JP2003131693A - Device and method for voice recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To recognize a word which is not shown in dictionary data. SOLUTION: Provided are a 1st decoder 7 and a 2nd decoder 8. The 1st decoder 7 decodes a feature quantity supplied from a voice analysis part 2 into a phoneme sequence by using a sound model and a 1st language model, compares the phoneme sequence with words and documents shown in the dictionary data, and outputs a word and/or document having a similar phoneme sequence as a recognition result of an inputted voice signal. The 2nd decoder 8, on the other hand, decodes the feature quantity supplied from the voice analysis part 2 into a phoneme sequence by using the sound model and a 2nd language model. Then a KANA(Japanese syllabary) conversion part 9 converts the phoneme sequence into KANA and outputs the conversion result as a recognition result of the inputted voice signal.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device and a voice recognition method applied when, for example, inputting voice and obtaining desired data from a database.

[0002]

2. Description of the Related Art In recent years, a system has been proposed in which information is input and a database is searched based on the input information to obtain desired information. As an example, there is a system in which information such as artist name and song name is input, and a database is searched based on the input information to obtain a desired song.

When inputting information such as artist name and song name in the system, voice input is performed in addition to manual operation input. Today, a voice recognition device is being developed in order to perform voice input more easily.

As shown in FIG. 8, a voice recognition device 100.
Is a voice analysis unit 101, an acoustic model storage unit 102, a language model storage unit 103, a dictionary 104, and a decoder 10.
5 and 5.

The voice analysis unit 101 performs frequency analysis on the input voice signal, for example, by extracting feature quantities necessary for recognition at predetermined sampling intervals.

The acoustic model storage unit 102 stores acoustic models. The acoustic model is obtained by modeling the characteristics of a speech signal by, for example, HMM (Hidden Markov Model) for each phoneme.

The language model storage unit 103 stores a language model. The language model is information indicating a phoneme string such as a word or a sentence to be recognized.

The dictionary 104 is a part of the language model 103 and stores dictionary data showing the correspondence between the phoneme sequence to be recognized and the text data. Information that connects acoustic models is modeled from this dictionary.

The decoder 105 uses the acoustic model and the language model to generate phoneme strings of words, sentences, etc. for the feature quantities supplied from the speech analysis unit 101, and displays the generated phoneme strings in dictionary data. When a word or sentence having a similar phoneme sequence is shown in the dictionary data, the word or sentence is output as a recognition result of the input voice signal. .

In the voice recognition apparatus 100 described above, first, the voice analysis unit 101 performs frequency analysis of the waveform of the input voice signal and extracts the feature amount. The extracted feature amount is supplied to the decoder 105.

Next, the decoder 105 models the feature amount supplied from the speech analysis unit 101 for each phoneme using the acoustic model.

Next, the decoder 105 generates a phoneme string by connecting the modeled phonemes using a language model, and the generated phoneme string is a phoneme string of words or sentences indicated in the dictionary data. When a word or sentence having a similar phoneme sequence is shown in the dictionary data, the word or sentence is output as a recognition result of the input voice signal.

[0013]

By the way, in the speech recognition apparatus 100, a phoneme sequence generated by converting an input voice signal is compared with a phoneme sequence of a word or a sentence shown in dictionary data, and the result is similar. When a word or a sentence having the selected phoneme sequence is shown in the dictionary data, the word or the sentence is used as the recognition result.

Therefore, it becomes difficult for the voice recognition apparatus 100 to recognize a word or a sentence not shown in the dictionary data as a recognition result. Further, when the user uses the voice recognition device 100 without knowing the words or sentences shown in the dictionary data, a word that is not a recognition result is input, and it is difficult to fully use the voice recognition device 100. Become.

Further, in the voice recognition apparatus 100, when a voice signal is input, a word or a sentence whose length is similar to the pronunciation is used as the recognition result. Therefore, even when words or sentences that partially match the input voice signal and have different lengths are shown in the dictionary data, the partially matching words or sentences are not regarded as the recognition result. In addition, a word or a sentence having a similar length often becomes a recognition result. That is, when a word or a sentence input by the user is partially stored, the speech recognition apparatus 100 causes the dictionary 104 to store the word or the sentence of the correct answer, but deviates from the correct answer. And sentences are often used as recognition results. For example, if the dictionary data shows "Mt. Fuji" and "Uji" but not "Fuji",
When the user inputs "Fuji" with the intention of "Mount Fuji",
"Uji" is recognized as the recognition result.

