JPH09244691A - Input speech rejecting method and device for executing same method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately reject vocalization which is not acoustically similar to an object vocabulary to be recognized and not reject the object vocabulary to be recognized when it is inputted by being vocalized ambiguously. SOLUTION: This method stores a character string (1) to be recognized, and generates a similar character string (3) which is acoustically similar to it (2) and also generates a rejected character string (5) which is acoustically similar to neither the character string to be recognized and the similar character string (4). Then a speech recognizing process is performed (9) among the character string to be recognized, similar character string, and character string for rejection for an input speech (6) and when the character string for rejection is present with higher than predetermined accuracy among several recognition results of high order predetermined in the decreasing order of recognition likelihood, it is judged (10) that a character string other than the character string to be recognized was vocalized.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an input voice reject method and an apparatus for implementing this method, and more particularly,
The present invention relates to an input speech reject method for determining whether or not a vocabulary input by utterance is a recognition target vocabulary and rejecting a vocabulary other than the recognition target vocabulary, and an apparatus for implementing this method.

[0002]

2. Description of the Related Art When determining whether or not a user has uttered a recognition target vocabulary, the likelihood of the recognition target vocabulary obtained as a result of input speech recognition is compared with a preset threshold, and this is smaller than the threshold. Had determined that the utterance was other than the recognition target vocabulary. However, the likelihood for the recognition target vocabulary varies due to the influence of noise, speaker difference, and other conditions. Therefore, considering that recognition of unspecified speakers with different speakers is performed in a noisy environment,
It is difficult to stably determine whether or not the recognition target vocabulary is uttered, depending only on the likelihood that changes due to the influence of noise or speaker differences or other conditions.

Therefore, a method of normalizing the likelihood for this recognition target vocabulary has been considered. This is different from a model in which the likelihood is high for a predetermined specific recognition target vocabulary, and a likelihood normalization that outputs a certain likelihood to all syllable or phoneme environments. It is a method that uses a model.The normalized model is used to normalize the likelihood for the recognition target vocabulary, and whether the recognition target vocabulary is uttered by comparing the normalized model with a threshold value. It is to determine. The normalized likelihood is the likelihood normalized using a garbage model, which is a recognition model constructed by speech different from the recognition target vocabulary, or the likelihood output from the speech typewriter. Used as an example.

[0004]

As described above, a model for stably normalizing the likelihood is required when recognizing an unspecified speaker or performing voice recognition in a telephone environment. For example, using the garbage model above,
This has to be created, but in order to create it, it is necessary to prepare a large amount of voice data that many speakers call from various regions and utter. However, it takes a lot of money and time to collect, prepare, and improve such voice data.

Even if the likelihood is stably normalized using the normalization model, if a syllable string that is not originally the recognition target vocabulary but is similar to the recognition target vocabulary is uttered, this is recognized. There is still a high possibility of misidentifying it as vocabulary. The present invention provides an input voice reject method and an apparatus for implementing the method, which solves the above problems.

[0006]

A character string to be recognized is stored in advance, a similar character string acoustically similar to the character string to be recognized is created, and at the same time acoustically similar to the character string to be recognized and the similar character string. Do not create a reject character string, perform voice recognition processing on the input voice between the recognition target character string, similar character string and reject character string, and set a predetermined higher order of recognition likelihood. The input voice reject method is configured to judge that the utterance is other than the recognition target character string when the reject character string is present with a certain accuracy or higher among the several recognition results.

Then, an input voice reject method is constructed in which a character string to be recognized is represented in syllables, and a part or all of each syllable is replaced with another syllable acoustically similar to the syllable to create a similar character string. . In addition, we constructed an input speech reject method that replaces some or all of syllables with consonants with their vowels.

Further, a syllable for creating a reject character string is stored, a plurality of syllables are randomly selected from the stored syllables to create reject character string candidates, and reject character string candidates are created. In this paper, we constructed an input speech reject method that excludes forbidden character strings from among the above and excludes character strings having overlapping syllables that are continuous with the recognition target character string and similar character strings to create a reject character string.

Further, a part or all of each syllable of the character string to be recognized is replaced with a plurality of predetermined syllables to create a reject character string candidate, and the character to be recognized in each of the generated character string candidates. We constructed an input speech reject method that rejects character strings that are common to strings and similar character strings and uses them as reject character strings. Here, the recognition target character string storage unit 1 is provided, and the similar character string creation unit 2 that generates a character string acoustically similar to each character string previously stored in the recognition target character string storage unit 1 is provided. Rejection character string creation that includes a similar character string storage unit 3 that stores the character string created by the similar character string creation unit 2 and creates a character string that is not acoustically similar to the recognition target character string and the similar character string A rejection character string storage unit 5 that stores the character string created by the rejection character string creation unit 4, and includes a recognition target character string storage unit 1, a similar character string storage unit 3, and a reject character. It is provided with a recognition dictionary creating section 7 for creating a dictionary used for voice recognition from a character string stored in the column storage section 5, a recognition dictionary storage section 8 and a voice input section 6 for receiving a voice input. Is input to the voice input unit 6. A voice recognition unit 9 for performing voice recognition processing on the recognized voices based on the recognition dictionary stored in the recognition dictionary storage unit 8 and outputting the upper N predetermined character strings in descending order of likelihood and their likelihoods. If the reject character string is present in the top N recognition results with a predetermined accuracy or higher, the input voice reject apparatus is configured to include the reject determination unit 10 that determines that the utterance is not a recognition target. did.

Then, the similar character string creating unit 2 recognizes a syllable conversion rule storage unit 21 in which replaceable syllables are described for each syllable, and a character to be recognized based on the syllables stored in the syllable conversion rule storage unit 21. An input speech rejecting device comprising a similar character string generation unit 22 which replaces a part or all of each syllable in the string is constructed. Further, the rewriting rule of the syllable conversion rule storage unit 21 constitutes an input voice reject device which is a rule for replacing each syllable having a consonant with its vowel.

Further, the reject character string creating section 2 includes a syllable storage section 41 that stores all syllables, and a random syllable output section 42 that randomly outputs the syllables stored in the syllable storage section 41. , A reject character string candidate generation unit 43 that generates a character string of a predetermined length using the syllable output from the random syllable output unit 42, and stores a predetermined unacceptable syllable arrangement. The character string generated by the reject character string candidate generation unit 43 does not include the character string stored in the prohibited part character string storage unit 44, and is a recognition target character. If there is no part that matches the character string stored in the string storage unit 1 and the similar character string storage unit 3 for a predetermined length or more on the syllable notation, this character string is written as a reject character string. The reject character string selection unit 45 for writing to the unit 5 is provided, and generation of new reject character string candidates is predetermined until the number of character strings stored in the reject character string storage unit 5 reaches a predetermined number. We constructed an input speech reject device that repeats a finite number of times.

Further, the reject character string creating section 4 comprises a reject syllable conversion rule storage section 4-1 in which replaceable syllables are described for each syllable, and the reject syllable conversion rule storage section 4-1. A reject character string candidate generation / selection unit 4-2 for generating a character string in which a part or all of each syllable of the character string stored in the recognition target character string storage unit 1 is generated based on the syllable stored in The character string generated by the reject character string candidate generation / selection unit 4-2 is the recognition target character string storage unit 1 and the similar character string storage unit 3.
If the character string is not stored in the input voice reject device, the character string is written in the reject character string storage unit 5.

[0013]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention will be described with reference to the example of FIG. Reference numeral 1 denotes a recognition target character string storage unit that stores in advance a set A of character strings to be recognized. Reference numeral 2 is a similar character string creation unit that creates a character string acoustically similar to the character string stored in the recognition target character string storage unit 1.

A similar character string storage unit 3 stores a set B of character strings generated by the similar character string generation unit 2. Reference numeral 4 denotes a reject character string creation unit that generates a character string that is not acoustically similar to both the character string stored in the recognition target character string storage unit 1 and the character string stored in the similar character string storage unit 3. is there.

A reject character string storage unit 5 stores a set C of character strings generated by the reject character string creating unit 4. Reference numeral 6 denotes a voice input unit that receives voice input during voice recognition. 7 is a recognition target character string storage unit 1, a similar character string storage unit 3, and a reject character string storage unit 5.
A recognition dictionary creating unit that creates a dictionary used for voice recognition from the character string stored in.

