained. For example, for the vowel i |
There are quite a few cases where the voice is muted by the speaker as seen in "bees" and "shichi". Recognition at the phoneme level must be strictly tested in the candidate word and its voice section to produce a result, so even if the vowel | i |
For each of the unvoiced | i |, several types of standard patterns are created. The same applies to other phonemes, for example, some have a path portion such as | g | and others do not. However, when creating a standard pattern for these phonemes, in order to create at least one standard pattern, about n ² + α voice data is required, where n is the number of dimensions of the feature vector of each frame.

Third Embodiment Further, as another example of identification by phoneme unit, a better result can be obtained by changing the phoneme unit. In the first embodiment, | a
As shown in |, | i |,…, | m |, | n |, and | r |, vowels and consonants, which are quite small, were treated separately.

In fact, continuous word voices uttered by humans rarely clearly utter the characteristics of individual phonemes in everyday life apart from an announcer and the like. Even when looking at the data, the part where | a | can be determined to be | m | here is considerably short in terms of time, and most are articulation coupling parts. (The articulation coupling unit is, for example, a (halfway) part that transits from the stationary part of “a” to the stationary part of “a” when “ear” is uttered.) Therefore, the unit of the phoneme is the articulation coupling unit. VCV type including
When CV is used for the head of a word, even when a plurality of candidate words appearing in the first embodiment appear, the ratio of cases where the order is different is reduced, so that it is easy to determine the final output word. (V: vowel Vowel, C: consonant, VCV is vowel-consonant-vowel, chain) Of course, the standard pattern of VCV is created from samples cut out from continuous speech.

Embodiment 4 In the above-described embodiment, the description has been given of the multiplication of the phoneme pattern stored in the phoneme standard pattern storage unit 106, the definition of the phoneme unit, and the method.

The same can be said for the word standard pattern storage unit 103. However, with regard to the word standard pattern, if the patterns are strictly categorized, the number of patterns may be too large. In addition, since it is not easy to collect and analyze the utterance samples of many speakers for each word, here, the categorization is performed according to the utterance time length of each word. In the first stage of the recognition system, a precondition is that 100% of the target word is included in the candidate words. This method basically performs time-expansion matching, which is a word having a utterance time length extremely deviating from the standard pattern, and is highly likely to be rejected.

Therefore, at least the recognition device checks the time length of the voice uttered by the cooperative speaker and multiplies the standard pattern so as to cover the entire time length. At the time of multiplexing, it is difficult to obtain a sample of an extremely long utterance, so that the average number of feature vector frames is set to 2 as shown in FIG.
It may be doubled or tripled.

FIG. 2 (e) shows an example in which the utterance time length of the reference pattern in units of the phoneme | a | m | u |

When expanding the utterance time length, it is necessary to be careful when, for example, plosive consonants such as | p |, | t |, | k | are included. As shown in this example, even if the utterance time length is longer for some consonants, the utterance time length of the consonant part does not change so much. Therefore, if means for individually changing the enlargement method by a consonant using a table or the like is provided, it is possible to easily and accurately create a standard pattern having a different time length.

In practice, a method of collecting voice samples having a long utterance time length and generating a standard pattern from these data can create a better standard pattern.

FIG. 2 (f) shows that one vowel frame is doubled, tripled,
9 is a table showing the overlap magnification of the consonant part frame when the standard pattern length is enlarged by overlapping with the double. FIG. 2 (g) shows a state in which the magnification of the "log" (of the vowel) is set to "3 times".

Further, the word standard pattern storage unit 103 in FIG. 1 is not limited to word units. It may be a phrase unit or a chain of meaningless syllables. In this case, the unit of the word standard pattern storage unit 103 is (V
CV, VCVCV, CV, VV, CVCV,...), And the unit (CV, VC, V, etc.) of the phoneme standard pattern storage unit 106.

Fifth Embodiment In the first embodiment, in the second part of the basic processing system configuration shown in FIG. 1, candidate words obtained as outputs of the first part are further divided into phoneme units (for example, C, V, CV, CVC, VCV, etc.)
Described the method of performing spotting processing such as continuous DP and outputting the result. However, in this embodiment, a method other than spotting will be described as a method of recognizing candidate words output by the first part in phoneme units. This is a method of matching a word formed by connecting phoneme standard patterns obtained from a plurality of speech samples in accordance with a phoneme sequence of a candidate word, and a feature vector of an input speech cut out as a speech section. Even with this method, a high recognition rate can be obtained.

FIG. 4 shows the basic configuration of a recognition processing system for each phoneme in this embodiment.

The candidate words determined by the candidate word determination unit 104 in FIG. 1 and the feature vector of the input voice cut out as a voice section are thereafter processed in the configuration shown in FIG. First,
The feature vector of the input speech is sent to the parameter storage unit 105, and the candidate words are sent to the standard pattern generation rule unit 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 of the feature vector of the input speech stored in the parameter storage unit 105. Pattern matching department
Perform at 111. A speech recognition result obtained by pattern matching is output from a result output unit 109.