Therefore, the voice recognition apparatus 100 becomes insufficient in accuracy. If the accuracy of the voice recognition device 100 is insufficient, it becomes difficult to obtain desired information when searching the database based on the input information, for example.

The present invention has been proposed in view of the conventional circumstances as described above, and partially matches the words and sentences not shown in the dictionary data and the words and sentences shown in the dictionary data. An object of the present invention is to provide a voice recognition device and a voice recognition method capable of recognizing an input word or sentence.

[0018]

A speech recognition apparatus according to the present invention comprises a speech analysis means for extracting a feature amount of an input voice signal and an acoustic storing a sound model in which the feature amount of the voice signal is modeled. Model storage means, language model storage means for storing a language model consisting of information indicating the probability of phoneme appearance, and dictionary data storage for storing dictionary data indicating phoneme strings of a plurality of words and / or sentences Means, and the language model and the acoustic model are applied to the feature amount supplied from the speech analysis means to generate a phoneme string, and the generated phoneme string is a word and / or a word indicated in the dictionary data. When a word and / or a sentence having a similar phoneme sequence is shown in the dictionary data by comparing with a phoneme sequence of a sentence, the word and / or the sentence is recognized as a recognition result of an input voice signal. A first decoding means for and outputs,
Second decoding means for applying the language model and the acoustic model to the feature quantity supplied from the speech analysis means to generate a phoneme string, and a phoneme string generated by the second decoding means. And a kana conversion means for outputting the result of conversion to kana as a recognition result of the input voice signal.

Further, the voice recognition apparatus according to the present invention detects a partial match between the word and / or sentence shown in the dictionary data and the word and / or sentence output from the kana conversion means. However, when a partially matched word and / or sentence is shown in the dictionary data, a partial match detection means for outputting the word and / or sentence as a recognition result of the input voice signal is provided. Characterize.

Further, the voice recognition device according to the present invention evaluates the reliability of the word and / or the sentence recognized by the first decoding means, and according to the reliability, recognizes the word and / or the sentence. , And a reliability detecting means for outputting as a recognition result of the input voice signal.

Further, the speech recognition method according to the present invention is such that an acoustic model modeling a feature amount of a speech signal, a language model consisting of information indicating the probability that a phoneme appears, and a phoneme of a plurality of words and / or sentences. The dictionary data indicating the sequence is stored in the storage unit, the feature amount of the input speech signal is extracted, the phoneme sequence is generated by applying the language model and the acoustic model to the feature amount, and is generated. The phoneme string is compared with the phoneme strings of the words and / or sentences shown in the dictionary data, and when words and / or sentences having similar phoneme strings are shown in the dictionary data, the word and / or sentence Or, while performing the first recognition process with the sentence as the recognition result,
After the phoneme sequence is generated by applying the language model and the acoustic model to the feature quantity, the phoneme sequence is converted into kana, and the kana-converted word and / or sentence is converted into the input speech signal. It is characterized in that a second recognition process is performed as a recognition result.

Further, the voice recognition method according to the present invention makes it possible to partially match the words and / or sentences shown in the dictionary data with the words and / or sentences obtained by the second recognition processing. When a word and / or a sentence that is detected and partially matched is shown in the dictionary data, the word and / or the sentence is used as a recognition result of the input voice signal.

Further, the speech recognition method according to the present invention evaluates the reliability of the word and / or the sentence generated by the first recognition processing, and determines the word and / or the sentence according to the reliability. , The recognition result of the input voice signal is used.

[0024]

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

First Embodiment Hereinafter, a first embodiment of the present invention will be described in detail with reference to the drawings.

The present invention can be applied to, for example, the voice recognition device 1 having the configuration shown in FIG.

The voice recognition device 1 includes a voice analysis unit 2 and a first
Language model storage unit 3, acoustic model storage unit 4, and second
Language model storage unit 5, a first decoder 7, a second decoder 8, and a Kana conversion unit 9.

The voice analysis unit 2 performs frequency analysis on the input voice signal, for example, by extracting feature quantities necessary for recognition at predetermined sampling intervals.
Specifically, the voice analysis unit 2 extracts the energy of the signal, the number of zero tolerances, the frequency characteristic, and the amount of change thereof.
As frequency analysis, linear prediction analysis (LPC), fast Fourier transform (FFT), band pass filter (BPF)
Are used. The voice analysis unit 2 extracts these feature quantities as a vector or quantizes and extracts them as a scalar.

The first language model storage unit 3 stores the first language model and dictionary data. Dictionary data is
It is data indicating a correspondence between a phoneme string to be recognized by the voice recognition device 1 and text data. Further, the first language model models the probability that a phoneme appears based on dictionary data.

The acoustic model storage unit 4 stores an acoustic model. The acoustic model is a HMM (Hidden Markov Model) for each phoneme of the feature amount of the input speech signal.
It is modeled by.

In this embodiment, PLU (pseudo phoneme unit) is adopted as a unit for modeling a voice. As shown in Table 1, PLU has 27 Japanese
The number of states is assigned to each phoneme. The number of states refers to the shortest number of continuous frames in subword units. For example, since the number of states of the phoneme "a" is "3", the phoneme "a" is at least three.
Means that the frame continues. The three states are pseudo representations of the rise of sound, the steady state, and the release state. It should be noted that plosive sounds such as phonemes "b" and "g" are set to two states because their original phonemes are short. Also, breathing (ASPIRATION) is set to 2 states. Furthermore, silence (SILENCE) is set to one state because there is no temporal change.

[0032]

[Table 1]

The second language model storage unit 5 stores the second language model. The second language model is a model in which each phoneme appears evenly or unevenly. In addition,
Information other than the second language model may be stored in the second language model storage unit 5. For example, information for determining whether or not the input voice signal is in Japanese may be stored.

The first decoder 7 uses the acoustic model and the first language model for the feature quantity supplied from the speech analysis unit 2 to generate a phoneme sequence of words and sentences, and the generated phoneme sequence. Is compared with the phoneme sequence of the word or sentence shown in the dictionary data, and when the word or sentence having a similar phoneme sequence is shown in the dictionary data, the word or sentence is Output as a recognition result.

When a plurality of words or sentences having similar phoneme strings are shown in the dictionary data, the first decoder 7 evaluates the reliability of each word or sentence and scores the score. The highest word or sentence,
Alternatively, a plurality of words or sentences whose scores are equal to or higher than a predetermined value may be output as the recognition result of the input voice signal.

As a method of evaluating the reliability of a word or a sentence, for example, it is possible to detect how much the length of each phoneme forming the input speech differs from a predetermined length. There is a way to evaluate. Specifically, when "kai" is entered, k, a,
There is a predetermined length for each phoneme of i, and the closer the length of each phoneme is to the predetermined length, the higher the reliability.

The second decoder 8 uses the acoustic model and the second language model for the feature amount supplied from the speech analysis unit 2 to generate a phoneme string such as a word or a sentence. Further, the second decoder 8 supplies the generated phoneme string to the kana conversion unit 9.

The kana conversion section 9 converts the phoneme string supplied from the second decoder 8 into kana. Also, the kana conversion unit 9
Outputs the result of kana conversion as the recognition result of the input voice signal.

The signal flow of the speech recognition apparatus 1 described above is as described below.

First, a voice is input to the voice recognition device 1. The input voice is supplied to the voice analysis unit 2.

Next, the voice analysis unit 2 performs frequency analysis on the input voice signal, for example, by extracting a feature amount necessary for recognition at a predetermined sampling interval, and then extracts the extracted feature amount. It is supplied to the first decoder 7 and the second decoder 8.

The first decoder 7 models the feature amount supplied from the speech analysis unit 2 for each phoneme using the acoustic model.

Next, the first decoder 7 generates a phoneme string by applying the first language model to the modeled phoneme, and the generated phoneme string is shown in the dictionary data. When a word or a sentence having a similar phoneme sequence is shown in the dictionary data by comparison with the phoneme sequence of the word or the sentence, the word or the sentence is output as the recognition result of the input voice signal.

On the other hand, the second decoder 8 models the feature amount supplied from the speech analysis unit 2 for each phoneme using the acoustic model.

Next, the second decoder 8 applies a second language model to the modeled phoneme to generate a phoneme string. The second decoder 8 supplies the generated phoneme string to the kana conversion unit 9.

Next, the kana conversion section 9 includes the second decoder 8
The phoneme string supplied from is converted into kana. Kana converter 9
Outputs the result of kana conversion as the recognition result of the input voice signal.

The hardware configuration of the voice recognition device 1 is as shown in FIG. 2, and the microphone 10, the analog-digital conversion unit 11, the central processing unit (Central Processing Unit).
Processing Unit; hereinafter referred to as CPU. ) 12, and
DRAM (Dinamic Random Access Memory) 13, flash ROM (Read Only Memory) 14, mask R
It consists of OM15.

The microphone 10 converts the input audio signal into an electric signal and supplies the electric signal to the analog-digital conversion unit 11. The analog-digital conversion unit 11 converts the electric signal supplied from the microphone 10 from an analog signal into a digital signal. The CPU 12 performs processing in the voice analysis unit 2, the first decoder 7, the second decoder 8, and the kana conversion unit 9.

The DRAM 13 temporarily stores the data calculated by the CPU 12. The flash ROM 14 is
Prepared for debugging. The mask ROM 15 stores an acoustic model, a first language model, and a second language model.

The voice recognition device 1 described above includes the first decoder 7, the second decoder 8 and the kana conversion section 9. In the voice recognition device 1, the first decoder 7 is
The feature quantity supplied from the speech analysis unit 2 is modeled for each phoneme using an acoustic model, and then connected using the first language model to generate a phoneme string. Then, the first decoder 7 compares the generated phoneme string with the phoneme strings of words or sentences shown in the dictionary data, and when words or sentences having similar phoneme strings are shown in the dictionary data. , The word or sentence is output as a recognition result of the input voice signal. Also, the second decoder 8 generates a phoneme string by modeling the feature amount supplied from the speech analysis unit 2 for each phoneme using the acoustic model and then connecting the feature quantity using the second language model. . Then, the kana conversion unit 9 converts the generated phoneme sequence into kana and outputs the result of conversion into kana as the recognition result of the input voice signal.

Therefore, the voice recognition device 1 can be used even when a word or a sentence not shown in the dictionary data is input.
It is possible to output the recognition result of the input voice signal. For example, even if the user does not know a word or a sentence shown in the dictionary data, the voice recognition device 1 can be fully used. Further, the voice recognition device 1 is
The accuracy is high, and it is easy to obtain desired information by searching the database based on the input information, for example.

The voice recognition device 1 to which the present invention is applied.
May include a reliability detector 20, as shown in FIG.

The reliability detecting section 20 detects the reliability of the recognition result evaluated by the first decoder 7 and outputs the words or sentences having a high score as the recognition results in preference to the words or sentences having a low score. Alternatively, the recognition result output from the kana conversion unit 9 is preferentially output as the recognition result rather than the words and sentences having a predetermined score or less.

For example, the data output from the first decoder 7 and the kana conversion unit 9 are as shown in Table 2 below, and the reliability detection unit 20 outputs the data more than words or sentences whose score is less than 60. When the word or sentence output from the kana conversion unit 9 is set to be preferentially output, the operation of the voice recognition device 1 is as described below.

[0055]

[Table 2]

As shown in FIG. 4, first, in step S1, the reliability detecting section 20 outputs "uji". Next, the process proceeds to step S2, and the user inputs a voice signal “uj (uj
i) ”. When the input voice signal is "uji", the recognition of the voice signal ends.

When the input voice signal is not "uji", the process proceeds to step S3, and the reliability detecting unit 20
Outputs "fui". Next, in step S4, the user determines that the input voice signal is “fu (fu).
i) ”. When the input voice signal is "fui", the recognition of the voice signal ends.

When the input voice signal is not "fui", the process proceeds to step S5, and the reliability detecting section 20
Outputs "Fuji". Then step S6
The user can confirm that the input voice signal is “Fuji (fuj).
i) ”. When the input voice signal is "Fuji", the recognition of the voice signal ends.

