JP3090344B2 - Voice recognition device - Google Patents

Abstract

PURPOSE:To obtain the speech recognition device which is adapted to a momentary change in environment and has a high recognition rate. CONSTITUTION:The speech recognition device has voice recognition dictionaries 5-1-5-n generated by superposing plural kinds of noise and is equipped with an input part 1 for inputting a voice, an analysis part 2 which analyzes the inputted voice, a storage part 6 stored previously with time information on, for example, time length, time, a period, etc., regarding the generation of a noise, selection parts 6 and 4 which selects the voice recognition dictionary generated by superposing the noise corresponding to the current time from a timer 61 according to the information in the storage part, and a recognition part 3 which collates the analysis result of the analysis part with the voice recognition dictionary selected by the selection part and recognizes the inputted voice.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus used for sorting luggage, ticket vending machines, and the like.

[0002]

2. Description of the Related Art In recent years, the technology of speech recognition / synthesis has been remarkably developed as a man-machine interface, and speech recognition devices have been used for baggage sorting, ticket vending machines, and the like. However, in many current speech recognition devices, the recognition rate decreases when the noise in the environment in which the device is used increases. Therefore, a method of using a speech recognition dictionary created from learning speech data in which the noise is artificially superimposed, or a method of removing noise has been used.

[0003] However, when using a speech recognition dictionary created from learning speech data in which noise is artificially superimposed, the surrounding environment cannot be improved unless the type of noise is limited. Since it is not uniform and constantly changes, it is difficult to use the speech recognition dictionary with limiting noise.

Further, even when a method of removing noise is used, since noise is generally non-stationary and colored, it is difficult to grasp the characteristics of noise as a sound wave and the timing of the generation of the noise. Therefore, when the surrounding environment changes, noise removal suitable for the generated noise cannot be sufficiently performed.

[0005]

As described above, in the conventional speech recognition apparatus, even if the above-described noise countermeasures are taken, it is not possible to accurately cope with noise whose surrounding environment changes every moment. This was a factor that reduced the recognition rate.

The present invention has been made in view of such circumstances, and an object of the present invention is to sufficiently improve a recognition rate with respect to noise that changes momentarily with changes in the surrounding environment. An object of the present invention is to provide a highly practical voice recognition device that can be achieved.

[0007]

According to a first aspect of the present invention, there is provided a speech recognition apparatus including a plurality of speech recognition dictionaries created by superposing a plurality of types of noises, an input unit for inputting speech, and an input unit for inputting speech. An analysis unit that analyzes voice, a storage unit that stores in advance temporal information such as a time length, a time, and a cycle related to the occurrence of noise, and a noise that corresponds to the current time from a timer and is created. And a recognition unit that recognizes the input speech by comparing the analysis result of the analysis unit with the speech recognition dictionary selected by the selection unit. It is characterized by having.

[0008] A speech recognition apparatus according to a second aspect of the present invention includes an input unit for holding a plurality of types of noise data and inputting speech.
A selection unit for selecting noise data corresponding to the current time using temporal information on the occurrence of noise stored in advance in the same manner as described above, and an input using the noise data selected by the selection unit. An analysis unit that removes noise from the obtained voice and analyzes the noise-removed voice, and a recognition unit that recognizes the input voice by comparing the analysis result of the analysis unit with a voice recognition dictionary. Things.

[0009]

According to the present invention, the time and the time length at which a certain noise occurs are obtained by referring to the temporal information of the noise occurrence stored in advance, and the same noise as the noise occurring at that time is superimposed. By selecting a speech recognition dictionary created from speech data, a speech recognition dictionary with limited types of noise can be used even under environmental noise that changes over time. Similarly, by referring to temporal information on the occurrence of noise stored in advance and selecting noise data corresponding to the noise generated at that time and performing noise removal, even in an environment where noise changes with time,
Noise can be appropriately removed from the input voice. Therefore, the performance of voice recognition can be improved. Such a noise countermeasure can be fully automatically performed by monitoring the time at which the target voice is input.

[0010]

An embodiment of the present invention will be described below with reference to the present invention. ○ Example 1

FIG. 1 is a block diagram of a speech recognition apparatus according to a first embodiment of the present invention. This device comprises a voice input unit 1, a sound analysis unit 2, a voice recognition unit 3, a recognition dictionary switch 4, a recognition dictionary unit 5, and a noise countermeasure unit 6.

The voice input unit 1 converts an input voice converted into an electric signal through a microphone or the like into a digital signal having a quantization bit number of 16 bits at a sampling frequency of 12 kHz, for example, and outputs the digital signal to the acoustic analysis unit 2. . The acoustic analysis unit 2
ET analysis (frequency analysis by fast Fourier transform), LP
For example, a characteristic parameter is obtained every 8 ms by any one of analysis methods such as C (linear prediction method) analysis, cepstrum analysis, and filter analysis, and the time series is output to the speech recognition unit 3. The voice recognition unit 3 compares the feature parameter obtained from the acoustic analysis unit 2 with the reference feature parameter of the recognition dictionary of the recognition dictionary unit 5, calculates the similarity, and performs voice recognition. For example, the frequency analysis pattern obtained by the FET analysis in the acoustic analysis unit is compared with the reference pattern of each word in the recognition dictionary, and the similarity is calculated by the composite similarity method to obtain the most similarity. A word with a high degree is output as a recognition result. The recognition dictionary used here is selected by the noise countermeasure unit 6 described later.

Here, the recognition dictionary unit 5 includes a plurality of recognition dictionaries 1.
To n. Each recognition dictionary superimposes the characteristic environmental sound unique to the environment where the voice recognition device is installed on the learning voice collected in a quiet environment, and obtains a pattern-deformed voice data. It was created. For example, if the device is installed in a school elevator,
Recognition dictionary 1 is designed to reduce environmental noise including footsteps and speech when the elevators are most crowded at school in the morning.
The recognition dictionary 2 is created by voice data obtained by superimposing environmental noises slightly containing the power sound of the elevators on the learning voice when the elevators are relatively less crowded in class. It is. Alternatively, for example, when the present apparatus is installed in a station yard, the recognition dictionary 1 includes environmental noise including a moving sound of a person accompanying the arrival and departure of a train, and the recognition dictionary 2 includes a bell sound generated by a home. The environmental noise and the recognition dictionary 3 are created by voice data in which environmental noise including train power noise is superimposed on the learning voice. Note that there is also a method in which these recognition dictionaries are created from voice data obtained by inputting and analyzing voices in which environmental sounds in each situation and learning voices are mixed.

The noise countermeasure unit 6 includes a timer (61), a time information management table 1 (62), and a countermeasure unit (63). For example, as shown in Table 1, the time information management table 1 holds the correspondence between time information (day of the week, time, time, etc.) to be used by each recognition dictionary and the recognition dictionary (dictionary number). ing. This table is created by referring to the time of departure and arrival of a train from a timetable, for example, in the latter example described above.

[0015]

[Table 1]

The operation of the countermeasure unit (63) will be described with reference to the flowchart of FIG. First, the time is monitored by the timer (61) (S1). Next, a time zone to which the time applies is searched from items such as a day of the week, a time, and a time length of the time information management table 1 (62), and a recognition dictionary number to be used for the searched time zone is extracted (S2). . For example, if the time from the timer is 8:20:00 on Wednesday, it corresponds to 8:00:00 to 8:24:59 on Monday to Friday in Table 1, so the recognition dictionary number 1 is extracted and the timer is extracted. -From 9:25 on Saturday:
If it is 00, it corresponds to 8:00 to 9:59:59 on Saturday to Sunday in Table 1, so the recognition dictionary number 4 is extracted. Next, the recognition dictionary switch 4 is controlled so that the recognition dictionary of the extracted recognition dictionary number is used in the voice recognition unit 3 (S
3). Then, the time from the timer is monitored again and the same processing is repeated.

As described above, in the present embodiment, the recognition dictionary created by the learning speech data in which the environmental noise reflecting the environment of the place is superimposed using the temporal information on the occurrence of the environmental noise. Since it is selectively used, appropriate noise countermeasures can be taken by effectively using environmental information, and recognition performance is improved.