A detailed configuration diagram of the standard pattern generation rule section 110 is shown in FIG. First, a phoneme sequence of a candidate word output as a result of the first part and a feature vector of an input voice cut out as a voice section are output. Here, for example, “tokusima
si (Tokushima City) "as a candidate word
simasi "," fukusimasi (Fukushima City) "," hirosimasi
(Hiroshima City) "will be described. First, these candidate words are divided into optimal phonemes for continuous speech recognition in the standard pattern generation rule unit 110. In this embodiment, The phoneme and CV (consonant + vowel) at the beginning of the word are VCV (vowel + consonant + vowel).

Next, the length of the feature parameter of the input speech is divided by the number of phonemes, and the average duration per mora is obtained by the average duration detector 152. Used to select a suitable phoneme standard pattern from among them.

FIG. 6A shows 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 table of correspondence between each phoneme and an address of a memory where a standard pattern is stored. The phoneme position label adding unit 151 is to select from a plurality of phoneme standard patterns corresponding to the phoneme positions of the candidate words. If the address notation is [D ₁ −D ₂ , D ₃ ], D ₁ becomes type of phonemes, D ₂ is the time length of the phoneme standard pattern, D ₃ is the type of a plurality of conditions of the phoneme standard pattern, for example a phoneme | a | standard pattern of has entered the address 001-1. Also,
Addresses 001-1 and 001-1 contain a standard pattern of unvoiced | a |. VCV type phonemes like | asa |
In addition to the standard ones in 1-1, the sound of the entire VCV is muted. Is 931-1 and 1 is a voice in which the CV sound is devoiced in VCV Is 931-1, 2 and VCV is a voice in which VC sound is devoiced. Are in 931-1 and 3. Also, not only this one
Each phoneme unit has multiple standard patterns.

FIG. 6 (b) is a diagram in which the phoneme when the time length (D ₂ ) of the phoneme standard pattern of three candidate words is 1 is selected, and its address is corresponded. Here, from the regulation that "the beginning and end of a word include a pattern in which the vowel part is devoiced,"
The word "kusimasi" can be combined with the four patterns shown in FIG. 6D using the phoneme addresses shown in FIG. 6B.

Here, when maintaining the standard pattern, connection cannot be established unless voiced and unvoiced voices of the latter half of the preceding phoneme and the former half of the following phoneme are complete. Class of a standard pattern of the phoneme, shown in Figure 6 the possible combinations connected by D ₃ (e). The FIG. 6 (e), the time length D ₂ of the reference pattern of a phoneme type D ₃
Only shown. For example, b / b at the top indicates a connection between b and b, which is a standard pattern of a certain phoneme, has a certain length of time (b =) and is voiced. Next stage b / b, 2
Is a certain time length of a standard pattern of a certain phoneme (= b)
In the standard pattern of a certain phoneme, the first half of a certain time length (b =) is a voiced sound, the latter half is an unvoiced sound, and b and 2 are connected. Here, the type of phonemes to be connected, and the second half of the previous phoneme connections, because translation may equal the first half of the phoneme after the connection portion is not necessary to show D ₁ to FIG. 6 (e), since the time length D ₂ of the standard pattern of the phoneme, calculated in the average duration per mora in one mora utterance duration detector 152, which is b, and the constant in the word within.

However, FIG. 6 (e) shows a part of the phoneme connection rules, and there are many other acoustic phoneme connection rules when uttering speech. FIG. 6 (d) shows only the combination of "tokusimasi", but the combination is similarly created for other candidate words. When the combination of phoneme standard patterns is completed, the phoneme standard patterns are connected in the phoneme standard pattern connection unit 153, and a word standard pattern is created. The connection method includes direct connection, linear interpolation, etc.
An example of connecting P, Q, and R is shown in FIG. 5 and will be described below.

FIG. 7 (a) shows an example in which the word OPQR is directly connected and the word OPQR is generated. FIG. 7 (b) is a diagram in which vowel parts are cut out from the phonemes O, P, Q, R by interpolation as O ', P ′,
In this example, Q 'and R' are used, and blank portions of these are filled in while linearly interpolating the elements of the parameters of each dimension to generate a continuous word standard pattern. Since the phoneme interpolation method is suitable or unsuitable depending on the properties of the parameters, the optimum interpolation method for the parameters is used here. Finally, a plurality of word standard patterns and input patterns output from the phoneme standard pattern connection unit 153 are stored in the pattern matching unit 111.
, And the word having the shortest distance is output from the result output unit 109 as, for example, a voice.