The input voice signal is "Fuji".
If not, the process proceeds to step S7, and the reliability detecting unit 2
0 outputs "kuji". Then step S
8, the user confirms that the input voice signal is “kuji (ku).
ji) ”. When the input voice signal is "kuji", the recognition of the voice signal ends. If it is not "kuji", the process proceeds to step S9, and it is determined that the recognition results are all incorrect. Then, the recognition of the voice signal ends.

As described above, by providing the reliability detecting section 20, the voice recognition device 1 can efficiently recognize the input voice signal.

Second Embodiment Next, a second embodiment of the present invention will be described in detail with reference to the drawings.

The present invention can also be applied to the voice recognition device 30 having the structure shown in FIG. 5, for example. In the voice recognition device 30, the same elements as those of the voice recognition device 1 are designated by the same reference numerals, and detailed description thereof will be omitted.

The speech recognition device 30 includes a speech analysis unit 2, a first language model storage unit 3, an acoustic model storage unit 4, a second language model storage unit 5, a first decoder 7, and a first decoder 7. Two
Decoder 8, Kana conversion unit 9, and partial match detection unit 31
With.

The partial match detection unit 31 detects whether or not the word or sentence converted by the kana conversion unit 9 partially matches the word or sentence shown in the dictionary data. Then, when the dictionary data shows a word or a sentence that partially matches the word or the sentence converted by the kana conversion unit 9, the partial match detection unit 31 uses the word or the sentence as the recognition result. Output.

The signal flow of the voice recognition device 30 described above is as described below.

First, a voice is input to the voice recognition device 30. The input voice is supplied to the voice analysis unit 2.

Next, the voice analysis unit 2 performs frequency analysis on the input voice signal, for example, by extracting a feature amount necessary for recognition at a predetermined sampling interval, and then extracts the extracted feature amount as a first value. It is supplied to the first decoder 7 and the second decoder 8.

The first decoder 7 models the feature amount supplied from the speech analysis unit 2 for each phoneme using the acoustic model.

Next, the first decoder 7 generates a phoneme string by applying the first language model to the modeled phoneme, and the generated phoneme string is shown in the dictionary data. When a word or a sentence having a similar phoneme sequence is shown in the dictionary data by comparison with the phoneme sequence of the word or the sentence, the word or the sentence is output as the recognition result of the input voice signal.

On the other hand, the second decoder 8 models the feature amount supplied from the speech analysis unit 2 for each phoneme using the acoustic model.

Next, the second decoder 8 applies a second language model to the modeled phoneme to generate a phoneme string. The second decoder 8 supplies the generated phoneme string to the kana conversion unit 9.

Next, the kana conversion section 9 uses the second decoder 8
The phoneme string supplied from is converted into kana. Kana converter 9
Outputs the result of kana conversion as a recognition result of the input voice signal and supplies the result to the partial match detection unit 31.

Next, the partial match detection section 31 detects whether or not the word or sentence converted by the kana conversion section 9 partially matches the word or sentence shown in the dictionary data. When a partially matching word or sentence is shown in the dictionary data, the word or sentence is output as the recognition result of the input voice signal.

The speech recognition device 30 described above is provided with the partial match detection section 31, and when a word or a sentence partially matching the input voice signal is shown in the dictionary data, the word is detected. Or a sentence can be used as the recognition result.

Further, the voice recognition device 30 compares the input voice signal with a word or a sentence having different lengths and partially matching, even when the dictionary data shows the word or the sentence. A sentence can be used as the recognition result.

Therefore, for example, even when the user inputs a partially stored word or sentence to the voice recognition device 30, the word or sentence partially matching the input word or sentence becomes the dictionary data. If so, the word or sentence can be output as a recognition result.

The voice recognition device 1 to which the present invention is applied.
May include a reliability detector 32, as shown in FIG.

The reliability detecting section 32 detects the reliability of the recognition result evaluated by the first decoder 7, and outputs the words or sentences with a high score as the recognition result with priority over the words or sentences with a low score. Alternatively, the recognition result output from the kana conversion unit 9 or the recognition result output from the partial match detection unit 31 is preferentially output over a word or a sentence having a predetermined score or less.

For example, the data output from the first decoder 7, the kana conversion unit 9, and the partial match detection unit 31 are as shown in Table 3 below, and the reliability recognizing unit 32 is used.
However, the words and sentences output from the partial match detection unit 31 are output with priority over the words and sentences with a score of less than 70, and the words and sentences with a score of less than 60 are output from the kana conversion unit 9. The operation of the voice recognition device 30 is as described below when the selected word or sentence is set to be preferentially output.

[0080]

[Table 3]

As shown in FIG. 7, first, step S21.
In, the reliability recognizing unit 32 outputs "uji". Next, the process proceeds to step S22, in which the user inputs the voice signal "U
ji) ”is confirmed. When the input voice signal is "uji", the recognition of the voice signal ends.

When the input voice is not "uji", the process proceeds to step S23, and the reliability recognizing section 32 outputs "fujisan". Next, proceeding to step S24, the user confirms whether or not the inputted voice is "Fujisan". When the input voice is "Fujisan", the recognition of the voice signal ends.

The input voice is "Fujisan".
n) ”, the reliability recognizing unit 32 outputs“ fui ”(S25). Next, proceeding to step S26, the user confirms whether or not the input voice is "fui". When the input voice is "fui", the recognition of the voice signal ends.

When the input voice is not "fui", the process proceeds to step S27, and the reliability recognizing section 32 outputs "fuji". Next in step S28
Then, the user inputs the inputted voice to "Fuji (fuj
i) ”. When the input voice is "Fuji", the recognition of the voice signal ends.

When the input voice is not "Fuji", the process proceeds to step S29, and the reliability recognizing unit 32 is used.
Outputs "kuji". Then step S3
0, the user inputs the input voice as “kuj (kuj
i) ”. When the input voice is "kuji", the recognition of the voice signal is ended. If it is not "kuji", the process proceeds to step S31, and it is determined that the recognition results are all incorrect. Then, the recognition of the voice signal is ended.

As described above, by providing the reliability detecting section 32, the voice recognition device 30 can efficiently recognize the input voice signal.

[0087]

The speech recognition apparatus and the speech recognition method according to the present invention extract a feature amount of an input voice signal and apply a language model and an acoustic model to the feature amount to generate a phoneme sequence. , Compare the generated phoneme string with the phoneme string of the word or sentence shown in the dictionary data, and enter the word or sentence when the word or sentence having a similar phoneme string is shown in the dictionary data. A first recognition process is performed to obtain a recognition result of the generated voice signal. Further, a voice recognition device and a voice recognition method according to the present invention, a language model for the feature amount,
After the input voice signal is converted into a phoneme sequence by applying the acoustic model, the phoneme sequence is converted into kana and the second recognition processing is performed to obtain the recognition result of the input voice signal.

That is, the voice recognition apparatus and the voice recognition method according to the present invention can output the recognition result of the input voice signal even when a word or a sentence not shown in the dictionary data is input. Becomes Therefore, the voice recognition device and the voice recognition method according to the present invention can accurately recognize the input voice signal.

Further, the voice recognition device and the voice recognition method according to the present invention compare the word or sentence obtained by the second recognition processing with the word or sentence shown in the dictionary data,
When partially matching words or sentences are shown in the dictionary data, the words or sentences are regarded as the recognition result of the input voice signal.

Therefore, the voice recognition apparatus and the voice recognition method according to the present invention obtain the recognition result even when the dictionary data shows a word or a sentence partially matching the input voice signal. It becomes possible. Further, the voice recognition device and the voice recognition method according to the present invention, even when a word or a sentence which is different in length and partially coincides with each other is shown in the dictionary data as compared with the input voice signal. It is possible to use the word or sentence as the recognition result.

Further, the voice recognition device and the voice recognition method according to the present invention evaluate the reliability of the word or sentence generated by the first recognition processing, and recognize the recognition result of the voice signal input according to the reliability. And

Therefore, according to the voice recognition device and the voice recognition method of the present invention, the input voice signal can be efficiently recognized.

[Brief description of drawings]

FIG. 1 is a block diagram showing a voice recognition device to which the present invention is applied.

FIG. 2 is a diagram showing a hardware configuration of the voice recognition device.

FIG. 3 is a block diagram showing another voice recognition device to which the present invention is applied.

FIG. 4 is a flowchart for explaining an operation of a reliability detection unit in the voice recognition device.

FIG. 5 is a block diagram showing still another voice recognition device to which the present invention is applied.

FIG. 6 is a block diagram showing still another voice recognition device to which the present invention is applied.

FIG. 7 is a flowchart for explaining the operation of the reliability detecting unit in the voice recognition device.

FIG. 8 is a block diagram showing a conventional voice recognition device.

[Explanation of symbols]

1 speech recognition device, 2 speech analysis unit, 3 first language model storage unit, 4 acoustic model storage unit, 5 second language model storage unit, 7 first decoder, 8 second decoder, 9 Kana conversion unit

Claims (8)

[Claims] 1. A voice analysis unit for extracting a feature amount of an input voice signal, an acoustic model storage unit for storing an acoustic model modeled on the feature amount of a voice signal, and information indicating a probability that a phoneme appears. A language model storage means for storing a language model consisting of, a dictionary data storage means for storing dictionary data indicating a phoneme sequence of a plurality of words and / or sentences, and a feature amount supplied from the speech analysis means. For the above, the language model and the acoustic model are applied to generate a phoneme string, the generated phoneme string is compared with the phoneme strings of the words and / or sentences shown in the dictionary data, and a similar phoneme string is obtained. A first decoding means for outputting the word and / or the sentence as a recognition result of the input voice signal when the word and / or the sentence included therein is shown in the dictionary data; Second decoding means for generating a phoneme string by applying the language model and the acoustic model to the feature amount supplied from the analyzing means, and converting the phoneme string generated by the second decoding means into kana. ,
A voice recognition device, comprising: a kana conversion unit that outputs the result of Kana conversion as a recognition result of an input voice signal. 2. A word and / or a sentence output from the first decoding means and a word and / or a sentence output from the kana conversion means are output as a recognition result of an input voice signal. The voice recognition device according to claim 1, wherein 3. A partial match between a word and / or a sentence shown in the dictionary data and a word and / or a sentence output from the kana conversion means is detected, and a partially matched word and a sentence are detected. 2. A voice according to claim 1, further comprising: partial match detection means for outputting the word and / or the sentence as a recognition result of the input voice signal when the sentence is indicated in the dictionary data. Recognition device. 4. The reliability of a word and / or a sentence recognized by the first decoding means is evaluated, and the word and / or the sentence is recognized according to the reliability, as a result of recognition of an input voice signal. The voice recognition device according to claim 1, further comprising a reliability detection unit that outputs the above. 5. A storage unit that stores an acoustic model that models a feature amount of a speech signal, a language model that includes information that indicates the probability that a phoneme appears, and dictionary data that indicates a phoneme string of a plurality of words and / or sentences. The phoneme string is generated by applying the language model and the acoustic model to the feature quantity, and the generated phoneme sequence is indicated in the dictionary data. A word and / or sentence having a similar phoneme sequence, and when a word and / or a sentence having a similar phoneme sequence is shown in the dictionary data, the word and / or the sentence is used as a recognition result. While performing the recognition process of the above, and after generating the phoneme sequence by applying the language model and the acoustic model to the feature amount, the phoneme sequence is converted into kana, and the word and / or the sentence converted into kana is converted into kana. Entered A speech recognition method, characterized in that a second recognition processing is performed to obtain a recognition result of the generated speech signal. 6. A speech input with both a word and / or a sentence obtained by performing the first recognition processing and a word and / or a sentence obtained by performing the second recognition processing. 6. The signal recognition result is used as a recognition result.
The voice recognition method described. 7. A partial match between a word and / or a sentence shown in the dictionary data and a word and / or a sentence obtained by the second recognition processing is detected, and the partially matched word is detected. 6. The voice recognition method according to claim 5, wherein when the dictionary data indicates a word and / or a sentence, the word and / or the sentence is used as a recognition result of the input voice signal. 8. The reliability of a word and / or a sentence generated by the first recognition processing is evaluated, and the word and / or the sentence is recognized as a result of recognition of an input voice signal according to the reliability. The voice recognition method according to claim 5, wherein: