JP4986301B2 - Content search apparatus, program, and method using voice recognition processing function - Google Patents

Description

  The present invention relates to a content search apparatus, program, and method using a speech recognition processing function.

  2. Description of the Related Art Conventionally, music distribution services for searching for music as content using a terminal such as a mobile phone or a personal computer have been provided. According to this service, a user inputs a keyword such as a music title or artist name into the terminal. The terminal can search the appropriate music piece by transmitting the keyword to the content search server via the network.

  In some cases, Japanese, English, numbers, and the like are mixed in the music name and artist name as keywords for searching for music. In such a case, for example, an operation for inputting a keyword to the mobile phone is troublesome for the user. In order to eliminate such trouble, some terminals have a voice input function.

  FIG. 1 is a functional configuration diagram of a content search apparatus in the prior art.

  According to FIG. 1, the content search device 1 includes a voice input unit 101, an acoustic feature amount extraction unit 102, an acoustic model storage unit 103, a language model storage unit 104, a voice recognition decoder 105, and a content search unit 106. And have. These functional components other than the voice input unit 101 are realized by executing a program that causes a computer installed in the apparatus to function.

  The voice input unit 101 inputs voice generated by a user and converts it into an electric signal (sound wave waveform). The converted sound wave waveform is output to the acoustic feature quantity extraction unit 102.

  The acoustic feature quantity extraction unit 102 extracts an acoustic feature quantity for extracting the acoustic feature quantity x from the input speech waveform. For example, a mel cepstrum coefficient (MFCC) weighted by human recognition sensitivity based on a difference in frequency may be used.

The acoustic model accumulating unit 103 accumulates an acoustic model, and outputs an acoustic probability P (x | ω) at which the acoustic feature amount x extracted from the input speech is observed with respect to the recognition result candidate word string ω. Here, ω (ω = ω ₁ , ω ₂ ,..., Ω _m ) means a word string, and ω _m means a word.

  The language model accumulating unit 104 accumulates language models and outputs statistical / grammatical language probabilities Pn-gram (ω) / Pcfg (ω). As the language model, a statistical language model or a description grammar is generally used. Note that the language model storage unit 104 also normally stores word dictionaries to be recognized.

  Pn-gram (ω) is the language probability of a statistical language model called n-gram, and is an estimate of the language probability P (ω) that a word string ω appears based on the language statistics of a large number of sentences. It is. Mainly used in dictation and interactive speech recognition.

  Pcfg (ω) is a language probability of a grammar rule-based model called context-free grammar, and is a description of syntax rules by hand based on the knowledge of the language and the analysis result of the recognition task.

The speech recognition decoder 105 includes an acoustic feature quantity x, an acoustic probability P (x | ω), and a language probability P (ω) using statistical / grammatical language probabilities Pn-gram (ω) / Pcfg (ω). Based on the above, the following evaluation functions are output in the highest N recognition result word string ω ^ from the highest or highest order.
P (ω) × P (x | ω) (ω∈W, x∈X)
The searched recognition result word string ω ^ is output to the content search unit 106. For example, the recognition result word string ω ^ with the maximum probability is expressed by the following expression.
ω ^ = arg max {P (ω) × P (x | ω)} (ω∈W, x∈X)

  A so-called beam search algorithm is used for the search processing of the recognition result word string ω ^. The beam search algorithm is a search process in which only a high evaluation function P (ω) × P (x | ω) is left and a low one is pruned using a predetermined search beam width for word string candidates. . The trade-off between calculation time and recognition accuracy can be controlled by setting the beam width.

  The content search unit 106 searches the content database or the download server address from the content database using the recognition result word string ω ^ as a search key.

  As a technique using such a content search apparatus, there is a music search system in which a user connects an artist name and a song name with “no”, utters, recognizes a keyword from the sound waveform, and searches for a song. (For example, refer to Patent Document 1).

  There is also a program designating device that searches for programs using a word dictionary (see, for example, Patent Document 2). The word dictionary is obtained by deleting words that do not match the user's preference information from a word dictionary registered in advance. Recognition accuracy can be improved by reducing the search space for word strings.

Japanese Patent Application Laid-Open No. 2002-189483 JP 2004-120767 A

  According to the technique described in Patent Document 1, a word string is directly extracted from a sound wave waveform uttered, and does not reflect the user's intention or preference. In addition, all music title words registered in the word dictionary are searched with the same priority. Therefore, when a commercial music database in which about tens of thousands to hundreds of thousands of music titles are recorded is to be searched, the music name words registered in the word dictionary increase and the search space becomes enormous. For this reason, searching for all music pieces with the same priority not only takes a very long search time, but also often outputs a music name that is different from the user's intention or preference as a recognition result.

  Moreover, according to the technique described in Patent Literature 2, a reduced word dictionary is created by deleting words in a word dictionary that does not match user preference information. Therefore, a program that does not correspond to the reduced word dictionary and deviates from the preference cannot be recognized at all. In addition, when the estimation accuracy of the preference level is not high, the recognition performance may be significantly reduced. Note that the preference level should be manually input by the user every time the user uses the evaluation score.

  Therefore, the present invention has an object to provide a content search device, a program, and a method that can search for content that makes the user feel that recognition accuracy is high for each user in consideration of user preferences. .

According to the present invention, acoustic feature quantity extraction means for extracting the acoustic feature quantity x from the input speech waveform;
An acoustic model accumulating means for accumulating an acoustic model and outputting an acoustic probability P (x | ω) at which an acoustic feature quantity x is observed for a word sequence ω of word recognition result candidates composed of one or more words ω _m ,
Language model storage means for storing language models and outputting statistical / grammatical language probabilities Pn-gram (ω) / Pcfg (ω);
A speech recognition decoder that outputs a recognition result word string ω ^ based on an acoustic feature quantity x, an acoustic probability P (x | ω), and a statistical / grammatical language probability Pn-gram (ω) / Pcfg (ω);
In a content search apparatus having content search means for searching for content from a content database using the recognition result word string ω ^ as a search key,
The similarity between two contents S _{i, j} in all contents and the word ω _m (= content name M _n (M ₁ , M ₂ ,..., M) for each content searched in the past for each user U _v )) a user information storage means for storing
First weight alpha (M _q, U) of the user U with respect to the content name M _q a, between the contents between the the content name Mq, each content name M _n retrieved in the past by the user U A preference probability P ^* (ω _m ) is calculated for each word ω _m calculated from the sum of the similarities S _{Mq and Mn} and given the first weight α (M _q , U) , and one or more words ω The preference probability P ^* (ω) for each word string ω including _m is expressed as P ^* (ω) = P ^* (ω ₁ , ω ₂ ,..., ω _m ) = P ^* (ω ₁ ) × P ^* preference degree probability calculating means for calculating by (ω ₂ ) ×... × P ^* (ω _m ) ;
A language probability calculating means for outputting a language probability P (ω) obtained by weighting a preference probability P ^* (ω) to a statistical / grammatical language probability Pn-gram (ω) / Pcfg (ω). And

According to another embodiment of the content search device of the present invention,
The user information accumulating means accumulates the content category k and the search frequency among all users for the content included in each category k.
The preference degree probability calculation means calculates the second weight βk based on the ratio of all other search frequencies to the search frequency of the user U in the category k corresponding to the content name M _q , A preference probability P ^* (ω _m ) is calculated for each word ω _m by adding the second weight β k to the weight α (M _q , U).
It is also preferable.

According to another embodiment of the content search device of the present invention,
In the user information storage means, a plurality of users U are classified into categories C, and the plurality of users U included in each category C have been searched in the past by the content name Mq and the users U. Accumulating the first weight α (M _q , U _n ) of the user Un for the content name M _q calculated from the sum of the inter-content similarity S _{Mq, Mn} with each content name M _n And
Preference probability calculation means, for the category C which the user U belongs, the content name M first weighted alpha (M _q, U _n) of the user Un with respect to _q third on the basis of the sum of the weights gamma (M _q , U), and the third weight γ (M _q , U) is added to the first weight α (M _q , U) and / or the second weight βk, and the preference probability for each word ω _m Calculate P ^* (ω _m )
It is also preferable.

According to another embodiment of the content search apparatus of the present invention, the recognition result word string ω ^ is displayed to the user, and the user's correct / incorrect evaluation input operation for the recognition result word string ω ^ is received to evaluate the error. In the case of an input operation, it is preferable to further include a user evaluation unit that causes the preference level probability calculation unit to recalculate the preference level probability P ^* (ω).

  According to another embodiment of the content search device of the present invention, the speech recognition decoder uses a recognition candidate word string in which the probability that the language probability P (ω) is weighted to the acoustic probability P (x | ω) is equal to or less than a predetermined threshold. It is also preferable to use a beam search method for pruning and finally output only the top N recognition result word strings ω ^ from the highest or highest weighted probability.

  According to another embodiment of the content search device of the present invention, the content is preferably music.

According to the present invention,
A program for causing a computer installed in an apparatus for searching for content from a content database to function.
Acoustic feature quantity extraction means for extracting the acoustic feature quantity X from the input speech waveform;
An acoustic model accumulating means for accumulating an acoustic model and outputting an acoustic probability P (x | ω) at which an acoustic feature quantity x is observed with respect to a recognition result candidate word string ω composed of one or more words ω _m ;
Language model storage means for storing language models and outputting statistical / grammatical language probabilities Pn-gram (ω) / Pcfg (ω);
A speech recognition decoder that outputs a recognition result word string ω ^ based on an acoustic feature quantity x, an acoustic probability P (x | ω), and a statistical / grammatical language probability Pn-gram (ω) / Pcfg (ω);
In a content search program for causing a computer to function as a content search means for searching for content from a content database using the recognition result word string ω as a search key,
The similarity between two contents S _{i, j} in all contents and the word ω _m (= content name M _n (M ₁ , M ₂ ,..., M) for each content searched in the past for each user U _v )) a user information storage means for storing
First weight alpha (M _q, U) of the user U with respect to the content name M _q a, between the contents between the the content name Mq, each content name M _n retrieved in the past by the user U A preference probability P ^* (ω _m ) is calculated for each word ω _m calculated from the sum of the similarities S _{Mq and Mn} and given the first weight α (M _q , U) , and one or more words ω The preference probability P ^* (ω) for each word string ω including _m is expressed as P ^* (ω) = P ^* (ω ₁ , ω ₂ ,..., ω _m ) = P ^* (ω ₁ ) × P ^* preference degree probability calculating means for calculating by (ω ₂ ) ×... × P ^* (ω _m ) ;
The computer further functions as a language probability calculation means for outputting a language probability P (ω) obtained by weighting the preference probability P ^* (ω) to the statistical / grammatical language probability Pn-gram (ω) / Pcfg (ω). It is characterized by that.

According to the present invention, the acoustic feature quantity x is extracted from the input speech waveform, and the acoustic feature quantity x is obtained for the acoustic feature quantity x and the recognition result candidate word string ω composed of one or more words ω _m. Based on the observed acoustic probability P (x | ω) and the statistical / grammatical language probabilities Pn-gram (ω) / Pcfg (ω), the recognition result word sequence ω ^ is output, and the recognition result word sequence ω In the content search method in the device for searching for content from the content database using ^ as a key,
The similarity between two contents S _{i, j} in all contents and the word ω _m (= content name M _n (M ₁ , M ₂ ,..., M) for each content searched in the past for each user U _v )) has a user information storage unit for storing
First weight alpha (M _q, U) of the user U with respect to the content name M _q a, between the contents between the the content name Mq, each content name M _n retrieved in the past by the user U A preference probability P ^* (ω _m ) is calculated for each word ω _m calculated from the sum of the similarities S _{Mq and Mn} and given the first weight α (M _q , U) , and one or more words ω The preference probability P ^* (ω) for each word string ω including _m is expressed as P ^* (ω) = P ^* (ω ₁ , ω ₂ ,..., ω _m ) = P ^* (ω ₁ ) × P ^* a first step calculated by (ω ₂ ) ×... × P ^* (ω _m ) ;
A second step of outputting a language probability P (ω) obtained by weighting the preference probability P ^* (ω) to the statistical / grammatical language probability Pn-gram (ω) / Pcfg (ω). And

  According to the content search apparatus, program, and method of the present invention, in a speech recognition process, a word string is extracted using a language probability weighted by a user's preference level, and the content is searched using the word string. By this, it can be made to feel that recognition accuracy is high for every user.

  Hereinafter, the best mode for carrying out the present invention will be described in detail with reference to the drawings.

  FIG. 2 is a functional configuration diagram of the content search apparatus according to the present invention.

  According to FIG. 2, compared with FIG. 1, the content search device 1 further includes a language probability calculation unit 111, a preference probability calculation unit 112, a user information storage unit 113, and a user evaluation unit 114. It has further. These functional components are realized by executing a program that causes a computer installed in the apparatus to function.

  The user information storage unit 113 stores user information. The user information includes user search history information, similarity between contents, content access degree, content freshness, and / or user attribute information. It is also preferable to update these information with information from the content database 2.

The preference degree probability calculation unit 112 represents the preference degree probability P ^{* of the} word string ω from the preference degree probability P ^* (ω _m ) for each word ω _m that represents the degree of preference of the user for the content based on the user information ^. (ω) = P ^* (ω ₁ , ω ₂ ,..., ω _m ) = P ^* (ω ₁ ) × P ^* (ω ₂ ) × ... × P ^* (ω _m ) is calculated. Further, preference probability calculation unit 112, the calculated preference score probability P ^* (omega) accumulates, and outputs the ^* preference probability P language probability calculation unit 111 navel (omega).

The language probability calculation unit 111 outputs the language probability P (ω) weighted by the preference probability P ^* (ω) to the statistical / grammatical language probability Pn-gram (ω) / Pcfg (ω). The characteristic point of the present invention is that the language probability P (ω) weighted by the preference probability P ^* (ω) is used as the language probability P (ω). Specifically, for example, when a statistical language model is used, the language probability P (ω) weighted by the preference level is calculated by the following equation. Here, Pn-gram (ω) is weighted with a preference probability P ^* (ω).
P (ω) = P ^* (ω) × Pn-gram (ω)

In the case of recognition processing using a description grammar, the language probability P (ω) weighted by the preference level is calculated by the following equation. Here, Pcfg (ω) is 1 or 0.
P (ω) = P ^* (ω) × Pcfg (ω)

  Here, the probability P (ω) weighted by the preference level is normalized so that the sum ΣP (ω) regarding all the word strings ω becomes 1.

The user evaluation unit 114 displays the recognition result word string ω ^ that is the output of the speech recognition decoder to the user. In addition, an input operation of correctness evaluation from the user for the recognition result word string ω ^ is accepted. In the case of a correct evaluation input operation, the content search unit 106 is instructed to perform a search using the recognition result word string ω ^ as a search key. In the case of an error evaluation input operation, the preference probability calculation unit 112 recalculates the preference probability P ^* (ω).

  Further, the speech recognition decoder 105 according to the present invention uses a beam search method for pruning recognition candidate word strings in which the probability of weighting the language probability P (ω) to the acoustic probability P (x | ω) is equal to or less than a predetermined threshold. Finally, the top N ranked recognition result word strings ω ^ are output in order from the highest or highest weighted probability.

FIG. 3 is an explanatory diagram showing steps of calculating the preference probability P ^* (ω) in the preference probability calculation unit.

  According to FIG. 3, the user information accumulation unit 113 accumulates user search history information, content information, and user attribute information. The user search history information is search history information such as the music searched by the user and the search frequency thereof. The content information has a similarity between contents, a content access degree, and a content freshness. The similarity between contents (songs) is a similarity between contents calculated in advance based on music information such as genre and artist. The content access degree is, for example, the search frequency of all users of each music piece. The content freshness is, for example, a release (release) date. The user attribute information is user profile information based on the user's age group, gender, occupation, and the like.

Using these pieces of user information, the preference probability calculation unit 112 calculates various weighting coefficients α, β, and γ, and the preference degree for each word of the content name (music title, artist name, etc.) based on the weighting coefficients. The probability P ^* (ω) is calculated.

  The preference degree probability calculation unit 112 calculates a history weight α based on the similarity between contents using the user search history information. The history weight α reflects the user's search preference for all song names to be recognized, and changes so as to be directly proportional to the amount of user history information.

A calculation example of the history weight α will be described. First, the similarity S _{i, j} between music i and music _j is calculated for all music. i and j are music numbers. The value of S _{i, j} is directly proportional to the degree of correlation between the genre information of music i and j. Also, the similarity S _{i, j} can be calculated using the degree of correlation in the artist information, lyric information, and melody information of the music. For user U's search history (searched music M _i ) R (U) = {M ₁ , M ₂ ,..., M _V }, the history preference weight α (Mq, U) for music M _q is Calculated by the formula. If the user's search history is blank, α is 0.
α (Mq, U) = ΣS _{Mq, Mn} M _n ∈ R (U)
Further, when the music pieces in the database are classified into categories and the user U searches many music pieces belonging to the category k, α (M, U) (M∈k) of the music pieces in the category k with respect to the user U. ) Can be increased.

  In addition, the preference probability calculation unit 112 calculates the trend weight β based on the content access degree and the content freshness.

An example of calculating the fashion weight β will be described. The trend weight β uses the search frequency of all users of the music in order to represent the fashion of the music. As a calculation example, first, in the music category, for all songs in a certain category k, the horizontal axis is the elapsed time t from the release date of each music, and the vertical axis is the time of each music in that time zone. Statistics as daily search frequency. From the statistical result, the relational expression f _k = F _k (t) between the search frequencies f _k and t of music belonging to the category k is estimated. Based on this relationship, the trend weight β _k of the music piece in category _k is calculated by the following equation.
β _k = F _k (t) / max {F _k (t)}

  Further, the preference probability calculation unit 112 calculates an attribute weight γ based on the similarity between contents using the user attribute information. The attribute weight γ is set to be high for content that is relatively frequently searched and content that has a high degree of similarity with the content to which the user belongs.

An example of calculating the attribute weight γ will be described. The attribute weight γ estimates the preference of the user U to the music with reference to the search history (searched music information) of other users similar to the profile information of the user U. First, as a result of clustering processing of all user profile information, it is assumed that the user U belongs to the category C of the profile. The attribute weight γ (Mq, U) of the music M _{q for} the users U ₁ , U ₂ ,..., U _N belonging to the category C other than the user U is calculated by the following equation. N is the number of users in category C (excluding user U).
γ (Mq, U) = (1 / N) Σα (Mq, Un) Un, U∈C

The history weight α, the fashion weight β, and the attribute weight γ obtained in this way are set to higher values as the user's preference is met. Then, the preference degree probability calculation unit 112 uses the history weight α, the fashion weight β, and / or the attribute weight γ to calculate the preference degree probability P of the word string ω from the preference degree probability P ^* (ω _m ) for each word ω _m. ^* (ω) = P ^* (ω ₁ , ω ₂ ,..., ω _m ) = P ^* (ω ₁ ) × P ^* (ω ₂ ) × ... × P ^* (ω _m ) is calculated. The word ω _m is, for example, a content name (music name, artist name, etc.).

The preference probability P ^* (ω) is expressed as P ^* (ω) = F (α + β + γ) using a general function F related to the argument (α + β + γ) based on the history weight α, the fashion weight β, and the attribute weight γ. Can be calculated. As a specific example, calculation by the following two formulas is shown, for example.
P ^ ^* (ω) = (α + β + γ) ^P
Here, the index p is a constant set for each user. The index p is determined by a speech recognition experiment based on prerecorded speech waveforms and user information.
P ^* (ω) = P ^ ^* (ω) / (ΣP ^ ^* (ω)) (normalization of P ^* (ω))

  The index p may be adjusted according to the result of speech recognition. Next, a method for adjusting the index p will be described.

  The user evaluation unit 114 shown in FIG. 2 presents the recognition result word string ω ^ that is the output of the speech recognition decoder 105 to the user. If the recognition result word string ω ^ is correct, a “search button” in the user evaluation unit 114 is pressed by the user's operation (Yes), and a search is performed using the recognition result word string ω ^ as a search key. .

If the recognition result word string ω ^ is incorrect, the “retry button” in the user evaluation unit 114 is pressed by the user's operation (No), and the pressing operation information is notified to the preference probability calculation unit 112. Is done. As a result, the preference probability calculation unit 112 has an appropriate balance between the linguistic probability and the preference probability in the weighted language model P (ω) based on the preference probability P ^* (ω). The value of the index p is adjusted so as to decrease.

An example of automatic updating of the index p is given. When the evaluation operation that the recognition result is incorrect is received, the speech recognition decoder 105 executes the recognition process again without using the preference probability P ^* (ω). When the re-recognition result is different from the previous recognition result and the recognition score S of the re-recognition result is higher than the preset threshold value R, the preference probability P ^* (ω) is inappropriate for the user. Presumed to be (S is normalized score, value is between 0 and 1). Therefore, until the same result as the result of re-recognition is obtained, the index p is adjusted by repeating the calculation of the following equation in order to bring the value of P ^* (ω) close to 1.
n: number of repetitions p _n : value of index p adjusted for n-th time p ₀ : initial value obtained by speech recognition experiment p _n = (1-S) p _n−1

With such an operation, the speech recognition rate can be increased by gradually correcting the preference probability P ^* (ω), thereby enabling an adaptive operation.

  FIG. 4 is a system configuration diagram according to another embodiment of the present invention.

  According to FIG. 4, a terminal 3 operated by a user, a recognition server 4, an information management server 5, and a content server 6 are connected via the Internet.

  The terminal 3 includes a voice input unit 101, an acoustic feature amount extraction unit 102, and a user evaluation unit 114. The acoustic feature quantity x output from the acoustic feature quantity extraction unit 102 is transmitted to the recognition server 4 via the network.

The recognition server 4 includes an acoustic model storage unit 103, a language model storage unit 104, a language probability calculation unit 111, and a speech recognition decoder 105. The speech recognition decoder 105 receives the acoustic feature quantity x from the terminal 3 and receives the preference probability P ^* (ω) from the information management server 5. Then, the speech recognition decoder 105 transmits the recognition result word string ω ^ to the content server 6 via the network.

The information management server 5 includes a preference probability calculation unit 112 and a user information storage unit 113. The preference degree probability calculation unit 112 stores the preference degree probability P ^* (ω) calculated in advance for each user. In addition, when the recognition result word string ω ^ is incorrect according to an instruction from the user evaluation unit 114 of the terminal 3, the preference degree probability calculation unit 112 calculates the preference degree probability P ^* (ω) for each user. You may make it recalculate.

In the terminal 3, the user's utterance is converted into a speech waveform by the speech input unit 101, the acoustic feature amount x is extracted by the acoustic feature amount extraction unit 102, and the acoustic feature amount x is transmitted to the recognition server 4. At the same time, the user identification number is transmitted from the terminal 3 to the information management server 5. The information management server 5 transmits the preference probability P ^* (ω) corresponding to the user identification number to the recognition server 4.

The language probability calculation unit 111 of the recognition server 4 calculates the language probability P (ω) weighted by the preference probability P ^* (ω) using the received preference probability P ^* (ω). On the other hand, the speech recognition decoder 105 executes speech recognition processing based on the acoustic feature amount x received from the terminal 3, the acoustic probability P (x | ω), and the weighted language probability P (ω).

  The recognition result word string ω ^ output from the speech recognition decoder 105 is transmitted to the user's terminal 3 via the network. The terminal 3 displays the recognition result word string ω ^ to the user using the user evaluation unit 114. When the recognition result word string ω ^ matches the content desired by the user (in the case of correct answer), the recognition result word string ω ^ is used as a search key in the content server or the WEB search server by the user's operation. It is transmitted to the content search unit 106. The search result is returned to the terminal 3.

If the recognition result word string ω ^ does not match the content desired by the user (in the case of an incorrect answer), the user performs a voice re-input operation or does not search any more. In the case of a re-input operation, the re-input operation information is fed back to the preference probability calculation unit 112 of the information management server 5. Thus, the preference probability calculation unit 112 is used for automatic updating of the calculation model of P ^* (ω).

  With such a distributed configuration, it is possible to flexibly share the functions between the terminal and each server, and it is easy to adapt to the search performance and the user scale.

  As described above in detail, according to the content search device, program, and method of the present invention, in the speech recognition process, using the acoustic probability and the language probability weighted by the user's preference level, By extracting the word string having the highest recognition score from the word string and searching for the content using the word string, it is possible to make each user feel that the recognition accuracy is high. Even if the speech recognition decoder searches by a beam search with a strict beam width, the recognition accuracy does not deteriorate and the calculation time can be shortened. Furthermore, since the vocabulary of the word dictionary is not reduced based on the preference level, music other than the preference is also searched.

  In the various embodiments of the present invention described above, various changes, modifications, and omissions in the scope of the technical idea and the viewpoint of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.

It is a functional block diagram of the content search apparatus in a prior art. It is a functional block diagram of the content search apparatus in this invention. It is explanatory drawing showing the calculation step of the preference degree probability P ^* ((omega)) in a preference degree probability calculation part. It is a system configuration figure in other embodiments of the present invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Content search apparatus 101 Speech input part 102 Acoustic feature-value extraction part 103 Acoustic model storage part 104 Language model storage part 105 Speech recognition decoder 106 Content search part 111 Language probability calculation part 112 Preference degree probability calculation part 113 User information storage part 114 User evaluation unit 2 Content database 3 Terminal 4 Recognition server 5 Information management server 6 Content server

Claims (8)

Acoustic feature quantity extraction means for extracting the acoustic feature quantity x from the input speech waveform;
An acoustic model accumulating means for accumulating an acoustic model and outputting an acoustic probability P (x | ω) at which the acoustic feature quantity x is observed for a word string ω of word recognition result candidates composed of one or more words ω _m When,
Language model storage means for storing language models and outputting statistical / grammatical language probabilities Pn-gram (ω) / Pcfg (ω);
Speech recognition that outputs a recognition result word string ω ^ based on the acoustic feature amount x, the acoustic probability P (x | ω), and the statistical / grammatical language probability Pn-gram (ω) / Pcfg (ω) A decoder;
In a content search apparatus having content search means for searching for content from a content database using the recognition result word string ω ^ as a search key,
The similarity between two contents S _{i, j} in all contents and the word ω _m (= content name M _n (M ₁ , M ₂ ,..., M) for each content searched in the past for each user U _v )) a user information storage means for storing
First weight alpha (M _q, U) of the user U with respect to the content name M _q a, between the contents between the the content name Mq, each content name M _n retrieved in the past by the user U A preference probability P ^* (ω _m ) is calculated for each word ω _m calculated from the sum of the similarities S _{Mq and Mn} and given the first weight α (M _q , U) , and one or more words ω The preference probability P ^* (ω) for each word string ω including _m is expressed as P ^* (ω) = P ^* (ω ₁ , ω ₂ ,..., ω _m ) = P ^* (ω ₁ ) × P ^* preference degree probability calculating means for calculating by (ω ₂ ) ×... × P ^* (ω _m ) ;
Language probability calculation means for outputting a language probability P (ω) obtained by weighting the preference probability P ^* (ω) to the statistical / grammatical language probability Pn-gram (ω) / Pcfg (ω). Content search device characterized by the above. The user information accumulating unit accumulates content categories k and search frequencies of all users for content included in each category k,
The preference probability calculation means calculates a second weight βk based on the ratio of all other search frequencies to the search frequency of the user U in the category k corresponding to the content name M _q , The preference probability P ^* (ω _m ) for each word ω _m is calculated by adding the second weight β k to the weight α (M _q , U) of
The content search apparatus according to claim 1. In the user information storage means, a plurality of users U are classified into categories C, and the plurality of users U included in each category C are searched in the past by the content name Mq and the users U. content similarity S _Mq between each content name M _{n was,} was calculated from the sum of _Mn, the first weighting α (M _q, U _n) of the user Un with respect to the content name M _q accumulate And
For the category C to which the user U belongs, the preference probability calculation means calculates a third weight γ (based on the sum of the first weight α (M _q , U _n ) of the user Un for the content name M _q . M _q , U) is calculated, and the third weight γ (M _q , U) is added to the first weight α (M _q , U) and / or the second weight βk, and the preference for each word ω _m Calculate the probability P ^* (ω _m )
The content search device according to claim 1, wherein the content search device is a content search device. The recognition result word string ω ^ is displayed to the user, and the user's correct / incorrect evaluation input operation is performed on the recognition result word string ω ^. 3. The content search apparatus according to claim 2, further comprising user evaluation means for recalculating the preference probability P ^* (ω).   The speech recognition decoder uses a beam search method for pruning a recognition candidate word string whose probability that the language probability P (ω) is weighted to a predetermined threshold or less by using the acoustic probability P (x | ω). 5. The content search apparatus according to claim 1, wherein only the top N recognition result word string ω ^ from the highest or highest weighted probability is output. The content search apparatus according to claim 1 , wherein the content is music. A program for causing a computer installed in an apparatus for searching for content from a content database to function.
Acoustic feature quantity extraction means for extracting the acoustic feature quantity X from the input speech waveform;
An acoustic model accumulating means for accumulating an acoustic model and outputting an acoustic probability P (x | ω) at which the acoustic feature quantity x is observed with respect to a recognition result candidate word string ω composed of one or more words ω _m ; ,
Language model storage means for storing language models and outputting statistical / grammatical language probabilities Pn-gram (ω) / Pcfg (ω);
Speech recognition that outputs a recognition result word string ω ^ based on the acoustic feature amount x, the acoustic probability P (x | ω), and the statistical / grammatical language probability Pn-gram (ω) / Pcfg (ω) A decoder;
In a content search program for causing a computer to function as content search means for searching for content from a content database using the recognition result word string ω as a search key,
The similarity between two contents S _{i, j} in all contents and the word ω _m (= content name M _n (M ₁ , M ₂ ,..., M) for each content searched in the past for each user U _v )) a user information storage means for storing
First weight alpha (M _q, U) of the user U with respect to the content name M _q a, between the contents between the the content name Mq, each content name M _n retrieved in the past by the user U A preference probability P ^* (ω _m ) is calculated for each word ω _m calculated from the sum of the similarities S _{Mq and Mn} and given the first weight α (M _q , U) , and one or more words ω The preference probability P ^* (ω) for each word string ω including _m is expressed as P ^* (ω) = P ^* (ω ₁ , ω ₂ ,..., ω _m ) = P ^* (ω ₁ ) × P ^* preference degree probability calculating means for calculating by (ω ₂ ) ×... × P ^* (ω _m ) ;
A computer is further provided as a language probability calculating means for outputting a language probability P (ω) obtained by weighting the preference probability P ^* (ω) to the statistical / grammatical language probability Pn-gram (ω) / Pcfg (ω). A program for content search, characterized by functioning. An acoustic feature quantity x is extracted from the input speech waveform, and the acoustic feature quantity x is observed with respect to the acoustic feature quantity x and a recognition result candidate word string ω composed of one or more words ω _m. A recognition result word string ω ^ is output based on the probability P (x | ω) and the statistical / grammatical language probabilities Pn-gram (ω) / Pcfg (ω), and the recognition result word string ω ^ is a key. In a content search method in an apparatus for searching for content from a content database,
The similarity between two contents S _{i, j} in all contents and the word ω _m (= content name M _n (M ₁ , M ₂ ,..., M) for each content searched in the past for each user U _v )) has a user information storage unit for storing
First weight alpha (M _q, U) of the user U with respect to the content name M _q a, between the contents between the the content name Mq, each content name M _n retrieved in the past by the user U A preference probability P ^* (ω _m ) is calculated for each word ω _m calculated from the sum of the similarities S _{Mq and Mn} and given the first weight α (M _q , U) , and one or more words ω The preference probability P ^* (ω) for each word string ω including _m is expressed as P ^* (ω) = P ^* (ω ₁ , ω ₂ ,..., ω _m ) = P ^* (ω ₁ ) × P ^* a first step calculated by (ω ₂ ) ×... × P ^* (ω _m ) ;
A second step of outputting a language probability P (ω) obtained by weighting the preference probability P ^* (ω) to the statistical / grammatical language probability Pn-gram (ω) / Pcfg (ω). Content search method characterized by the above.

Images