JPH04233599A - Method and device speech recognition - Google Patents

Abstract

PURPOSE:To perform detailed speech recognition and increase the recognition rate by deriving a voice section and candidate words by word-by-word spotting in a 1st stage of the voice recognition and comparing them standard patterns of phenomes prepared with characteristics of a voice in the next stage. CONSTITUTION:A voice signal is inputted from a microphone and its input waveform is converted into a digital value, which is transferred to a voice analysis part 101. The voice analysis part 101 analysis and compresses the voice to connect to a time series of feature vectors. a word distance calculation part 102 calculates the distance if each frame from the time series of feature vectors and a standard pattern by using continuous Makaranobis' DP and the candidate words are discriminated by a candidate word decision part 104 and stored in a storage part 105 together with parameters. Then the input voice information is matched with word information by a spotting method and the most proper word is identified (108) in each matched phenome sequence and outputted (109).

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition device, and more particularly to a speech recognition device that recognizes, with a high recognition rate, speech such as words uttered consecutively by any speaker.

0002

2. Description of the Related Art Several recognition methods for speaker-independent recognition have been devised, but the currently most common speaker-independent recognition system will be described.

[0003] Conventionally, a recognition system aiming at an unspecified large vocabulary has a configuration as shown in FIG. The voice input from the voice input section 1 is processed by the voice analysis section 2 to obtain characteristic parameters such as a filter bank output including the power term of the voice, LPC cepstrum, etc. -dimensional compression by L transformation etc.) is also performed. (Since the analysis is performed frame by frame, the feature parameters after compression are hereinafter referred to as feature vectors).

Next, the phoneme boundary detection section 3 performs processing for determining phoneme boundaries from the continuous speech. 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.

[0005] Normally, the phoneme boundary detection section 3 uses a discriminant function or the like, and the phoneme identification section 4 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 do like 7 and 8
Information such as wn is used to narrow down the words to the final word.

[0006]

[Problem to be Solved by the Invention] However, since the basic configuration of the conventional recognition device described above is a bottom-up type, if an error occurs at a certain point in the recognition process,
It is in a form that is likely to have a negative impact 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.

[0007] In particular, when constructing a recognition device for unspecified speakers, it is extremely difficult to set threshold values for judgment in each process. If you set a threshold so that at least 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 multiple candidate words. was. Also,
When a recognition device is used in a real environment, there is a considerable amount of non-stationary noise, etc., and the recognition rate is low even for a recognition device for a small number of words, making it difficult to use.

[0008]

[Means for solving the problem] In order to solve the above problem, when voice information is input and the voice information is recognized, word information used as a reference and phoneme information classified according to voice characteristics are stored. , determine the characteristics of the input speech information, match the input speech information and word information using the spotting method, derive a candidate word and a speech interval of the candidate word, and perform a process for the derived speech interval. A speech recognition method is provided, characterized in that phoneme information corresponding to the candidate word is retrieved from the storage means according to the determined characteristics of the speech and matched with the input speech.

In order to solve the above problems, we have provided an input means for inputting voice information, a storage means for storing word information used as a reference and phoneme information classified according to voice characteristics when recognizing the voice information, and the input means. a determining means for determining the characteristics of the derived speech information; a deriving means for matching input speech information and word information using a spotting method to derive a candidate word and a speech interval of the candidate word; The present invention provides a speech recognition device comprising a phoneme recognition means for reading out phoneme information corresponding to the candidate word from the storage means according to the determined characteristics of the speech and performing matching with the input speech.

[0010] In order to solve the above problem, preferably, the characteristics of the voice differ depending on the speaker who utters the voice.

