JP2005249874A - Device, method, and program for speech recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognition device capable of realizing speech recognition with higher precision than a conventional device. <P>SOLUTION: The speech recognition device 10 is equipped with a signal state classifying means 11 of analyzing a feature of an input signal and classifying the signal state of the input signal into one of a plurality of predetermined signal states, a state generation frequency calculating means 12 of calculating the generation frequency of a signal state, a sound model storage memory 13 which stores a plurality of sound models and data by the plurality of sound models used when the plurality of sound models are created, a sound model mixing rate calculating means 14 of calculating the mixing rate of data according to the generation frequency, a sound model composing means 15 of composing a sound model by mixing data stored in the sound model storage memory 13 at the mixing rate, and a speech recognition means 16 of performs speech recognition processing for the input signal by using the sound model composed by the sound model composing means 15. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a voice recognition device, a voice recognition method, and a voice recognition program for recognizing voice.

  Conventionally, speech recognition is realized by matching processing using an acoustic model close to the characteristics of an input signal. Since the characteristics of the input signal change depending on the speaker environment, utterance style, and the like when the speech is recognized, it is necessary to recognize the speech according to these changes.

  For example, when attention is paid to the noise in the speaker's environment, in an environment where noise exists, a large number of acoustic models generated using voice data close to the type of noise and the SN ratio are prepared and matching processing is performed. Can perform high-performance speech recognition processing.

  In addition, the speaker's utterance style is different between when the speaker speaks quickly and when speaking slowly, and when the speaker reads out the newspaper and when speaking in the usual spoken language. Even if the speaker speaks the same content, there are individual differences in the utterance style depending on the gender, age, and loudness of the speaker. On the other hand, even if the same speaker is speaking on the phone or speaking in a room with reverberation such as a bathroom, the speaking style of the speaker is different. Since the speaker's speech characteristics will be different if the speaker's utterance style is different, an acoustic model generated using speech data close to the speaker's utterance style to achieve high recognition performance It is necessary to prepare a large number of matching processes.

  As described above, in the field of speech recognition, robust speech recognition that can cope with various changes in the speaker environment, utterance style, etc. (hereinafter referred to as “environment etc.”) during speech recognition processing is required. Various proposals have been made for this requirement.

For example, the on-vehicle speech recognition device disclosed in Non-Patent Document 1 uses running noise and idling noise as in-vehicle noise superimposed on learning speech, and uses an acoustic model created with speech data including these two noises. The voice uttered by the passenger can be recognized.
Hiroaki Ogikubo, Akio Amano, Nobuo Hataoka “Noise Countermeasures and Evaluation in Vehicle Voice Recognition” IEICE Transactions, D-II Vol.J83-D-II, No.11, pp.2190-2197, 2000 November

  However, in such a conventional speech recognition apparatus, if speech recognition processing is to be performed with high accuracy in an actual environment where the speech recognition environment changes in various ways, the speech data of each assumed environment or the like is used. It is necessary to prepare an enormous acoustic model that has been created.

  Therefore, in the conventional speech recognition device, it is difficult to prepare an acoustic model that can follow changes in various environments due to restrictions on the capacity of the memory that can be installed in the speech recognition device and the time required for speech recognition processing. There has been a problem that the accuracy of voice recognition is lowered due to changes in the environment and the like.

  The present invention has been made in order to solve such a problem, and a speech recognition apparatus capable of realizing speech recognition with higher accuracy than conventional apparatuses even when the environment for speech recognition processing changes. A speech recognition method and a speech recognition program are provided.

  The speech recognition apparatus according to the present invention is characterized by analyzing characteristics of an input signal and classifying the signal state of the input signal into one of a plurality of predetermined signal states, and generation of the signal state A state occurrence frequency calculating means for calculating a frequency, and an acoustic model for calculating a mixing ratio based on the occurrence frequency for mixing data used for each of the acoustic models when a plurality of predetermined acoustic models are created Mixing ratio calculating means, acoustic model synthesizing means for newly synthesizing an acoustic model by mixing the data at the mixing ratio, and speech recognition for performing speech recognition processing on the input signal using the synthesized acoustic model And means.

