JPH09258776A - Voice recognition method, voice recognition device, and translating method and translating device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve accuracy of translation and the processing speed. SOLUTION: An intention key 3 is composed of various keys operated when voice is intended for a request, a desire, a question, an explanation, or the like, for instance. Among these keys, a user operates the key corresponding to an intention expressed by voice to be inputted. An operation signal corresponding to this operation is fed from a control key group 3 to a system control part 4 and received. In the system control part 4, the user's intention(talk intention) expressed by the voice of the user who is about to talk is recognized. When voice spoken in Japanese is inputted to a microphone 1, this voice is recognized according to the talk intention recognized in the system control part 4. In the system control part 4, the voice recognized result is translated into English also according to the talk intention.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method, a speech recognition apparatus, a translation method and a translation apparatus. In particular, by translating an input sentence to be translated according to an intention expressed by the input sentence, a voice recognition method and a voice recognition device that can accurately and quickly translate the input sentence. , And a translation method and a translation device.

[0002]

2. Description of the Related Art In recent years, research and development of an automatic translation device for translating an input sentence input in a predetermined language into another language have been actively conducted in various companies and research periods. Furthermore, recently, so-called voice translators have been researched and developed which can input input sentences to be translated by voice by incorporating a voice recognition device in the translator.

[0003]

By the way, in the present technology, even when translating an input sentence input from a keyboard, for example, there is a certain limit to the accuracy of the translation, and the accuracy of the translation is limited. For recognition accuracy in recognition,
There are still some limits.

For this reason, the translated sentence obtained by recognizing the input sentence by voice and translating the voice recognition result may be inferior in accuracy.

The present invention has been made in view of such a situation, and improves both the accuracy of translation and the accuracy of speech recognition, thereby improving the accuracy of translation of an input sentence input by voice. It enables you to do it.

[0006]

According to a first aspect of the present invention, there is provided a voice recognition method, which comprises an input step of inputting an utterance intention represented by a voice, and when recognizing the voice based on a characteristic parameter of the voice, The feature is that the voice is recognized according to the utterance intention.

According to another aspect of the present invention, there is provided a voice recognition device, which comprises input means for inputting an utterance intention expressed by the voice, and the voice recognition means recognizing the voice based on the characteristic parameter of the voice according to the utterance intention. , Is characterized by recognizing voice.

The translation method according to claim 4 comprises an input step of inputting an intention represented by the input sentence, and a translation step of translating the input sentence into another language according to the intention represented by the input sentence. Is characterized by.

A translation device according to a seventh aspect of the present invention comprises an input means for inputting an intention represented by the input sentence, and a translation means for translating the input sentence into another language according to the intention input by the input means. It is characterized by

In the voice recognition method according to the first aspect, the utterance intention represented by the voice is input, and the voice is recognized according to the utterance intention.

In the voice recognition apparatus according to the third aspect, the input means can input the utterance intention represented by the voice, and the voice recognition means recognizes the voice according to the utterance intention. It is designed to do.

In the translation method according to the fourth aspect, the intention represented by the input sentence is input, and the input sentence is translated into another language according to the intention represented by the input sentence.

In the translation device according to the seventh aspect, the input means can input the intention represented by the input sentence, and the translation means inputs the input in accordance with the input intention by the input means. It is designed to translate sentences into other languages.

[0014]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below, but before that, in order to clarify the correspondence between each means of the invention described in the claims and the following embodiments. The features of the present invention are described as follows by adding a corresponding embodiment (however, an example) in parentheses after each means.

That is, the speech recognition apparatus according to the third aspect is
An acoustic analysis unit (for example, the acoustic analysis unit 13 shown in FIG. 2) that acoustically analyzes the input voice and outputs a characteristic parameter of the voice, and a voice recognition unit that recognizes the voice based on the characteristic parameter ( For example, a voice recognition device including the recognition processing unit 14 shown in FIG. 2), further comprising an input unit (for example, an intention key 3D shown in FIG. 1) for inputting an utterance intention expressed by voice, and voice recognition Means
The feature is that the voice is recognized according to the utterance intention.

A translation device according to a seventh aspect is a translation device for translating an input sentence input in a predetermined language into another language, and input means for inputting an intention expressed by the input sentence (for example, FIG. 1) and the translation means for translating the input sentence into another language according to the intention input by the input means (for example, the system control unit 4 shown in FIG. 1). Characterize.

Of course, this description does not mean that each means is limited to the above.

FIG. 1 shows the configuration of an embodiment of a translation apparatus (speech translator) to which the present invention is applied. This translation device is adapted to translate a voice input in Japanese or the like into English or the like.

That is, the microphone 1 is adapted to convert the voice input thereto into a voice signal as an electric signal and output it to the voice recognition section 2. The voice recognition unit 2 recognizes the voice input to the microphone 1, for example, as a word. That is, the voice recognition unit 2 includes, for example, as shown in FIG. 2, an A / D converter 12, an acoustic analysis unit 13, a recognition processing unit 14, and a large-scale word dictionary 15.

The A / D converter 12 is adapted to sample an analog voice signal output from the microphone 1 at a predetermined clock timing, quantize it, and convert it into a digital voice signal (digital data). There is.
The acoustic analysis unit 13 acoustically analyzes the voice signal output from the A / D converter 12, and thereby, for example, the power of the voice for each predetermined band, the voice of the linear prediction coefficient (LPC), the cepstrum coefficient, and the like. It is designed to extract characteristic parameters. That is, the acoustic analysis unit 13 obtains the sound power for each predetermined band by filtering the sound signal for each predetermined band by a filter bank and rectifying and smoothing the filtering result. . Alternatively, the acoustic analysis unit 13 performs, for example, linear prediction analysis processing on the input voice,
The linear prediction coefficient is calculated, and the cepstrum coefficient is calculated from the linear prediction coefficient.

The characteristic parameters obtained by the acoustic analysis unit 3 are output to the recognition processing unit 14 as they are or after being vector-quantized if necessary.

The recognition unit processing 14 is based on the characteristic parameter (or the symbol obtained by vector-quantizing the characteristic parameter) from the acoustic analysis unit 13, for example, a dynamic programming (DP) matching method or a hidden Markov model (HMM). According to a voice recognition algorithm such as the above, a voice recognition process is performed by referring to a large-scale word dictionary 15 described later, and one or more words included in the voice input to the microphone 1 are output as a voice recognition result. ing.

The large-scale word dictionary 15 stores, for example, standard patterns (or models, etc.) of words (hence, in this embodiment, Japanese words) to be subjected to voice recognition. In the recognition processing unit 14, the large-scale word dictionary 1
Speech recognition is performed for the words stored in No. 5.

Returning to FIG. 1, the control key group 3 is composed of various keys, and is operated when a predetermined input is given to the system control section 4. That is, the control key group 3 includes a confirmation / conversion key 3A, a re-search / non-conversion key 3B, which causes the system control unit 4 to recognize whether the sentence output from the output unit 7 described later is correct or incorrect. It is composed of a clear key 3C for clearing the output state (display state) of the output unit 7, an intention key 3D described later, and the like.

The system control unit 4 controls each block constituting the device and retrieves sentences stored in each of the first language sentence storage unit 5 and the second language sentence storage unit 6 to be described later, Translate from Japanese to English,
Further, the output unit 7 is configured to supply and output a sentence.

The first language sentence storage unit 5 stores a large number of sentences in the language of the voice input to the microphone 1, that is, Japanese sentences in this embodiment. The second language sentence storage unit 6 stores a sentence obtained by translating a Japanese sentence stored in the first language sentence storage unit 5 into a target language, that is, an English sentence in this embodiment. ing.

The output unit 7 is composed of, for example, a voice synthesizer, a speaker, a monitor, etc., and according to the control of the system control unit 4, the Japanese sentences stored in the first language sentence storage unit 5 and the second language. The English sentence stored in the sentence storage unit 6 is output (output with a synthetic sound or displayed).

Next, referring to the flowchart of FIG.
The operation will be described. In this translation device,
When you input a Japanese sentence (input sentence) that you want to translate into English by voice, for example, by inputting a list of words such as independent words, after the list of words is recognized by voice, it will be translated into Japanese as it is. There is. Then, this Japanese sentence is presented to the user once, and when it is confirmed that it is correct, the Japanese sentence is translated into English. Further, in this translation device, the intention of the user represented by the input sentence input by voice (the intention of the speaker (user) of the voice represented by the voice corresponding to the input sentence) can be input. When this intention is input, voice recognition and translation are performed according to the intention.

Specifically, in step S0, the user inputs the intention (speech intention) represented by the voice (input sentence) to be input by operating the intention key 3D. That is, the intention key 3 is composed of various keys that are operated when the voice is intended, for example, for a request, a desire, a question, an explanation, etc. Operate the one corresponding to the intention expressed by the voice.

An operation signal corresponding to this operation is supplied from the control key group 3 to the system control unit 4 and received. As a result, the system control unit 4 causes the user's intention to be expressed by the voice that the user is about to speak (hereinafter, as appropriate,
The utterance intention) is recognized.