Reference numeral 8 is a recognition dictionary storage unit for storing the dictionary created by the recognition dictionary creation unit 7. Reference numeral 9 performs a voice recognition process on the voice input to the voice input unit 6 based on the recognition dictionary stored in the recognition dictionary storage unit 8, and a predetermined upper N character string and its likelihood are arranged in descending order of likelihood. It is a voice recognition unit that outputs degrees. Reference numeral 10 denotes a reject determination unit, in which the character strings belonging to the set C stored in the reject character string storage unit 5 are present with a certainty L or more in the top N recognition results given as recognition results. If so, the input unit determines that the input voice is other than the recognition target, and otherwise determines that it is within the recognition target.

The above similar character string creating unit 2 is stored in the recognition target vocabulary storage unit 1 according to the syllable conversion rule storage unit 21 in which replaceable syllables are described for each syllable and the syllable conversion rule storage unit 21. It is composed of a similar character string generation unit 22 which replaces a part or all of each syllable of the character string. And
The reject character string creation unit 4 includes a syllable storage unit 41 that stores all syllables, a random syllable output unit 42 that randomly outputs the syllables stored in the syllable storage unit 41, and a syllable output from the random syllable output unit 42. Character string candidate generation section 43 for generating a character string of a predetermined length using, a prohibited partial character string storage section 44 for storing a predetermined and unacceptable syllable arrangement, and a character string candidate generation for rejection. The character string created by the unit 43 does not include the character string stored in the prohibited part character string storage unit 44, and the character string stored in the recognition target character string storage unit 1 and the similar character string storage unit 3 If there is no matching part in the syllable notation for a predetermined length or longer, this character string is written in the reject character string storage unit 5, otherwise it is not written reject Consisting of a string selection unit 45.

Here, the processing contents and operations of the above-mentioned respective parts will be described more specifically. The recognition target vocabulary is assumed to be four words, "yes", "no", "so-do", and "difference". In this case, the image of the contents of the recognition target character string storage unit 1 is as shown in FIG. The image of the contents of the syllable conversion rule storage unit 21 in which replaceable syllables are described for each syllable is as shown in FIG. 3, assuming that each syllable whose rule has consonants is replaced by its vowel. . That is, looking at this in the first line, "ka", "sa",
"Ta", "na", "ha", "ma", "ya", "ra",
"Wa", "ga", "za", "da", "ba", "pa" are
This means rewriting only the vowel "A". It should be noted that the rule of FIG. 3 is deficient when foreign words are also taken into consideration, but this point is omitted for simplification of description.

Assuming that the syllable conversion rule storage unit 21 is shown in FIG. 3, the similar character string generation unit 22 causes a part or all of each syllable of the character string of FIG. 2 stored in the recognition target character string storage unit 1. Is replaced based on the conversion rule of the storage unit 21.
The result is written in the similar character string storage unit 3 as a character string as shown in FIG. In FIG. 4, there is no similar character string corresponding to “No”. Further, in this example, similar character strings obtained for each recognition target character string do not have exactly the same character strings, but if there are the same character strings, processing to leave only one is necessary.

Now that FIG. 2 and FIG. 4 have been decided, a method of generating a reject character string will be described. First, the image of the syllable storage unit 41 storing the syllables for creating the character string for reject is shown in FIG.
The syllable shown in FIG. 5 is different from the syllable in the ordinary sense, but each syllable has a normal length of about one character and represents most of the sequence of sounds that can appear as a whole. It does not matter which one is used. For example, "one",
This is why "tsu", "a", "i", "e", and "o" are inserted.

Although 117 syllables are described in FIG. 5, if four syllables stored in the syllable storage unit 41 are randomly output to create 31 character string candidates for reject. The results shown in FIG. 6 are obtained. This result is influenced by the arrangement of syllables written in the syllable storage unit 41. The column of "property of prohibited character string" in FIG. 6 indicates whether the character string is a prohibited character string described later. When it is not a prohibited character string, it is marked with X, and when it is a prohibited character string, it is marked with O. In addition, the column of "Pros and Cons of Overlap" is
2 consecutively with the character string to be recognized and the similar character string in FIG.
It indicates whether or not it has more than the syllable overlap. If there is an overlap, it is marked with ○, and if it does not, it is marked with ×.