  With this configuration, in the speech recognition apparatus of the present invention, the acoustic model mixing ratio calculation unit calculates the mixing ratio for mixing the data used for each acoustic model based on the occurrence frequency of the signal state, and the speech recognition unit Since the speech recognition process is performed on the input signal using the acoustic model learned from the data mixed at the mixing ratio suitable for the environment at that time, even if the environment for speech recognition processing changes, the accuracy is higher than that of conventional devices. Voice recognition can be realized.

  Further, in the speech recognition apparatus of the present invention, the acoustic model mixture ratio calculation unit is configured to use the mixing ratio of the data used for each of the acoustic models corresponding to the plurality of signal states classified by the signal state classification unit. Is increased as the frequency of occurrence increases.

  With this configuration, the speech recognition apparatus of the present invention is adapted to the signal state of the input signal because the acoustic model mixture ratio calculation means increases the mixture ratio of the data used for each acoustic model as the frequency of occurrence increases. Voice recognition processing can be realized.

  Furthermore, the speech recognition apparatus of the present invention has a configuration characterized in that the acoustic model mixture ratio calculation means calculates the mixture ratio based on a speech recognition result of the speech recognition means.

  With this configuration, in the speech recognition apparatus according to the present invention, the acoustic model mixture ratio calculation unit calculates the data mixture ratio based on the speech recognition result of the speech recognition unit. It can be reset to increase the accuracy of voice recognition.

  Furthermore, in the speech recognition apparatus of the present invention, when the acoustic model mixture ratio calculation means has a score representing the probability of the speech recognition result as a numerical value is equal to or less than a predetermined value, the mixing is performed based on the change tendency of the occurrence frequency. It has the structure characterized by calculating a ratio.

  With this configuration, in the speech recognition apparatus of the present invention, the acoustic model mixture ratio calculation means calculates the mixture ratio based on the change tendency of the occurrence frequency when the score of the speech recognition result is a predetermined value or less. The mixing ratio adapted to the signal state and the score of the speech recognition result can be calculated, and speech recognition can be realized with higher accuracy than conventional devices.

  Furthermore, the speech recognition apparatus of the present invention has a configuration characterized by comprising acoustic model mixture ratio input means for inputting the mixture ratio.

  With this configuration, in the speech recognition apparatus of the present invention, since the acoustic model mixture ratio input means inputs the mixture ratio, the acoustic model mixture ratio in the initial state or the like can be easily set.

  Furthermore, the speech recognition apparatus of the present invention has a configuration characterized by comprising test data generating means for generating test data for setting the mixing ratio.

  With this configuration, since the test data generating means generates test data for setting the mixing ratio, the voice recognition device of the present invention can grasp the relationship between the signal state and the score of the voice recognition result, and the voice recognition It is possible to improve the performance.

  Furthermore, in the speech recognition apparatus of the present invention, the data includes at least one of time waveform data of voice data, a mel cepstrum coefficient, a delta mel cepstrum coefficient, and a delta logarithmic power coefficient when the acoustic model is created. It has the structure characterized by this.

  With this configuration, in the speech recognition apparatus of the present invention, the acoustic model storage memory stores the time waveform data, mel cepstrum coefficient, etc. of the speech data when creating the acoustic model, so these data are mixed at a mixing ratio. By synthesizing the acoustic model, speech recognition can be realized with higher accuracy than conventional devices even if the environment for speech recognition processing changes.

  Furthermore, in the speech recognition apparatus of the present invention, the acoustic model has a hidden Markov model structure, and the acoustic model storage means includes data of transition probability and output probability defined by the hidden Markov model. As a storage feature.

  With this configuration, since the acoustic model has a hidden Markov model structure, the speech recognition environment of the present invention changes the environment for speech recognition processing, etc. by calculating the mixing ratio of transition probability or output probability data. Even so, speech recognition can be realized with higher accuracy than conventional devices.

  The speech recognition method of the present invention analyzes the characteristics of an input signal, classifies the signal state of the input signal into one of a plurality of predetermined signal states, and then creates a predetermined acoustic model. The mixing ratio for mixing the data used in the calculation is calculated based on the occurrence frequency of the signal state, the data is mixed at the mixing ratio to synthesize the acoustic model, and the synthesized acoustic model is used. A voice recognition process is performed on the input signal.