There are many pattern matching methods such as linear expansion and contraction, and DP matching, and good results are obtained with DP matching. Here, a statistical distance scale represented by a Mahalanobis distance or the like is used as the distance scale.

[Effects of the Invention] As described above, according to the present invention, there is provided a speech recognition method using a word dictionary storing dictionary word data and a phoneme dictionary storing dictionary phoneme data. And matching the dictionary word data stored in the word dictionary with the input voice data using a spotting method, so that a voice section of the input voice data and a candidate word corresponding to the voice section are input. , Phoneme data for dictionary of phonemes that can constitute the selected candidate word is selectively extracted from the phoneme dictionary, and is matched with input speech data of the selected voice section, and the phoneme for dictionary is selected. By determining the recognition result of the input voice data based on the result of the matching between the data and the input voice data, a word dictionary and a phoneme dictionary,
Matching with dictionary data created in different sections is performed in two steps, and the second-stage matching uses the results obtained in the first-stage matching, so that speech recognition with high efficiency and high recognition rate is performed. Can be provided.

[Brief description of the drawings]

FIG. 1 is a diagram showing the basic configuration of a processing system according to a first embodiment of the present invention. FIG. 2 (a) is a flowchart showing the flow of an operation for creating a standard pattern. FIG. 2 (b) is a flowchart showing a continuous Mahalanobis DP. FIG. 2 (c) is a diagram showing phoneme matching, FIG. 2 (d) is a diagram showing a relationship between a plurality of candidate words and an input signal, and FIG. 2 (e) is a utterance time. FIG. 2 (f) is a diagram showing a state of a standard pattern whose length is doubled, FIG. 2 (f) is a diagram showing a magnification corresponding to a phoneme due to a change in magnification of the utterance time length, and FIG. 2 (g) is FIG. 1 (f). FIG. 3 is a diagram showing a state in which the utterance time length is tripled according to the magnification of FIG. 3, FIG. 3 is a configuration diagram of a conventional speaker-independent speech recognition system, and FIG. 4 is a configuration of a second phoneme recognition process of the present invention. FIG. 5 is a configuration diagram of a standard pattern generation rule unit. FIG. 6 (a) is a diagram showing a state of phoneme decomposition of a candidate word. ) Shows the address of the standard pattern of each phoneme of the candidate word, FIG. 6 (c) shows an example of the address according to the type of the phoneme standard pattern, and FIG. 6 (d) shows the combination of the generated standard patterns. FIG. 6E is a diagram showing an example of combinations of connectable standard patterns, and FIG. 7 is a diagram showing an interpolation method. In the figure, 1 is a voice input device, 2 is a voice analysis unit, 3 is a phoneme boundary detection unit, 4 is a phoneme identification unit, 5 is a phoneme standard pattern storage unit, 6 is a word identification unit, 7 is a word dictionary, and 8 is a correction. Rules department,
9 is a recognition result display unit, 100 is a speech input unit, 101 is a speech analysis unit, 102 is a distance calculation unit using continuous Mahalanobis DP, 103 is a word standard pattern storage unit, 104 is a candidate word discrimination unit, 105 is a parameter storage unit, 106 is a phoneme standard pattern storage unit, 107 is a distance calculation unit based on continuous Mahalanobis DP, 108 is a recognition unit based on recognition results in phoneme units, 109 is a result output unit, 110 is a standard pattern generation rule unit, 111 is a pattern matching unit, Reference numeral 150 denotes a phoneme division processing unit, 151 denotes a phoneme label addition unit, 152 denotes a 1-molar utterance time length detection unit, and 153 denotes a phoneme standard pattern connection unit.

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-63-165900 (JP, A) JP-A-60-121499 (JP, A) JP-A-63-46496 (JP, A) 798 (JP, B2) (58) Field surveyed (Int. Cl. ⁶ , DB name) G10L 3/00-9/18 JICST (JOIS)

Claims (4)

(57) [Claims] 1. A word dictionary storing dictionary word data,
A speech recognition method using a phoneme dictionary that stores dictionary phoneme data, wherein speech data is input, and the dictionary-use word data stored in the word dictionary is input using the spotting method. By performing matching, a voice section of the input voice data and a candidate word corresponding to the voice section are selected, and phoneme data for a dictionary of phonemes that can constitute the selected candidate word is selected from the phoneme dictionary. And matching with the input voice data of the selected voice section, and determining a recognition result of the input voice data based on a result of matching between the dictionary phoneme data and the input voice data. Voice recognition method. 2. The speech recognition method according to claim 1, wherein matching between the input speech data and dictionary word data is performed by continuous Mahalanobis DP. 3. The speech recognition method according to claim 1, wherein the matching between the input speech data and the phoneme data for dictionary is performed by continuous Mahalanobis DP. 4. The voice recognition method according to claim 1, wherein said voice data is inputted from a microphone.