Note that the "timer" here means a certain period (for example, one month, one week,
Although this is a device that counts time in a certain unit (second, minute, hour) during one day or the like, it may be replaced with a normal clock (a clock with date or day of the week in some cases). ○ Example 2

FIG. 2 is a block diagram of a speech recognition apparatus according to the second embodiment. This speech recognition device includes a speech input unit 1, an acoustic analysis unit 2, a speech recognition unit 3, a noise removal unit 8, a noise data unit 9, a noise data switch 10, a recognition dictionary 11, and a noise countermeasure unit 7. .

The functions of the voice input unit 1, the sound analysis unit 2, and the voice recognition unit 3 are the same as in the first embodiment. However, the noise removal unit 8 in the sound analysis unit 2 removes environmental noise included in the input voice.

Here, a case will be described in which a technique called a spectral subtraction method is used as a method for removing noise. First, environmental noise is taken in when no speech to be recognized is input, and a power spectrum for each band frequency of the noise is obtained as noise data. That is, a power spectrum sequence for each band frequency: | x ₁ || x ₂ | ... | x _n | (n is the number of bands,
| X _i | is the power spectrum at the band frequency i) as the noise data. The environmental noise taken in to create this noise data is a characteristic environmental sound unique to the environment in which the present voice recognition device is installed. A plurality (1 to n) of noise data is held in the noise data unit 9. For example, if this device is installed in the station yard, noise data 1
Is a power spectrum obtained by inputting a bell sound emitted from a bell at a fixed position according to the arrival and departure of a train from a microphone at a fixed position. This is a power spectrum obtained by inputting a moving noise.

The noise counter 7 comprises a timer (71), a time information management table 2 (72), and a counter (73). For example, as shown in Table 2, the time information management table 2 holds the correspondence between time information (day of the week, time, time length, etc.) at which each noise data is to be used and the noise data. I have. This table is created with reference to the timetable in the example described above.

[0023]

[Table 2]

As in the first embodiment, the countermeasure unit (73) monitors the current time by the timer (71), and reads the noise data to be used at a certain time from the time management information table 2 (72). -Extract the data number. For example, when the current time is 8:00: 00 on Thursday, the time from Monday to Friday, 8:00:
The noise data number 1 corresponding to 00 to 8:00:05 is extracted. Next, the noise data of the extracted noise data number is used so that the noise data is used by the noise removing unit 8.
Switch 10 is controlled.

The noise elimination unit 8 obtains | x _{i of the} noise data selected by the noise countermeasure unit 7 from the power spectra of all the band frequencies of the voice input to the voice input unit 1.
| (I = 1, 2,..., N) is subtracted. That is, (power spectrum of input voice | y _i | -above | x
_i |) is performed.

When the speech to be recognized is input together with noise, as described above, the noise component is removed by the acoustic analyzer to estimate the power spectrum of the speech signal to be recognized. The voice recognition unit 3 performs recognition by collating with the reference pattern in the recognition dictionary 11 in the voice recognition unit 3.

Other methods for removing noise include active noise control using an adaptive filter. This is to remove noise using an input unit for inputting speech mixed with noise, an input unit for inputting only noise, and an adaptive filter. In this case, the noise corresponding to each time zone is corresponded. Set the initial value of the adaptive filter to noise data 1
Ｎn.

As described above, when the environmental sound changes depending on the time zone, noise is removed by selectively using noise data reflecting the local environment using temporal information on the occurrence of environmental noise. Therefore, the recognition performance can be improved by increasing the S / N ratio.

In the first and second embodiments, the environmental noise to be superimposed at the time of creating the recognition dictionary and the environmental noise taken at the time of creating the noise data correspond to the situation as described above. It is possible to collect things and make them correspond to time later when creating the time information management table, or to first describe the temporal change of the situation in the form of the time information management table, The environmental noise in each time zone may be collected, and a recognition dictionary or noise data of the number may be created. ○ Example 3

FIG. 4 is a block diagram of a speech recognition apparatus according to the third embodiment. This speech recognition device comprises a first and a second
The present embodiment is a combination of the above embodiments, and includes both a recognition dictionary unit 5, a recognition dictionary switch 4, a noise removing unit 8, a noise data unit 9, and a noise data switch 10. Further, the noise countermeasure unit 67 holds both time information management tables 1 and 2.

The processing flow will be briefly described. The noise countermeasure unit 67 knows the time when the sound was input from the timer,
The noise data of the noise data section 9 is selected with reference to the time information management table 2, and the recognition dictionary of the recognition dictionary section 5 is selected with reference to the time information management table 1. Here, a table in which the time information management tables 1 and 2 are combined and which has a set of time information items, noise data numbers, and recognition dictionary numbers may be provided. Selected noise data
The sound analysis unit 2 and the noise removal unit 8 therein.
Performs noise removal together with the analysis, and outputs the result to the speech recognition unit 3. The speech recognition unit 3 performs recognition using the selected recognition dictionary. Here, the recognition dictionary is created based on speech data obtained by superimposing noise remaining after the noise removal using the corresponding noise data on the learning speech.

In this embodiment, even when the effect of noise removal in the second embodiment is insufficient, the recognition performance is further improved by selectively using a recognition dictionary corresponding to the noise that has not been completely removed. improves. ○ Example 4

FIG. 5 is a block diagram of a speech recognition apparatus according to the fourth embodiment. This speech recognition apparatus is a combination of the first embodiment and a function of learning a recognition dictionary by a word spotting method. The configuration of the first embodiment includes a learning speech data file 12, It has a configuration in which a voice data processing unit 13 and a recognition dictionary creating unit 14 are added.

The learning of the recognition dictionary by the word spotting method can use the method described in Japanese Patent Application No. 1-255270. When the voice to be recognized is not input from the voice input unit 1 (only environmental noise is input), the environmental noise input from the voice input unit 1 and analyzed by the acoustic analysis unit 2 and the learning voice data The learning voice data of the data file 12 is superimposed on the learning voice data processing unit 13, and the result is sent to the recognition dictionary creating unit 14 through the acoustic analysis unit 2.

In the recognition dictionary creating section 14, the determination as to which recognition dictionary of the recognition dictionary section 5 is to be learned is as follows.
This is based on information on the correspondence between the time at which a certain recognition dictionary is to be used and the dictionary held in the noise suppression unit 6. For example, a recognition dictionary of a dictionary number whose day, time, and time length of the time information management table 1 match the time from the timer 7 is selected, the recognition dictionary switch 4 is controlled, and the selected recognition dictionary is stored in the corresponding dictionary. Learning with environmental noise at the time.

In this embodiment, since the recognition dictionary is created using the noise peculiar to the environment in the completely same environment when the speech recognition apparatus is actually used, the effect of the first embodiment is further added. The performance of speech recognition can be improved. ○ Example 5

FIG. 6 is a block diagram of a speech recognition apparatus according to the fifth embodiment. This voice recognition device has a third and a fourth
The processing flow is the same as that of the third and fourth embodiments. ○ Example 6

FIG. 7 is a block diagram of a speech recognition apparatus according to the sixth embodiment. Up to the fifth embodiment, the recognition performance is improved by taking measures against noise. However, in the present embodiment, the recognition performance is improved by limiting the number of speakers. This voice recognition device includes a voice input unit 1,
It comprises an acoustic analysis unit 2, a speech recognition unit 3, a recognition dictionary 11, a speaker prediction unit 15, and a vocabulary / speaker correspondence table 16. The functions of the voice input unit 1 and the acoustic analysis unit 2 are the same as in the first embodiment.

The vocabulary / speaker correspondence table 16 is, for example, a vocabulary / speaker correspondence table (Table 3) in which a set of persons mainly speaking the vocabulary is associated with each vocabulary. is there. For example, the vocabulary "Mom-chan" is associated with a set of elementary school students.

[0040]

[Table 3]

The speaker predictor 15 is a person having a high probability of becoming the speaker of the input speech of the present speech recognition device at that time as shown in the time information management table 3 (152) (Table 4). Holds the information corresponding to the set of. Then, the time is monitored by the timer (151), and a set of persons who are likely to be speakers at that time (this is called a predicted speaker) is selected from the time information management table 3 and transmitted to the speech recognition unit 3. . The predicted speaker is, for example, dividing a person into several groups according to a certain criterion, summing up the frequency at which persons belonging to the group input voices to the apparatus during the time zone, for each group. is there.

[0042]

[Table 4]

The speech recognition section 3 calculates the similarity of the vocabulary to be recognized by the compound similarity method as described in the first embodiment. Then, for example, when recognizing a word, a vocabulary (word) in which the set of the target predicted speaker of the vocabulary / speaker correspondence table 1 coincides with the current predicted speaker reported from the speaker predictor 15 is checked The similarity of a matching vocabulary (word) is weighted to increase, and the similarity of a non-matching vocabulary (word) is reduced. Next, the recognition word is determined based on the weighted similarity.

For example, at a store at 15:30 from Monday to Friday
When a large number of elementary school students always arrive at about 17:00, the speaker predicting unit 15 uses the time information management table 3 to transmit information representing the set of elementary school students to the speech recognition unit 3. In the voice recognition unit 3, similarities are weighted for vocabularies such as "candy", "oniichan", and "mother" in the recognition dictionary 11 because the predicted speaker is an elementary school student from the vocabulary / speaker correspondence table 1. Since it becomes large, it is easy to be determined as a recognized word. Therefore, for example, when an elementary school student calls a male clerk around the age of 20 "Oniichan", the voice can be recognized and the clerk can be notified.

As described above, according to this embodiment, it is possible to obtain a high recognition rate by weighting the vocabulary similarity using the information of a person who has a high probability of being a speaker. Further, by using information of a person who has a high probability of being a speaker in association with time, it is possible to perform recognition adapted to an ever-changing environment. ○ Example 7

FIG. 8 is a block diagram of a speech soft device according to a seventh embodiment. This voice recognition device has first and sixth
Which includes both the recognition dictionary section 5, the recognition dictionary switch 4, and the vocabulary / speaker correspondence table 16, and has the functions of the noise countermeasure section 6 and the speaker prediction section 15. It has an environment adaptation unit 17. The processing flow is the same as in the first and sixth embodiments. ○ Example 8

FIG. 9 is a block diagram of a speech recognition apparatus according to the eighth embodiment. This voice recognition device has a second and a sixth
Which includes both a noise removing section 8, a noise data section 9, a noise data switch 10, and a vocabulary / speaker correspondence table 16, and includes a noise countermeasure section 7 and a speaker predictor. Environment adaptation unit 1 having the function of unit 15
Have eight. The processing flow is the same as in the second and sixth embodiments. ○ Example 9

FIG. 10 is a block diagram of a speech recognition apparatus according to the ninth embodiment. Up to the eighth embodiment, it has been assumed that the environment changes according to the schedule of the time information management table. However, this embodiment is also applicable to the time shift of the environmental change. It is made to be able to cope.
This speech recognition device includes a speech input unit 1, an acoustic analysis unit 2, a speech recognition unit 3, a recognition dictionary switch 4, a recognition dictionary unit 5, a noise countermeasure unit 19, and a temporary event selection unit 20.

The environmental changes include noise caused by the occurrence of observable "events" such as a train arriving at the station premises and a bell ringing as the train arrives and departs, and a crowd generated by the crowd. There is a "situation" noise, such as a sound or a voice of a bird singing in the early morning, which cannot be clearly understood by the concept of an event because the causal relationship between the noise and the cause of the noise is not clear. Until the eighth embodiment, a change in the environment is captured in the frame of time, and noise measures including the former and the latter are taken. However, this method cannot cope with a case where an event does not occur at a predetermined time.

Therefore, the noise countermeasure unit 19 of this embodiment is
The time information management table 4 (193) holds information in which the time and the event occurring at that time are associated with each other. For example, an event B (for example a train arrives) at time t _1, it is stored in time t ₄ (rings, for example a siren) event E is scheduled to occur. The recognition dictionary unit 5 includes recognition dictionaries a, b, c,... Created for each noise and a spare general-purpose recognition dictionary (a general recognition dictionary not specific to a certain noise). In each of the recognition dictionaries, for example, a is a noise including the voice of a bird singing in the early morning, b is noise when an event B occurs, and c is a bell sound and a rush sound of a home that occur after the event B. , D is a noise including footsteps of a person coming and going during the day, and e is generated based on voice data obtained by superimposing a noise when an event E occurs on a learning voice. Further, the noise countermeasure unit 19 of the present embodiment holds information in which the event and the recognition dictionary to be used when the event occurs are stored in the event / dictionary table (194). Here, the recognition dictionaries b and e associated with the event are registered, and the recognition dictionaries a and d corresponding to the situation where the event does not occur are not registered.

Next, the temporary event selection unit 20 has a function of transmitting a current or future event to the noise countermeasure unit 19. For example, when the arrival of the train is likely to be delayed by 10 minutes, the noise countermeasure unit 19 is manually notified that the train arrives 10 minutes after the scheduled time. Alternatively, the current event is observed using a monitoring camera, and the result (whether or not the arriving train is visible or not yet visible) is transmitted to the noise suppression unit 19.

The operation of the noise countermeasure unit 19 will be described with reference to the flowchart of FIG. The current time t is monitored by the timer (191), and the time information management table 4
An event scheduled to occur at the current time is detected from (193), and when this matches the event occurrence information sent from the temporary event selection unit 20 (in the above example, the event B is detected at t = t _1). Event E occurs at t = t ₄ , and t 1t ₁ ,
If the t ₄ when event occurrence is not observed), as in the first embodiment, the time and the time information management Te and the recognition dictionary in correspondence - with Bull 1 (192), one out of all recognition dictionary The recognition dictionary switch 4 is controlled in order to select (S10).
3, S107, S115).

Here, if there is a time lag in the occurrence of an event, for example, the arrival of a train is delayed, the temporary event selection unit 20
The contents of the time information management table 1 are temporarily changed according to the information that the train will not arrive. For example, when t = t ₁ to the event B does not occur, the time information management Te - Bull 1, the recognition dictionary of the recognition dictionary b is supposed at a time t ₁ ~t ₂ used corresponding to the event B Term a
Or, it is rewritten as general purpose (S104). This is because the fact that the event has not occurred means that the situation immediately before that is continued, and a recognition dictionary suitable for this situation is selected. If the immediately preceding one is not a recognition dictionary suitable for the situation, a general-purpose recognition dictionary is selected. Event E at t = t ₄
Does not occur, the terms of the recognition dictionary of the time information management table 1 are rewritten in the same way (S108). In addition,
recognition dictionary c supposed used t ₂ ~t _3, since those in line with situations in which subsequent to the event B, also terms of the recognition dictionary of time when the event B does not occur t ₂ ~t ₃ a Or rewrite it as generic. Further, when the event B has occurred in t ≠ t ₁ is event-Dictionary Te - selecting a recognition dictionary b in question with reference to the table (194) (S110), the time information management Te - Bull 1, The time t ₁ to t ₂ at which the recognition dictionary b corresponding to the event B was to be used is rewritten so as to be shifted by (t−t ₁ ) (S111). t ≠ t ₄ the same even if the event E has occurred in the event, dictionaries tape - to select the recognition dictionary e by Bull (S113), the time information management Te -
The term of time when the recognition dictionary e of the bull 1 was used is rewritten (S114).

As described above, according to the present embodiment, when a change in environment is caused by an event or a situation that can be regarded as a situation is enclosed in a time frame to take measures against noise, the event occurs at a predetermined time. Can respond to no emergencies.

[0055]

As described above, according to the present invention, for a non-stationary environmental noise, the temporal information of the noise is stored in advance and used to select a speech recognition dictionary suitable for the environment. This makes it possible to provide a voice recognition device capable of improving the recognition performance by practically using it and by appropriately removing noise.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a speech recognition device according to a first embodiment.

FIG. 2 is a configuration diagram of a speech recognition device according to a second embodiment.

FIG. 3 is a flowchart showing the operation of a noise suppression unit 6;

FIG. 4 is a configuration diagram of a speech recognition device according to a third embodiment.

FIG. 5 is a configuration diagram of a speech recognition device according to a fourth embodiment.

FIG. 6 is a configuration diagram of a speech recognition device according to a fifth embodiment.

FIG. 7 is a configuration diagram of a speech recognition device according to a sixth embodiment.

FIG. 8 is a configuration diagram of a speech recognition device according to a seventh embodiment.

FIG. 9 is a configuration diagram of a speech recognition device according to an eighth embodiment.

FIG. 10 is a configuration diagram of a speech recognition device according to a ninth embodiment.

FIG. 11 is a flowchart showing the operation of the noise suppression unit 19;

[Explanation of symbols]

 Reference Signs List 1 voice input unit 2 sound analysis unit 3 voice recognition unit 4 recognition dictionary switch 5 recognition dictionary unit 6, 7, 67, 19 noise suppression unit 61, 71, 151, 191 timer 62, 192 time information management table 1 72 Time information management table 2 193 Time information management table 4 194 Event / dictionary table 8 Noise removal unit 9 Noise data unit 10 Noise data switch 11 Recognition dictionary 12 Learning voice data file 13 Learning voice Data processing unit 14 Recognition dictionary creation unit 15 Speaker prediction unit 152 Time information management table 3 16 Vocabulary / speaker correspondence table 17, 18 Environment adaptation unit 20 Temporary event selection unit

──────────────────────────────────────────────────続 き Continuation of the front page (72) Inventor Hirofumi Kanazawa 1st address, Komukai Toshiba-cho, Sachi-ku, Kawasaki-shi, Kanagawa Prefecture Toshiba Research Institute, Inc. (56) References JP-A-59-168496 (JP, A) JP-A-62-103699 (JP, A) JP-A-59-34595 (JP, A) JP-A-59-34596 (JP, A) JP-B-61-27758 (JP, B2) JP-B-63-67197 (JP, A) , B2) (58) Fields surveyed (Int. Cl. ⁷ , DB name) G10L 15/00-17/00

Claims (2)

(57) [Claims] An input means for inputting voice, comprising: a plurality of voice recognition dictionaries; an analysis means for analyzing voice input by the input means; and associating time with a voice recognition dictionary to be used for the time. A storage means for storing the time, and a time detecting means for detecting a time at which the voice is input by the input means; and a voice recognition dictionary corresponding to the time detected by the time detecting means, using the storage means. Selecting means for selecting; a recognizing means for recognizing the input voice by collating the analysis result by the analyzing means with the voice recognition dictionary selected by the selecting means; and a time and an event scheduled to occur at the time Event storage means for storing in advance, a preliminary storage means for storing in advance an event and a speech recognition dictionary to be used when the event occurs, and the time detection means An event detecting means for detecting an event occurring at a time detected by the detecting means, and a preliminary selecting means for selecting a voice recognition dictionary corresponding to the event detected by the event detecting means using the preliminary storing means, An event scheduled to occur at the time detected by the time detecting means is checked from the event storage means. If the scheduled event does not match the event detected by the event detecting means, the selecting means A speech recognition apparatus, wherein the recognition means is operated by switching to the preliminary selection means. 2. An input means for inputting a voice, an analyzing means for analyzing the voice input by the input means, and collating the analysis result by the analyzing means with a voice recognition dictionary to recognize the input voice. A first storage unit that stores in advance a vocabulary in the voice recognition dictionary and a utterance set that utters the vocabulary, and a time in advance, Second storage means for storing a utterance set having the highest probability of inputting a voice; time detecting means for detecting a time at which the voice is input by the input means; A search unit for searching a corresponding set of speakers from the second storage unit, wherein when the recognition unit compares the analysis result with the speech recognition dictionary, the first storage unit The speech recognition apparatus characterized in that the utterance set stored in the utterance set is recognized with a higher priority than the vocabulary that does not match the vocabulary that matches the utterance set searched by the search means.