On the other hand, an image of the forbidden substring storage unit 44 for storing an unacceptable syllable arrangement is, for example, as shown in FIG. In FIG. 7, * represents an arbitrary character string, for example, rule 1 prohibits character strings beginning with "n",
Rule 2 also indicates that a character string beginning with "tsu" is prohibited, and rule 3 indicates that a character string including the sequence "n" is not allowed. According to the rule of FIG. 7, only the 28th character string of the character strings of FIG. 6 is prohibited. Here, it is assumed that the number of reject character strings determined to be necessary in advance is 30, and the recognition target character string is
If the overlapping of consecutive syllables with the similar character string is not more than two and it is not allowed, all the character strings in FIG. 6 have no problem with the overlapping, so the 30 strings except the 28th character string are excluded. The character string will be written in the reject character string storage unit 5 as a reject character string.

In the actual reject process, the character string shown in FIG.
The reject character string selection unit 45 generates the individual character strings and determines whether the character strings are the prohibited character strings in the third column in FIG. 6 and whether the recognition target character string and the similar character string in the fourth column overlap two or more consecutive phoneme strings. It is checked whether or not there is any, and it is determined whether or not to write in the reject character string storage unit 5. This process is performed until the number of character strings stored in the reject character string storage unit 5 reaches a predetermined number, or the number of times of this repetition reaches a predetermined number.

Next, the recognition dictionary creating unit 7 creates a dictionary to be used for voice recognition based on the character strings in the recognition target character string storage unit 1, the similar character string storage unit 3, and the reject character string storage unit 5. This dictionary is stored in the recognition dictionary storage unit 8. This completes the preparation for recognizing the input voice. Here, a case where the word “Maruchimedia” not included in the recognition target character string storage unit 1 is input will be described as an example. Top 3 when the character string "Maruchimedia" is input to the voice recognition device described in the following document.
The recognition result up to the rank is shown in FIG. The top two candidates for all three times were reject character strings. All of these three utterances are rejected because the first-ranked recognition result is the reject character string. Here, there are some character strings for rejection, which are targeted for the top several candidates, for example, the top three candidates, and have a small likelihood difference of less than or equal to a predetermined accuracy, for example, 2000 or less with respect to the first place candidate. In this case, if it is determined that this input voice is a voice other than the recognition target, the three voices "Maruchimedia" will be reliably rejected. Even if the recognition result of the first place is the recognition target character string, if there is a reject character string whose likelihood difference with the first place candidate is 2000 or less, similarly, the input voice is uttered other than the recognition target. (References: Yamada, Noda, Imoto, Sagayama: "HMM-LR continuous speech recognition system with client / server configuration and its application", IPSJ SIG-S
LP94-5, pp. 39-46, 1995.2. ).

When the present invention is carried out, in addition to the reject character string created by the above-described procedure, in addition to this, the vocabulary that is likely to be uttered by the speaker by mistake in the scene of use is rejected. It is considered that the effect of reject will be improved by registering it as a character string. When the present invention is applied to, for example, a telephone, it is considered preferable to add an idiom such as "Hello!" Used in ordinary telephones to the speech recognition at the beginning.

Another embodiment of the present invention will be described with reference to FIG. The embodiment of FIG. 9 and the embodiment of FIG. 1 differ in the reject character string creation unit 4, but all other configurations are common. The reject character string creation unit 4 stores a predetermined rule for replacing each syllable with another syllable different from it, and the reject syllable conversion rule storage unit 4
-1, and a reject character string in which a part or all of each syllable of the character string previously stored in the recognition target character string storage unit 1 is replaced according to this rule is generated, and the character string is generated for each generated character string. Is included in the recognition target character string storage unit 1 and the similar character string storage unit 3 only, and includes a reject character string candidate generation / selection unit 4-2 which performs an operation of writing in the reject character string storage unit 5.

Here, the recognition target character strings are only two, "Yes" and "No", out of the four character strings in the previous embodiment. In this case, the image of the recognition target character string storage unit 1 is shown in FIG. 10, and the image of the similar character string storage unit 3 is shown in FIG.
It becomes 1. Hereinafter, the syllable conversion rule storage unit for reject 4-
1 and the movements of the reject character string candidate generation / selection unit 4-2, and an example of the case where they are recognized will be described.