Then, in step S1, the microphone 1
On the other hand, when a voice corresponding to the utterance intention input in step S0 is input, the voice is converted into a voice signal as an electric signal and output to the voice recognition unit 2. Here, the voice may be a utterance of a sentence, or may be a list of independent words (words) in the sentence as described above. That is, to obtain an English translation of a sentence such as "Please tell me your favorite restaurant", you can enter the word string "You, like, restaurant, teach." Therefore, the user does not feel much burden in inputting voice. The order of the words to be input does not have to match the order of the words appearing in the sentence. That is, in the above example,
For example, you can enter a word string such as "Teach, you, like, restaurant".

In the voice recognition unit 2 (FIG. 2), step S2
At, voice recognition of each word constituting the word string input by voice into the microphone 1 is performed. That is, the voice signal from the microphone 1 is output to the recognition processing unit 14 as a characteristic parameter (or symbol) through the A / D converter 12 and further the acoustic analysis unit 13. In the recognition processing unit 14, based on the output of the acoustic analysis unit 13, the speech stored in the large-scale word dictionary 15 is subjected to speech recognition as a speech recognition target.

Here, since the input voice is a word string of independent words as described above, in general, silence is inserted between words for a short period of time. Therefore,
The recognition processing unit 14 detects the silent part,
Each word constituting the input word string is recognized by speech by recognizing a word and a word break. Therefore, the recognition processing unit 14 is not limited to the continuous speech recognition, as long as it can perform simpler isolated word speech recognition, so that it is possible to prevent the increase in size and cost. As a result, it is possible to prevent the translation apparatus from increasing in size and cost.

Further, the recognition processing section 14 is supplied with the utterance intention recognized by the system control section 4, and the recognition processing section 14 performs voice recognition in accordance with the utterance intention. That is, the words stored in the large-scale word dictionary 15 are, for example, requests, desires, questions,
The recognition processing unit 14 is categorized into those that represent explanations.
Of the words stored in the large-scale word dictionary 15,
Only the words classified as representing the utterance intention from the system control unit 4 are subjected to voice recognition, and the voice input to the microphone 1 is voice-recognized.

Therefore, the voice input to the microphone 1 is voice-recognized only for the word used for expressing the utterance intention represented by the voice, and the recognition accuracy can be improved. Furthermore, in this case, the number of words to be subjected to voice recognition is reduced, so that the recognition processing speed can be improved.

It should be noted that the words stored in the large-scale word dictionary 15 are classified into, for example, ones representing the utterance intention of any one of a request, a desire, a question, an explanation, etc.
Classifying into two or more utterance intentions, that is, 2
It is also possible to classify the above intentions redundantly.

The word string consisting of one or more words obtained as a result of the voice recognition is output to the system control unit 4. From the voice recognition unit 2 (recognition processing unit 14), the system control unit 4
When one or more words as the speech recognition result are received, in step S3, the sentence most similar to the combination of the words is written in the Japanese sentence stored in the first language sentence storage unit 5 (hereinafter, appropriately Japanese Search the word).

Here, the search in step S3 is performed as follows, for example. That is, the system control unit 4 searches the first language sentence storage unit 5 for a Japanese sentence including all the words (hereinafter, appropriately referred to as a recognized word) obtained as a result of the voice recognition. If such a Japanese sentence exists, the system control unit 4 reads the Japanese sentence from the first language sentence storage unit 5 as the one most similar to the combination of the recognized words. If none of the Japanese sentences stored in the first language sentence storage unit 5 include all the recognized words, the system control unit 4 includes all the words except one of them. Search for Japanese sentences. If such a Japanese sentence exists, the system control unit 4 reads the Japanese sentence from the first language sentence storage unit 5 as the one most similar to the combination of the recognized words. If such a Japanese sentence does not exist, the system control unit 4 searches for a Japanese sentence including all of the recognized words except for any two words. In the same manner, the Japanese sentence most similar to the combination of recognized words is searched.

The first language sentence storage unit 5 stores a Japanese sentence together with the intention represented by the sentence, and the system control unit 4 stores the Japanese sentence stored in the first language sentence storage unit 5. Of the above, only the object corresponding to the utterance intention input in step S0 is subjected to the above-described search.

Therefore, the Japanese sentence most similar to the combination of words obtained as a result of the speech recognition is searched for only the Japanese sentence stored in the first language sentence storage unit 5 that corresponds to the utterance intention. Since it is performed as a target, the search accuracy can be improved. Further, in this case, the number of Japanese sentences to be searched is reduced, so that the search processing speed can be improved.