  By this method, the mixing ratio for mixing the data used for each acoustic model is calculated based on the occurrence frequency of the signal state, and the speech recognition process is performed on the input signal by the acoustic model learned from the data mixed at the mixing ratio. Therefore, voice recognition can be realized with higher accuracy than conventional devices even if the environment or the like for voice recognition processing changes.

  The speech recognition program according to the present invention includes a step of analyzing a feature of an input signal, a step of classifying the signal state of the input signal into one of a plurality of predetermined signal states, and a predetermined acoustic model Calculating a mixing ratio based on the frequency of occurrence of the signal state, mixing the data at the mixing ratio, and synthesizing the acoustic model. Performing a voice recognition process on the input signal using the acoustic model.

  With this program, it is possible to cause a computer to execute voice recognition processing with higher accuracy than a conventional program.

  The present invention can provide a speech recognition device having an effect that speech recognition can be realized with higher accuracy than conventional devices.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(First embodiment)
First, the configuration of the speech recognition apparatus according to the first embodiment of the present invention will be described.

  As shown in FIG. 1, the speech recognition apparatus 10 according to the present embodiment analyzes a feature of an input signal and classifies the signal state of the input signal into one of a plurality of predetermined signal states. Classifying means 11, state occurrence frequency calculating means 12 for calculating the occurrence frequency of signal states classified by signal state classifying means 11, and a plurality of predetermined acoustic models and a plurality of acoustic models are used. An acoustic model storage memory 13 for storing data for each of the plurality of acoustic models, an acoustic model mixture ratio calculation means 14 for calculating a data mixture ratio based on the occurrence frequency calculated by the state occurrence frequency calculation means 12, The acoustic model synthesis means 15 for synthesizing the acoustic model by mixing the data stored in the acoustic model storage memory 13 at the mixing ratio, and the acoustic model synthesis means 15 Use made acoustic model and a speech recognition unit 16 for performing speech recognition processing on the input signal.

  Here, the signal state classified by the signal state classification means 11 refers to a state classified by characteristics such as noise type, SN ratio, speaker gender, age, and speaking speed.

  The signal state classification means 11 analyzes the characteristics of the input signal and classifies it into one of signal states based on, for example, a preset S / N ratio of the input signal.

  More specifically, the signal states to be classified by the signal state classification unit 11 include, for example, “signal state A whose SN ratio is 25 dB or more”, “signal state B whose SN ratio is 15 or more and less than 25 dB” and “SN ratio”. When the three signal states C of less than 15 dB are preset, the signal state classification unit 11 calculates the signal-to-noise ratio of the input signal for each predetermined time frame, thereby changing the signal state of the input signal to the time frame. Each is classified into one of signal states A, B and C.

  In the following description, the speech recognition apparatus 10 of the present embodiment will be described as performing speech recognition on the input signals in the signal states A, B, and C described above.

  The state occurrence frequency calculation means 12 calculates a time ratio in which the input signal is in the signal states A, B, and C within a predetermined time (hereinafter referred to as “occurrence frequency calculation time”). The frequency of occurrence is calculated.

  Specifically, for example, when the occurrence frequency calculation time is 100 seconds, the signal state A with an SN ratio of 25 dB or more, the signal state B with an SN ratio of 15 or more and less than 25 dB, and the signal state C with an SN ratio of less than 15 dB The state occurrence frequency calculation means 12 calculates the occurrence frequencies of the signal state A, the signal state B, and the signal state C, respectively, at 30%, 60%, and 10 seconds, respectively. % Is calculated.

  The occurrence frequency calculation time is not limited to the above-mentioned 100 seconds, and may be 20 seconds or 10 minutes. Further, the state occurrence frequency calculation unit 12 may be configured to calculate the occurrence frequency by setting the occurrence frequency calculation time according to, for example, a temporal change in the SN ratio. With this configuration, the state occurrence frequency calculating means 12 calculates the occurrence frequency every 5 seconds when the SN ratio changes in a short time, and every 20 minutes when the SN ratio hardly changes in time, for example. Since the occurrence frequency is calculated, resources such as a CPU and a memory can be used effectively.

  As shown in FIG. 2, the acoustic model storage memory 13 includes, for example, an acoustic model A learned with an SN ratio of 30 dB, an acoustic model B learned with an SN ratio of 20 dB, an acoustic model C learned with an SN ratio of 10 dB, an acoustic model A, The audio data A, B, and C used for learning when creating B and C are stored.

  Specifically, the acoustic model storage memory 13 is recorded in a sound model A created by learning a plurality of sound data A and sound data A recorded in an environment with an SN ratio of 30 dB, and an environment with an SN ratio of 20 dB. A plurality of voice data B and an acoustic model B created by learning with the voice data B, and an acoustic model C created by learning with a plurality of voice data C and voice data C recorded in an environment with an SN ratio of 10 dB And. Note that the acoustic models A, B, and C shown in FIG. 2 include audio data A, B, and C, respectively.

  The acoustic model mixture ratio calculation means 14 has a sound data mixture ratio corresponding to each acoustic model when the acoustic models A, B, and C are synthesized as the occurrence frequency calculated by the state occurrence frequency calculation means 12 increases. Calculations are made to increase.

  For example, as shown in FIG. 2, when the occurrence frequencies of the signal states A, B, and C calculated by the state occurrence frequency calculation unit 12 are 30%, 60%, and 10%, respectively, the acoustic model mixture ratio calculation unit 14 Is designed to calculate the respective mixing ratios as 0.3, 0.6 and 0.1.

  The acoustic model synthesizing unit 15 uses the mixing ratio calculated by the acoustic model mixture ratio calculating unit 14 as a weighting factor, mixes the audio data A, B, and C stored in the acoustic model storage memory 13 and creates a new acoustic model. It comes to synthesize. The acoustic model synthesized by the acoustic model synthesis unit 15 is hereinafter referred to as “synthetic acoustic model”.

  Specifically, as shown in FIG. 2, the acoustic model synthesizing unit 15 obtains weight coefficients 0.3, 0.6, and 0.1 from data A, B, and C each composed of a plurality of data. A synthetic acoustic model is generated by learning using each data at a rate of.

  The voice recognition means 16 performs voice recognition processing on the input signal using the synthetic acoustic model, and outputs the result of the voice recognition processing.

  Next, the operation of the speech recognition apparatus 10 according to the present embodiment will be described. A case where speech recognition is performed on the input signals in the signal states A, B, and C will be described as an example.

  First, the characteristics of the input signal are analyzed by the signal state classification unit 11 and the signal state of the input signal is classified into one of the signal states A, B, and C.

  Next, the occurrence frequency of the signal states A, B, and C within the occurrence frequency calculation time is calculated by the state occurrence frequency calculation means 12.

  Further, the acoustic model mixture ratio calculation means 14 calculates a mixture ratio for mixing the audio data A, B, and C stored in the acoustic model storage memory 13 based on the occurrence frequency.

  Subsequently, the acoustic model synthesizing unit 15 mixes the audio data A, B, and C stored in the acoustic model storage memory 13 using the mixing ratio as a weighting factor, thereby generating a synthetic acoustic model.

  Then, the speech recognition unit 16 performs speech recognition on the input signal using the synthetic acoustic model, and outputs a speech recognition result.

  In this embodiment, the signal states are classified based on the SN ratio. However, the present invention is not limited to this, and is classified according to characteristics such as noise type, speaker gender, age, and speaking speed. May be.

  For example, when the speech recognition device 10 of the present embodiment is applied to a portable voice input type car navigation device, the signal state classification means 11 is set so as to classify the signal state of the input signal according to the type of noise, and the home If you prepare an acoustic model and its data learned from the noise inside the vehicle and the vehicle, you can use it even when you move the car navigation device from the home to the car and continue to use it after setting the route with voice in the home Accurate speech recognition corresponding to the type of noise can be performed without depending on changes in the environment or the like.

  Moreover, the speech recognition apparatus 10 of this Embodiment is comprised, for example with a computer. In the case of this configuration, the computer reads out a program created to operate as the voice recognition device 10 from a storage medium or receives it via a network and executes it, so that the voice can be obtained with higher accuracy than a conventional device. Recognition can be realized.

  As described above, according to the speech recognition apparatus 10 of the present embodiment, the acoustic model synthesis unit 15 mixes the speech data A, B, and C stored in the acoustic model storage memory 13 with a mixing ratio according to the signal state. Since the synthesized acoustic model is generated by mixing, the speech recognition means 16 is configured to perform speech recognition processing on the input signal using the synthesized acoustic model. Speech recognition can be realized with higher accuracy than the apparatus.

  In the present embodiment, data for creating an acoustic model has been described as speech data. However, the present invention is not limited to this, and time waveform data, mel cepstrum of speech data for creating an acoustic model is not limited to this. Similar effects can be obtained with data such as a coefficient, a delta-mel cepstrum coefficient, and a delta logarithmic power coefficient.

  Further, when the acoustic model in the present embodiment has a hidden Markov model structure, a synthetic acoustic model is generated by using a weighting factor for the transition probability or output probability data of the acoustic models A, B, and C described above. Even if it comprises, the same effect is acquired.

(Second Embodiment)
First, the configuration of the speech recognition apparatus according to the second embodiment of the present invention will be described. However, the same components as those of the speech recognition apparatus 10 according to the first embodiment of the present invention are denoted by the same reference numerals and description thereof is omitted.

  As shown in FIG. 3, the speech recognition apparatus 20 according to the present embodiment includes an acoustic model mixture ratio calculation unit that calculates a mixture ratio based on the calculation result of the state occurrence frequency calculation unit 12 and the speech recognition result of the speech recognition unit 16. 14 is provided.

  Next, the operation of the speech recognition apparatus 20 according to the present embodiment will be described. However, description of operations similar to those of the speech recognition apparatus 10 according to the first exemplary embodiment of the present invention is omitted.

  The acoustic model mixture ratio calculation means 14 inputs the occurrence frequency calculated by the state occurrence frequency calculation means 12 and the speech recognition result output by the speech recognition means 16.

  Here, when the score representing the probability of the speech recognition result as a numerical value is equal to or greater than a predetermined value, the acoustic model mixture ratio calculation means 14 determines that the calculation result of the mixture ratio is adapted to the signal state of the input signal. Then, as described in the first embodiment of the present invention, the mixing ratio is calculated based on the calculation result of the state occurrence frequency calculation means 12.

  On the other hand, when the score of the speech recognition result is less than the predetermined value, the acoustic model mixture ratio calculation means 14 determines that the calculation result of the mixture ratio is not adapted to the signal state of the input signal, and the occurrence frequency tends to change. Based on the mixing ratio.

  For example, in a scene where the signal state of the input signal changes every moment, the calculation result of the mixing ratio may not be adapted to the signal state of the input signal, and the score of the speech recognition result may be lowered. Therefore, the acoustic model mixture ratio calculation means 14 calculates the mixture ratio based on the temporal change tendency of the occurrence frequency of the signal state so that the score of the speech recognition result becomes a predetermined value or more.

  Specifically, the acoustic model mixture ratio calculation means 14 has the occurrence frequencies of the signal states A, B, and C at a certain time of 10%, 50%, and 40%, respectively, and the time of the signal state C. Suppose that there is the largest change tendency. When the score of the voice recognition result at this time is lower than a predetermined value, the acoustic model mixture ratio calculation means 14 temporarily sets the occurrence frequencies of the signal states A, B, and C to 10%, 40%, and 50%, respectively. Calculate the mixing ratio. That is, the acoustic model mixture ratio calculation means 14 calculates a mixture ratio obtained by raising the frequency of occurrence of the signal state C having the largest temporal change rate from the actual value so as to increase the score of the speech recognition result for the input signal. Operate.

  Note that the speech recognition apparatus 20 according to the present embodiment can also have a configuration in which the acoustic model mixture ratio calculation unit 14 learns the relationship between the occurrence frequency and temporal change of the signal state and the score of the speech recognition result. With this configuration, the speech recognition apparatus 20 according to the present exemplary embodiment allows the speech recognition result score to be set to a predetermined value in a short time even if the speech recognition environment suddenly changes and the speech recognition result score temporarily decreases. It can return to the above.

  As described above, according to the speech recognition apparatus 20 of the present embodiment, the acoustic model mixture ratio calculation unit 14 is based on the calculation result of the state occurrence frequency calculation unit 12 and the speech recognition result of the speech recognition unit 16. Therefore, the mixing ratio adapted to the signal state of the input signal and the score of the speech recognition result can be calculated, and speech recognition can be realized with higher accuracy than the conventional device.

  In the present embodiment, the calculation of the mixing ratio based on the change tendency of the occurrence frequency of the signal state and the score of the speech recognition result has been described. However, the present invention is not limited to this, and for example, the signal state Even when the score of the speech recognition result is less than a predetermined value in spite of the environment in which the sound is almost constant, the acoustic model mixture ratio calculation means 14 calculates the mixture ratio with reference to the score of the speech recognition result. Thus, speech recognition performance can be improved.

(Third embodiment)
First, the configuration of the speech recognition apparatus according to the third embodiment of the present invention will be described. However, the same components as those of the speech recognition apparatus 10 according to the first embodiment of the present invention are denoted by the same reference numerals and description thereof is omitted.

  As shown in FIG. 4, the speech recognition apparatus 30 according to the present embodiment includes an acoustic model mixture ratio input unit 31 that inputs an acoustic model mixture ratio.

  The acoustic model mixture ratio input means 31 includes, for example, a switch as shown in FIG. 5A and a memory (not shown) that stores a table as shown in FIG. 5B, and a switch lever S1. The mixing ratio is input by a switch pattern based on a combination of the positions of S3.

  For example, when the acoustic model mixture ratio input means 31 inputs the mixture ratio with the switch pattern “1, 1, 0” shown in FIG. 5A, the acoustic model mixture ratio calculation means 14 displays the result shown in FIG. Based on the table shown in Fig. 4, the mixing ratio of the audio data A, B, and C corresponding to each acoustic model is set to 0.4, 0.4, and 0.2, respectively.

  Next, the operation of the speech recognition apparatus 30 according to the present embodiment will be described. However, description of operations similar to those of the speech recognition apparatus 10 according to the first exemplary embodiment of the present invention is omitted.

  The switch pattern information set by the switch is read by the acoustic model mixture ratio input means 31, and the switch pattern information and table information are output to the acoustic model mixture ratio calculation means 14.

  Then, the acoustic model mixture ratio calculation means 14 sets the mixture ratio of the audio data corresponding to each acoustic model based on the switch pattern information and the table information.

  As described above, according to the speech recognition apparatus 30 of the present embodiment, the acoustic model mixture ratio input unit 31 is configured to input the acoustic model mixture ratio based on the switch pattern information set by the switch. The acoustic model mixture ratio can be easily set when the signal state can be predicted or in the initial state after power-on and reset.

  In the present embodiment, an example in which the acoustic model mixture ratio input means 31 is configured by a switch and a memory has been described. However, the present invention is not limited to this, for example, a keyboard or a pointing device, and a display The same effect can be obtained when the acoustic model mixture ratio input means 31 is configured with the above and the acoustic model mixture ratio is input on the input screen displayed on the display.

(Fourth embodiment)
First, the configuration of the speech recognition apparatus according to the fourth embodiment of the present invention will be described. However, the same components as those of the speech recognition apparatus 20 according to the second embodiment of the present invention are denoted by the same reference numerals and description thereof is omitted.

  As shown in FIG. 6, the speech recognition apparatus 40 of the present embodiment includes test data generation means 41 that generates test data that is used by switching to an input signal that is a target of speech recognition.

  The test data generating means 41 generates various signals assumed as input signals, for example, signals in which the SN ratio, noise type, etc. are set as test data.

  Next, the operation of the speech recognition apparatus 40 according to this embodiment will be described. However, description of operations similar to those of the speech recognition apparatus 20 according to the second exemplary embodiment of the present invention is omitted.

  First, the test data generating means 41 generates test data with various signal-to-noise ratios, for example. The test data is processed in the same procedure as the voice recognition device 20 described above, and voice recognition processing is performed by the voice recognition means 16.

  Next, the acoustic model mixture ratio calculation unit 14 calculates the acoustic model mixture ratio based on the score of the speech recognition result and the classification result of the signal state of the test data.

  As described above, according to the speech recognition apparatus 40 of the present embodiment, the test data generation means 41 is configured to generate various test data assumed as input signals. The relationship with the score can be grasped, and the speech recognition performance can be improved.

  As described above, the speech recognition device according to the present invention has an effect that speech recognition can be realized with higher accuracy than conventional devices, and a speech recognition device, speech recognition method, and speech recognition that recognize speech. Useful as a program.

The block diagram of the speech recognition apparatus of the 1st Embodiment of this invention Explanatory drawing of the acoustic model synthesis | combination of the speech recognition apparatus of the 1st Embodiment of this invention The block diagram of the speech recognition apparatus of the 2nd Embodiment of this invention The block diagram of the speech recognition apparatus of the 3rd Embodiment of this invention (A) The figure which shows an example of the switch which comprises the acoustic model mixing ratio input means of the 3rd Embodiment of this invention (b) The table which set the acoustic model mixing ratio concerning the 3rd Embodiment of this invention Figure showing an example The block diagram of the speech recognition apparatus of the 4th Embodiment of this invention

Explanation of symbols

DESCRIPTION OF SYMBOLS 10, 20, 30, 40 Speech recognition apparatus 11 Signal state classification | category means 12 State generation frequency calculation means 13 Acoustic model storage memory 14 Acoustic model mixing ratio calculation means 15 Acoustic model synthesis means 16 Speech recognition means 31 Acoustic model mixing ratio input means 41 Test data generation means

Claims (10)

Analyzing the characteristics of the input signal, classifying the signal state of the input signal into one of a plurality of predetermined signal states, and calculating the state occurrence frequency for calculating the occurrence frequency of the signal state Means, and an acoustic model mixture ratio calculating means for calculating a mixing ratio for mixing data used for each acoustic model when a plurality of predetermined acoustic models are created, based on the occurrence frequency, and the data An acoustic model synthesizing unit that newly synthesizes an acoustic model by mixing at a mixing ratio, and a speech recognition unit that performs speech recognition processing on the input signal using the synthesized acoustic model. Voice recognition device. The acoustic model mixture ratio calculation means increases the mixing ratio of the data used for each of the acoustic models corresponding to the plurality of signal states classified by the signal state classification means as the generation frequency increases. The speech recognition apparatus according to claim 1. The speech recognition apparatus according to claim 1, wherein the acoustic model mixture ratio calculation unit calculates the mixture ratio based on a speech recognition result of the speech recognition unit. The acoustic model mixture ratio calculation means calculates the mixture ratio based on a change tendency of the occurrence frequency when a score representing the probability of the speech recognition result by a numerical value is a predetermined value or less. The speech recognition device according to any one of claims 1 to 3. The speech recognition apparatus according to claim 1, further comprising an acoustic model mixture ratio input unit that inputs the mixture ratio. 6. The speech recognition apparatus according to claim 1, further comprising test data generating means for generating test data for setting the mixing ratio. 2. The data according to claim 1, wherein the data includes at least one of time waveform data, mel cepstrum coefficient, delta mel cepstrum coefficient, and delta logarithmic power coefficient of voice data when the acoustic model is created. Item 7. The speech recognition device according to any one of Items 6 to 6. The acoustic model has a hidden Markov model structure, and the acoustic model storage means stores the acoustic model as transition probability and output probability data defined by the hidden Markov model. The speech recognition device according to any one of claims 1 to 7. Analyzing the characteristics of the input signal, classifying the signal state of the input signal into one of a plurality of predetermined signal states, and then mixing the data used when a predetermined acoustic model is created A mixing ratio is calculated based on the frequency of occurrence of the signal state, the data is mixed at the mixing ratio to synthesize the acoustic model, and speech recognition processing is performed on the input signal using the synthesized acoustic model A speech recognition method characterized by performing. Analyzing the characteristics of the input signal; classifying the signal state of the input signal into one of a plurality of predetermined signal states; and data used when a predetermined acoustic model is created Calculating a mixing ratio based on the frequency of occurrence of the signal state, mixing the data at the mixing ratio to synthesize the acoustic model, and using the synthesized acoustic model Performing a voice recognition process on an input signal.

Images