For the sake of simplicity, the contents of the reject syllable conversion rule storage unit 4-1 are considered to be those shown in FIG. In this case, the reject character string candidate generation / selection unit 4
For -2, the reject character string candidates shown in FIG. 13 are generated. In this example, the character strings shown in FIG. 13 are different from the character strings shown in FIG. 10 and the character string shown in FIG. If the vowels "ei" and "ou" continue,
Do you want to pronounce the vowel exactly as it is?
It is necessary to consider whether to lengthen the sound or allow both, like this, but in this example, it is considered that the sound is accurately pronounced.

Next, a dictionary used for voice recognition is created by the recognition dictionary creating unit 7 from the character strings in the recognition target character string storage unit 1, the similar character string storage unit 3, and the reject character string storage unit 5, and the dictionary is It is stored in the recognition dictionary storage unit 8, and preparation for recognizing the input voice is completed. Here, a case will be described as an example where "Maruchimedia" is input as a word that is not in the recognition target character string, as in the previous embodiment. FIG. 14 shows the recognition results of the top three ranks when this character string is applied to the voice recognition device described in the above document. In all three cases, the top three candidates were all reject character strings. When the rejected character strings that have a difference of 2000 or less from the first-ranked candidates are targeted, and the input voice is determined to be a utterance other than the recognition target. "Maruchimedia" is the three vocalizations
Will be correctly rejected.

When the present invention is embodied, in addition to the reject character string created by the above-mentioned procedure, in addition to this, the vocabulary that is likely to be erroneously uttered by the speaker in the use scene is rejected. It is considered that the effect of reject will be improved by registering it as a character string. When the present invention is applied to, for example, a telephone, it is considered preferable to add an idiom such as "Hello!" Used in ordinary telephones to the speech recognition at the beginning.

[0031]

As described above, according to the present invention, a similar character string acoustically similar to a recognition target character string is created, and a reject character string acoustically dissimilar to the recognition target character string and the similar character string. When the voice recognition process is performed between the recognition target character string, the similar character string and the reject character string for the input voice, and the reject character string exists in the recognition result, the recognition target character It is assumed that the utterance is other than the line. In this way, the utterance is rejected based on the fact that the rejected character string that is not acoustically similar to the recognition target character string and the similar character string is recognized, so that a utterance that is not acoustically similar to the recognition target vocabulary is made. In this case, this can be properly rejected, and when the present invention is used in a speaker recognition device or other recognition device, its malfunction can be prevented.

A vocabulary acoustically similar to the recognition target vocabulary is also stored in advance together with the recognition target vocabulary and is used as the recognition target vocabulary. Therefore, even when the recognition target vocabulary is slightly ambiguously input. It is possible to avoid rejecting this.

[Brief description of drawings]

FIG. 1 illustrates an embodiment.

FIG. 2 is a diagram showing an image of a recognition target character string.

FIG. 3 is a diagram showing an image of a syllable conversion rule storage unit.

FIG. 4 is a diagram showing an image of a similar character string storage unit.

FIG. 5 is a diagram showing an image of a syllable storage unit.

FIG. 6 is a diagram showing candidates for reject character strings.

FIG. 7 is a diagram showing an image of a prohibited character string storage unit.

FIG. 8 is a diagram showing a recognition result.

FIG. 9 is a diagram illustrating another embodiment.

FIG. 10 is a diagram showing an image of a recognition target character string.

FIG. 11 is a diagram showing an image of a similar character string storage unit.

FIG. 12 is a diagram showing an image of a syllable conversion rule storage unit.

FIG. 13 is a diagram showing rejected character string candidates.

FIG. 14 is a diagram showing a recognition result.