After the Japanese sentence most similar to the combination of the recognized words is retrieved as described above, step S
4, the Japanese sentence is supplied from the system control unit 4 to the output unit 7 and output. That is, the Japanese sentence stored in the first language sentence storage unit 5 is added with, for example, the character code of the characters forming the Japanese sentence, or the phonological information and the prosody information of the Japanese sentence. The data is supplied to the output unit 7. In the output unit 7, the character corresponding to the character code is displayed on the monitor. Alternatively, in the output unit 7, the voice synthesizer generates a synthetic sound based on the phoneme information and the prosody information, and outputs the synthesized sound from the speaker. That is, in the output unit 7, the Japanese sentence is displayed on the monitor, or the Japanese sentence is output as a synthetic voice (in the present specification, when displaying,
And, in any case, when outputting a synthetic sound, it is referred to as output).

After the Japanese sentence is output at the output unit 7, the user confirms whether the sentence is correct or not. That is, if the Japanese sentence output from the output unit 7 is correct, the user operates the confirmation / conversion key 3A of the control key group 3. If the Japanese sentence output from the output unit 7 is incorrect, the user operates the re-search / non-conversion key 3B of the control key group 3. Which key has been operated is determined by the system control unit 4 in step S5.

If it is determined in step S5 that the re-search / non-conversion key 3B has been operated, that is, if the Japanese sentence output in step S4 is incorrect, the process returns to step S1. Therefore, in this case, the voice is again input by the user, and the above-described processing is performed thereafter.

When it is determined in step S5 that the re-search / non-conversion key 3B has been operated, the process returns to step S2 or S3 instead of step S1 as shown by the dotted line in FIG. It is also possible. If the process returns to step S2, the voice recognition unit 2
From among the voice recognition result candidates obtained in the voice recognition processing of the voice input in step S1, the one with the next highest likelihood is output as the voice recognition result. That is, in the speech recognition unit 2, the recognition processing unit 14 calculates the likelihood that each word stored in the large-scale word dictionary 15 is close to the input speech, and normally, the likelihood is calculated. The highest word is output as the voice recognition result. Therefore, if the Japanese sentence output in step S4 is incorrect, it is possible that the word with the highest likelihood output as the speech recognition result was incorrect. , Is output as a voice recognition result, and thereafter, the processing from step S3 is repeated.

When the process returns to step S3, the Japanese sentences other than the Japanese sentences that have been output so far are searched for, and the Japanese sentences are searched. That is, the fact that the Japanese sentence output in step S4 is incorrect means that the search processing in step S3 may have been incorrect. The search is done.

In the present embodiment, as described above, since the voice recognition and the retrieval of the Japanese sentence are performed according to the utterance intention, it is considered that the number of cases where the Japanese sentence is incorrect is considerably small.

On the other hand, if it is determined in step S5 that the confirmation / conversion key 3A has been operated, that is, if the Japanese sentence output in step S4 is correct (step S4).
(If the meaning of the Japanese sentence output in step S1 matches the meaning of the voice input in step S1), the process proceeds to step S6, where the system control unit 4 composes the Japanese sentence output in step S4. Each word (Japanese word) to be converted into an English word, and the word most similar to the combination of the English words is the English sentence stored in the second language sentence storage unit 6 (hereinafter referred to as an English sentence as appropriate). Search).

This search is performed, for example, in step S3.
It is performed in the same manner as the case of searching for a Japanese sentence in. That is, as in the case of the first language sentence storage unit 5, the English sentence is also stored in the second language sentence storage unit 6 together with the intention represented by the sentence, and the system control unit 4
Of the English sentences stored in the second language sentence storage unit 6, only the sentence corresponding to the utterance intention input in step S0 is searched.

Therefore, in this case, since the search is performed for the English sentence according to the utterance intention, the search accuracy and the search processing speed can be improved. That is,
In this case, since the Japanese sentence is translated into the English sentence according to the utterance intention, the translation accuracy and the translation processing speed can be improved.

As described above, the English sentence, that is, confirmation /
When the English sentence corresponding to the Japanese sentence output from the output unit 7 is searched when the conversion key 3A is operated, the English sentence is supplied from the system control unit 4 to the output unit 7 in step S7. Is output and the processing ends. That is, in the second language sentence storage unit 6, as in the case of the first language sentence storage unit 5 described above, an English sentence is a character code of a character forming the English sentence, or phonological information of the English sentence and It is stored together with the prosody information, and these data are supplied to the output unit 7. Therefore, in step S7,
Similar to the case in step S4, the English sentence is displayed on the monitor, or alternatively, the English sentence is output as a synthetic sound.

The case where the present invention is applied to a translation device for translating a Japanese input sentence into English has been described above.
The present invention is also applicable to a translation device that translates a language other than Japanese into English, or translates Japanese into a language other than English.

In this embodiment, only the independent word is used as the voice recognition target, but words other than the independent word can be included in the voice recognition target.

Further, a voice recognition algorithm in the voice recognition unit 2, a method of generating a synthesized voice by the output unit 7, a method of searching a Japanese sentence similar to a recognized word, and an English sentence similar to a combination of English words are searched. The method is not limited to that described in this embodiment.

Further, in this embodiment, the Japanese words constituting the Japanese sentence are converted into English words to translate the Japanese sentence into the English sentence, but this translation method is also applicable. It is not particularly limited. That is, for example, the second language sentence storage unit 6 stores an English sentence in the first language sentence storage unit 5.
In step S6, the English sentence associated with the Japanese sentence output in step S4 is searched for in association with the Japanese sentence stored in. It is possible to translate into English. In this case, the Japanese sentence
Since it is associated with the translated English sentence, there is no mistake in translation.

Further, in the present embodiment, the speech recognition section 2 is adapted to perform the isolated word speech recognition. However, the speech recognition section 2 may be adapted to perform, for example, continuous speech recognition or word spotting. It is also possible to do so.

Further, in the present embodiment, the utterance intention is inputted by operating the intention key 3D, but the utterance intention can be inputted by other means such as voice. In this case, the translation device has the intention key 3D
Since it is not necessary to provide the device, the device can be downsized.

Further, in this embodiment, the first language sentence storage unit 5
Although the second language sentence storage unit 6 and the second language sentence storage unit 6 are configured separately, they may be configured integrally.

Further, in the present embodiment, the speech intention represented by the speech is inputted before the speech is inputted in step S1, but the speech intention is inputted by other means, for example, after the speech is inputted. It is also possible to

Further, in the present embodiment, the input sentence input by voice is translated, but it is also possible to translate the input sentence input by operating a keyboard or the like. is there.

[0060]

According to the voice recognition method of the first aspect and the voice recognition apparatus of the third aspect, the utterance intention represented by the voice is input, and the voice is recognized according to the utterance intention. Therefore, it is possible to improve the voice recognition accuracy and the processing speed thereof.

A translation method according to claim 4 and claim 7
According to the translation device described in (1), the intention represented by the input sentence is input, and the input sentence is translated into another language according to the intention represented by the input sentence. Therefore, it is possible to improve the translation accuracy and the translation processing speed.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of a translation device to which the present invention has been applied.

2 is a block diagram showing a detailed configuration example of a voice recognition unit 2 in FIG.

FIG. 3 is a flowchart for explaining the operation of the translation device of FIG. 1;

[Explanation of symbols]

1 microphone (microphone), 2 voice recognition unit,
3 control key group, 3D intention key, 4 system control unit, 5 first language sentence storage unit, 6 second language sentence storage unit, 7 output unit, 12 A / D converter, 1
3 acoustic analysis unit, 14 recognition processing unit, 15 large-scale word dictionary

Claims (7)

[Claims] 1. A voice recognition method comprising: an acoustic analysis step of acoustically analyzing input voice and outputting a characteristic parameter of the voice; and a voice recognition step of recognizing the voice based on the characteristic parameter. A voice recognition method, further comprising: an input step of inputting a utterance intention represented by the voice, wherein the voice recognition step recognizes the voice according to the utterance intention. 2. The voice recognition method according to claim 1, wherein the utterance intention is input by voice. 3. A voice recognition device comprising: an acoustic analysis unit that acoustically analyzes an input voice and outputs a characteristic parameter of the voice; and a voice recognition unit that recognizes the voice based on the characteristic parameter. A voice recognition device, further comprising an input unit for inputting a utterance intention represented by the voice, wherein the voice recognition unit recognizes the voice according to the utterance intention. 4. A translation method for translating an input sentence input in a predetermined language into another language, comprising: an input step of inputting an intention expressed by the input sentence; And a translation step of translating the input sentence into the other language. 5. The translation method according to claim 4, wherein the input sentence is input by voice, and the method further comprises a voice recognition step of recognizing the input sentence by voice. 6. The translation method according to claim 4, further comprising a voice recognition step of recognizing the intention represented by the input sentence by voice inputting the intention represented by the input sentence. 7. A translation device for translating an input sentence input in a predetermined language into another language, the input device inputting an intention expressed by the input sentence, and the intention input by the input device. According to the present invention, a translation device comprising: a translation unit that translates the input sentence into the other language.