[0011]

[Example] (Example 1) Fig. 1 is a basic configuration diagram of a speech recognition system according to the present invention, where 100 is a speech input section, 101 is an input section that analyzes and compresses input speech, and converts it into a time series of feature vectors. A speech analysis unit 103 stores standard patterns obtained from word data uttered by a large number of speakers in correspondence with phoneme notation, a word standard pattern storage unit 102 stores feature vector series and word standard patterns of the speech analysis unit 101. Storage part 10
word distance calculation unit using continuous Mahalanobis DP, which calculates the distance of each standard pattern stored in 3 using continuous Mahalanobis DP for each frame of input data;
Reference numeral 4 denotes a candidate word discriminator that discriminates candidate words from word standard patterns based on the value of the distance from each frame word standard pattern obtained by continuous Mahalanobis DP, and 105 a candidate word discriminator for one or more candidate word sections. A parameter storage unit that stores parameters of feature vectors; 106 is a phoneme standard pattern storage unit that stores standard patterns created for each phoneme from voices uttered by multiple speakers; 107 is a feature vector of candidate words; a phoneme distance calculation unit using continuous Mahalanobis DP that calculates the distance between input data and a phoneme standard pattern using continuous Mahalanobis DP for the series in units of phonemes;
Reference numeral 108 denotes an identification unit based on phoneme unit recognition results that identifies and outputs the most appropriate word from each phoneme sequence matched for each of one or more candidate words. 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 phoneme unit recognition unit within the candidate words.

[0012] 110 selects the optimal phoneme pattern for the speaker who is currently inputting speech from the speaker category in which a plurality of phoneme standard patterns are classified according to the characteristics of each speaker so as to correspond to a plurality of standard patterns by a plurality of speakers. A speaker category identification pattern storage section in which patterns for identifying speaker categories are stored.

111 selects a standard pattern to be compared with the input speech by the optimal speaker phoneme standard pattern storage section 102, which will be described later, and in the phoneme recognition in the second part shown in FIG. a processing selection unit that instructs to select an optimal phoneme group from among the selected phonemes and store it in the optimal speaker phoneme standard pattern storage unit 112;

Reference numeral 112 denotes an optimum speaker phoneme standard pattern storage unit that stores the phoneme standard pattern of the optimum speaker category according to instructions from the processing selection unit 111.

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 surrounding noise signals during the time when audio input is accepted, and transfers the audio input waveform to the audio analysis unit 101 as a waveform converted into a digital value. Voice analysis section 101
Then, the waveform that is always input is 10msec to 30m.
Analysis is performed with a window width of about 2 msec to 10 ms.
Types of feature parameters to obtain feature parameters for each frame with a length of ec include LPC cepstrum and LPC mel cepstrum, which can be analyzed relatively quickly, and FFT cepstrum and FFT mel cepstrum if you want to extract parameters with high precision. is common, and there are also filter bank output values.

[0016] Also, by using normalized power information or multiplying each dimension of the parameter by a weighting coefficient,
Each frame is analyzed with the parameters most appropriate to the system usage. Next, compression is performed on the dimensions of the analyzed feature parameters. The cepstrum parameters are usually extracted from the first to twelfth order terms of the coefficients by the required number of dimensions (for example, six dimensions), and are used as feature vectors.

[0017]Furthermore, spectral difference information, power information, etc. may be parameterized and used as a feature vector in accordance with the parameters obtained from the spectral information.

When the filter bank output is used as a feature parameter, the dimension is reduced by orthogonal transformation such as KL transformation or Fourier transformation, and lower-order terms are used. These compressed parameters for one frame are used as feature vectors, and the time series of the feature vectors after dimension compression is used as the feature vector series (
Alternatively, we will simply call it a parameter.

In this example, the analysis window length is 25.6 msec.
After calculating the mel cepstral coefficients from the envelope spectrum that passes near the peak of the FFT spectrum with a frame period of 10 msec, the first to eighth orders of the coefficients are used.

Furthermore, first-order regression coefficients are obtained as difference information between adjacent mel-cepstrum, and a total of 16 features are calculated using the first to eighth orders of the regression coefficients in the same way as the coefficients of the mel-cepstrum obtained earlier. Let it be a feature vector for a frame. Here, the zero-order term of the mel cepstrum represents power. (This example shows the case where power information is not used)
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, ni, lei, nana, yon, go, maru, shi, roku, ku, hachi, shichi, kyu, ichi" and "hai, ie" are used, including vocal variations. 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 5000 people are used to create a standard pattern for one word. (The larger the number of audio samples, the better.) Here, 17
We will discuss an example in which a standard pattern of 17 words is created and stored for the purpose of recognizing only words, but this is not limited to 17 words; if you create any number of patterns in the same way, you can recognize any speech. Be able to recognize.

Furthermore, as a word standard pattern, the average of each phoneme stored in the phoneme standard pattern is combined according to a predetermined rule to create a standard pattern of words such as phrases. It is also possible to do so. Further, there may be a plurality of these standard patterns for each speaker.

FIG. 2 shows a flowchart showing the standard pattern creation procedure.

First, a core pattern is selected as a temporary comparison target when creating a standard pattern from a voice sample (S200). The selection method uses words with the most average utterance duration and utterance pattern among the 5000 words. Next, a sample word is input (S201), time-axis expansion/contraction matching is performed between the input word and the core pattern, and the average vector and variance co-value are calculated for each frame along the matching path that minimizes the time-normalized distance. A dispersion matrix is created (S202). Here, DP matching is used as a time axis expansion/contraction matching method. Next, the speaker numbers of the input words are changed one after another (S204).
(i = 1 to 5000), calculate the average value of the feature vector and the variance-covariance matrix for each frame (S2
03, S205). In this way, standard word patterns are created for each of the 17 words in the same manner as described above and stored in the standard word pattern storage section 103.

Reference numeral 110 denotes a speaker category identification pattern storage section.

[0025] This recognition device recognizes words, sentences, etc. uttered by an unspecified speaker, but before actually recognizing the target speech, it determines which category the speaker who is currently inputting falls into. A recognition device with high recognition accuracy can be realized by learning in advance whether or not the speaker enters the speaker's voice, and performing recognition using the phoneme standard pattern most suitable for the speaker from among a plurality of phoneme standard pattern groups in the second part. A method for creating a speaker category identification pattern will be described below according to the flowchart shown in FIG. First, the feature vector series obtained by analyzing the voices slowly uttered by 5,000 speakers using the word "aiueo" is classified into multiple arbitrary categories. Here, we will divide it into n classes. Class division exists as a clustering method. Any of a wide variety of methods may be used. In FIG. 3, first, in steps S401 to S405, the most average speaker among all 5000 speakers is selected, and the speaker who uttered the voice with the feature vector with the largest DP distance from this speaker's feature vector is selected. This is then set as I2 (S406).

[0026] Next, the procedure of selecting speaker I3 with the largest DP distance (normalized value) between speakers I1 and I2 is repeated, and the DP distance value is set to a predetermined standard such as 0.05. Repeat until the value is below the value. In this example,
Nine speakers from I1 to I9 were selected as representative samples for the category. A conceptual diagram of this speaker category is shown in Figure 4.
FIG. 5 shows an example of symbol representation of the feature vector.

Next, a standard pattern for the continuously uttered word "aiueo" is created according to the flowchart shown in FIG. 2, using the feature vector series of the case speakers of these categories (hereinafter referred to as case speakers) as core patterns. D in 202
While performing P matching, the variance and covariance vector of the corresponding frame are determined according to the DP path, but the DP window restrictions, DP slope restrictions, etc. are slightly tightened and used as a standard pattern.

By restricting speakers, good standard patterns with relatively little variance can be generated for each category centered on case speakers.

A collection of speakers used to create standard patterns corresponding to speaker categories using case speakers as core patterns will hereinafter be referred to as a category speaker group.

The word distance calculation unit 102 using the continuous Mahalanobis DP stores the time series of feature vectors input one after another using the continuous Mahalanobis DP into the word standard pattern storage unit 103 or the speaker category identification pattern storage unit 1.
The distance is calculated by performing matching with all standard patterns of words or phoneme chains stored in 10 using continuous Mahalanobis DP.

Here, the processing selection unit 111 selects the target of matching with the input voice from the speaker category identification pattern storage unit 110 in order to identify which speaker category the currently inputting speaker belongs to. , word standard pattern storage section 103
Choose one.

FIG. 6 shows an internal configuration diagram for explaining the operation of the processing selection section 111.

Further, a flowchart showing the processing operation of the processing selection section 111 is shown in FIG.

[0034] At the start of the speech recognition process (S301),
Since the mode is speaker identification mode, the process advances to S304. However, in case the input speaker changes midway through, or if the speaker wants to switch to speaker identification mode again, the speaker himself/herself can set a mode flag. Therefore, the mode switching unit 12
1 mode flag is read (S302). If the mode flag is word recognition mode, the mode switching unit switches to word recognition mode (S303), and as described above, the input voice is regarded as the target word and word recognition is performed (S310).
). If it is determined that the mode is the speaker identification mode (S303), an instruction such as "Please utter 'Ai ewoo'" is given to the speaker using an instruction means such as a display or voice synthesis (S304). Search for the optimal speaker category (S305)
, here, the value of the distance is 0.1 or less,
A restriction is set (S306). If rejected in S306, it is determined that the speaker's utterance length, intensity, etc. are extremely different from standard values, and retry information is added (S307).
, prompts for input again (S304). At this time, the input voice instruction section changes the content to ``Please utter slowly and continuously like ``Aiueo.'' Now, go ahead.'' and gives an instruction to the speaker. In this way, after identifying a category from among speaker categories I1 to I9, a phoneme standard pattern created as a core pattern with the same speaker as the case speaker to be returned to that category is stored in the phoneme standard pattern storage unit 106 as the optimal phoneme standard pattern. It is transferred (stored) to the speaker phoneme standard pattern storage unit 112 (S308).

Once the speaker category is specified, the mode flag is set to word recognition mode (S309), and word recognition processing is started (S310).

Next, 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, and it is an innovative method that recognizes the target speech section at the same time as cutting it out. In this embodiment, unspecificity is absorbed by using the Mahalanobis distance for the distance within each frame of the continuous DP method.

FIG. 8 shows the continuous Mahalanobis DP matching between the standard pattern of the word "zero" and the time series analysis of feature vectors of the input speech when the word "zero" is uttered, including silent sections. The results are shown below. In the figure, areas where black is darker are areas where the distance between the standard pattern and the input pattern is large, and areas where black is lighter and closer to white are areas where the distance between the standard pattern and the input pattern is small. Below the matching results, changes in cumulative distance over time are shown. This cumulative distance indicates the distance of the DP path that ends at that point in time, and the DP path is determined and its value is stored in the memory. The DP path stored in this memory is used to find the start of the voice section. 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 will be smaller than an arbitrarily determined threshold, and the word of that standard pattern will be selected as a candidate. Recognize it as a word. Then, in order to cut out a voice section from the input pattern, the starting point of the voice section is found by recalling the DP path from the memory and backtracking from the point in time when the cumulative distance is smaller than the threshold value and is minimum. The time series of feature vectors of the voice section thus obtained is stored in the parameter storage unit 105.

The processing system described above first obtains a candidate word, a series of feature vectors obtained by analyzing its speech interval, and cumulative distance results by continuous Mahalanobis DP. Here, when a plurality of candidate words such as "shichi" and "shi" with overlapping vocal sections are selected, "shichi" is selected and "shi" is discarded. Similarly, for “roku” and “ku”, when most of the vocal interval of “ku” is included in “roku” (here, 80% or more), “ku” is truncated and “roku” is used. Verification is performed only on

In this embodiment, the phoneme standard pattern storage section 106
vowels (a, i, u, e, o) and consonants (z, s, n, r
, g, m, shi, k, h, ci).

Note that in this embodiment, since the purpose is to recognize the 17 words mentioned above, the phoneme standard pattern storage unit 106
The phonemes stored in the 15 types are the above, but as mentioned earlier, when expanding the recognition target and increasing the number of standard patterns, all the phonemes that make up the standard pattern are stored in the standard pattern in the same way. is created and stored in the phoneme standard pattern storage section 106.

[0041] Here, we classify each phoneme into the category speakers used to create the category-based standard pattern, cut out each phoneme from the words uttered by each speaker, and perform clustering on these same phoneme sets. etc., to create and store a plurality of phoneme standard patterns belonging to each class.

FIG. 9 shows this situation. Extract the phoneme part from the group of category speakers belonging to the speaker category (
For example, the phoneme /a/) is further subjected to processing such as clustering to create one or more standard patterns for the phoneme /a/. In the figure, when the speaker category is 1, /a/
are /a1/ and /a2/, /u/ are /u1/, /u2/
, /u3/, phoneme standard pattern sequences corresponding to a plurality of phoneme classes are stored. For example, /a1/ is a voiced “a”, and /a2/ is a devoiced “a”.
Even the same phoneme can undergo significant variations due to differences in the surrounding phonemes (phonetic environment) due to differences in the position of the phoneme in the word, and differences in the way the phoneme is uttered even by the same speaker.

[0043] As in this method, phonemes are extracted from words classified by speaker category, and by further creating multiple phoneme standard patterns through clustering, etc.
More accurate recognition results can be obtained.

In addition, the optimal speaker phoneme standard pattern storage unit 11
2, a group of phoneme standard patterns corresponding to the optimal speaker category selected from the speaker category identification pattern storage section 110 is transferred from the phoneme standard pattern storage section 106 by the processing selection section 111 and stored. .

The phoneme distance calculation section 107 using continuous Mahalanobis DP performs matching with each phoneme for the speech section cut out as a candidate word stored in the parameter storage section 105.

Similar to the word distance calculation unit 102 using continuous Mahalanobis DP, the phoneme interval is calculated from the position where the cumulative distance is the minimum. (Similar to the candidate word discrimination unit 104, the point in time when the cumulative distance is the minimum is the end of the phoneme,
In this embodiment, for example, if "zero" → "zero" is a candidate word, "z", "e", "r", "
Matching is performed only for the four types of phonemes "o".4
As a result of matching between the phoneme of the species and the voice section that became a candidate after being identified as "zero", the positional relationship of the point where the cumulative distance of each phoneme is the minimum and the average value of the minimum distance are calculated. 10.

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, “z
”, the minimum value is “j” and the frame position is “zf”. The recognition unit 108 based on the recognition result in units of phonemes performs final word identification based on the data sent from the phoneme distance calculation unit 107 using continuous Mahalanobis DP. First, the order (frame position) of the phoneme string of the candidate word is z
Check whether f<ef<rf<of. If this order is used, the recognized word will be “zero” and the average recognition distance will be “zero”.

[0048]

If the value of X is smaller than the threshold value H, "zero" is output as the recognition result.

FIG. 11 shows the output results of word candidates (output results of the candidate word discriminator 104). The word "Hachi" is output as a candidate for ■, the word "shichi" is output as a candidate for ■, and the word "shi" is output as a candidate for ■. However, as mentioned earlier, ■ is included in the interval of ■ by more than 80%, and the same “shi” is included in the interval of ■.
Since it exists in , identification at the phoneme level will be performed for ■■.

Case ■ Phoneme sequence of word S1 “/h/a
/c/i/” and the phoneme sequence of word S2 “/sh/i/c/i
As a result of matching for ``/'', if the order of both phonemes is the same as that of the candidate word, and if the distance of each phoneme is smaller than H (threshold) -> the smaller of the average cumulative distances X is output.

Case ■ When the distance between individual phonemes that are both in different orders is smaller than the threshold (H) -> DP matching is performed using the character string of the word and the phoneme string, and the distance is determined by the threshold (I).

[0052] Case ■ If the order is correct or the threshold of each phoneme does not clear (H) → Reject case ■ If the order is wrong and the threshold of the phoneme also does not clear → Reject Recognition result of phoneme unit The word identification method 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.

For example, the speech (word) outputted as the final result by the recognition unit 108 based on the recognition result of each phoneme is outputted by the result output unit 109. When recognizing only voice information from a telephone or the like, the recognition result is confirmed using a voice synthesis means, for example, saying, "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.

[0054] In this embodiment, a continuous Mahalanobis matching method using Mahalanobis distance as a distance measure that statistically absorbs unspecificity was used as a pattern matching method, but the second method is not limited to this.
It goes without saying that for departmental recognition, any matching method that uses a distance that uses probability and absorbs unspecificity, such as the Marcol model, may be used.

[0055] In this embodiment, the word "aiueo" uttered consecutively was used as the phonetic chain for identifying the speaker group, but the word for identifying the speaker group is not limited to this. Moreover, this may be plural. For example, word A
(Words that include vowels) are categorized based on the basic characteristics of speakers (length of formant peak, etc.). Words B and words that have the strongest characteristics for classifying speakers, such as whether they are easy to include or difficult to include, 2. length of consonants such as “p, t, k,” 3. average speaking speed, etc.) Starting from C, it is better to further classify the speaker groups.

(Embodiment 2) In the above embodiment 1, the phonemes stored in the phoneme standard pattern storage section were limited to the phonemes included in the recognition target words necessary for recognition by this recognition device. As for phoneme standard patterns that are always stored, phoneme standard patterns may be created for all Japanese phonemes for each speaker category and phoneme class (when recognizing Japanese). This increases the memory of 106, but when the recognition target word is changed, a standard pattern corresponding to the speaker category can be stored in 112 for phonemes (plurality) used in the target word.

Furthermore, as phonemes, we store not only the phonemes necessary for Japanese pronunciation, but also all the phonemes used in various languages (English, French, German, Chinese, etc.), and from among these, we can select the target word for recognition. You may choose.

[0058] As shown in Fig. 12, increase the phoneme types in Fig. 2.4,
Furthermore, it would be a good idea to prepare this for each country.

[0059]

As explained above, according to the present invention, in the first step of speech recognition, word-by-word spotting is performed to derive speech sections and candidate words, and in the second step, a plurality of candidate words are prepared depending on the characteristics of the speech. By comparing the phoneme pattern with the standard pattern of phonemes, more detailed speech recognition is performed in the second stage, resulting in an effect of increasing the recognition rate.

[Brief explanation of the drawing]

FIG. 1 is a basic block diagram of this embodiment.

[Figure 2] Standard pattern creation flow.

[Fig. 3] Speaker category creation flow.

FIG. 4 is a conceptual diagram of speaker categories.

FIG. 5 is a diagram illustrating a representation of a feature vector.

FIG. 6 is an internal configuration diagram of a processing selection section.

FIG. 7 is a flowchart showing the overall flow.

FIG. 8 is an exemplary diagram of matching using Mahalanobis distance.

FIG. 9 is a data format diagram of speaker categories.

FIG. 10 is an exemplary diagram of phoneme matching.

FIG. 11 is an exemplary diagram of a plurality of candidate words and input signals.

FIG. 12 is a data format diagram when having speaker categories for multiple languages.

FIG. 13 is a configuration diagram of a conventional speech recognition system.

Claims (4)

[Claims] 1. Storing word information used as a reference and phoneme information classified according to voice characteristics when inputting voice information and recognizing the voice information, and determining the characteristics of the input voice information; The input speech information and word information are matched using a spotting method, a candidate word and a speech interval of the candidate word are derived, and the phoneme information corresponding to the candidate word is determined for the derived speech interval. A voice recognition method, characterized in that the voice is retrieved from the storage means according to the characteristics of the input voice, and matching with the input voice is performed. 2. Claim 1, wherein the characteristics of the voice differ depending on the speaker who utters the voice.
The speech recognition method described in . 3. Input means for inputting voice information, storage means for storing word information used as a reference when recognizing the voice information and phoneme information classified according to voice characteristics, characteristics of the input voice information. a derivation means for matching input speech information and word information using a spotting method and deriving a candidate word and a speech interval of the candidate word; A speech recognition device comprising phoneme recognition means for retrieving corresponding phoneme information from the storage means according to the characteristics of the determined speech and performing matching with the input speech. 4. Claim 3, wherein the characteristics of the voice differ depending on the speaker who utters the voice.
The speech recognition device described in .