[Explanation of symbols]

 1 recognition target character string storage unit 2 similar character string creation unit 3 similar character string storage unit 4 reject character string creation unit 5 reject character string storage unit 6 voice input unit 7 recognition dictionary creation unit 8 recognition dictionary storage unit 9 voice recognition Division 10 Rejection judgment section

Claims (10)

[Claims] 1. A reject character string that stores a recognition target character string in advance, creates a similar character string acoustically similar to the recognition target character string, and does not acoustically resemble the recognition target character string and the similar character string. Create a speech recognition process between the recognition target character string, a similar character string, and the reject character string for the input speech, and recognize a predetermined number of higher-ranked recognition sequences in descending order of recognition likelihood. An input voice reject method, characterized in that when a reject character string is present in a result with a certainty or higher, the utterance is other than the recognition target character string. 2. The input voice reject method according to claim 1, wherein the recognition target character string is represented in syllables, and a part or all of each syllable is replaced with another syllable acoustically similar to the syllable. An input voice reject method characterized by creating a similar character string. 3. The input voice reject method according to claim 2, wherein a part or all of each syllable having a consonant is replaced with its vowel. 4. The input voice reject method according to claim 1, wherein a syllable for creating a character string for reject is stored,
Randomly select a plurality of stored syllables to create reject character string candidates, exclude prohibited character strings from reject character string candidates, and connect with the recognition target character string and similar character strings. Then, an input speech reject method is characterized in that a character string for rejection is created by eliminating character strings having overlapping syllables. 5. The input voice reject method according to claim 1, wherein a part or all of each syllable of the recognition target character string is replaced with a plurality of predetermined syllables. An input speech reject method characterized in that a character string candidate for rejection is created, and a character string common to a recognition target character string and a similar character string among the generated character string candidates is excluded to be a character string for rejection. . 6. A recognition target character string storage unit is provided, and a similar character string creation unit for generating a character string acoustically similar to each character string stored in advance in the recognition target character string storage unit is provided. A similar character string storage unit that stores the character string created by the character string creation unit is provided, and a reject character string creation unit that creates a character string that is not acoustically similar to the recognition target character string and the similar character string is provided. The character string storage unit for storing the character string created by the reject character string creation unit is provided, and the characters stored in the recognition target character string storage unit, the similar character string storage unit, and the reject character string storage unit. It is equipped with a recognition dictionary creation unit that creates a dictionary used for voice recognition from a string, a recognition dictionary storage unit, a voice input unit that receives voice input, and a voice recognition unit that recognizes the voice input to the voice input unit. Record The speech recognition processing is performed based on the recognition dictionary stored in the unit, and the speech recognition unit includes a speech recognition unit that outputs predetermined upper N character strings and their likelihoods in descending order of likelihood. An input voice rejecting device, comprising: a reject determining unit that determines that the rejected character string is a utterance that is not a recognition target when the reject character string exists with a certainty or higher. 7. The input voice reject apparatus according to claim 6, wherein the similar character string creating unit includes a syllable conversion rule storage unit in which replaceable syllables are described for each syllable, and a syllable conversion rule storage unit. An input speech rejecting device comprising a similar character string generation unit that replaces a part or all of each syllable of a character string to be recognized based on a stored syllable. 8. The input voice reject apparatus according to claim 7, wherein the rewriting rule of the syllable conversion rule storage unit is a rule for replacing each syllable having a consonant with its vowel. 9. The input voice reject device according to claim 6, wherein the reject character string creating unit includes a syllable storage unit that stores all syllables. Character string candidates for rejection that include a random syllable output unit that randomly outputs the syllables stored in the section, and generate a character string of a predetermined length using the syllables output from the random syllable output unit A prohibited part character string storage part for storing a predetermined sequence of unacceptable syllables, and a character string generated by the reject character string candidate generation part is stored in the prohibited part character string storage part. Character string that does not include a character string that is stored in the recognition target character string storage unit and the similar character string storage unit, and has a portion that matches the syllable notation for a predetermined length or more. A reject character string selection unit that writes this character string to the reject character string storage unit is provided, and new reject character string candidates are provided until the number of character strings stored in the reject character string storage unit reaches a predetermined number. The input voice rejecting device characterized in that the generation of the above is repeated a predetermined finite number of times. 10. The input voice reject device according to claim 6, wherein the reject character string creating unit describes a syllable for rejection in which replaceable syllables are described for each syllable. A conversion rule storage unit is provided, and a character string in which a part or all of each syllable of the character string stored in the recognition target character string storage unit is replaced is generated based on the syllable stored in the rejection syllable conversion rule storage unit. A character string candidate generation / selection unit for reject
If the character string generated by the reject character string candidate generation / selection unit is a character string that is not stored in the recognition target character string storage unit or the similar character string storage unit, write this character string in the reject character string storage unit. An input voice reject device characterized by: