JP4919282B2 - Unclear voice command recognition device and unclear voice command recognition processing method - Google Patents

Description

  The present invention can be used as a voice command even for voices that are unclearly spoken, such as voices of elderly people and persons with disabilities, or voices in a noisy environment, or voices that vary greatly from voice to voice. The present invention relates to an indistinct voice command recognition apparatus and an indistinct voice command recognition processing method that can be operated.

  Conventionally, techniques for recognizing voice commands have been researched and developed in order to operate devices using voice. For example, Patent Document 1 describes a voice recognition processing technique for operating a video recording device of home appliances using a word spotting voice recognition technique, and Patent Document 2 operates an in-vehicle information device such as a car navigation system. For example, Patent Document 3 describes a voice recognition processing technique for operating an electric wheelchair by voice instead of a joystick. In these speech recognition processing techniques, a recognition engine that recognizes speech from speech signals uses existing speech recognition processing.

  Currently, the most commonly used technique for recognizing a voice command is to recognize a voice from a voice signal by using an acoustic model and a description grammar described by a Hidden Markov Model (HMM). Among the commands registered in the dictionary, the command with the highest likelihood is output as the recognition result in light of the acoustic model and the description grammar.

  Usually, a dictionary of voice commands for operating a device using voice describes a phoneme string representing the standard utterance of each voice command, but for a speaker with a large deviation from the standard utterance In some cases, the HMM of a recognition engine that recognizes speech from a speech signal cannot recognize the deviation and recognizes it as a phoneme sequence significantly different from a standard phoneme sequence.

  In order to adapt to a speaker whose deviation from the standard utterance is large, there is also known a technique for improving the recognition accuracy by registering a plurality of phoneme sequences unique to the speaker.

  In addition, in the case of a speaker for which stable utterance is difficult for some reason, fluctuation for each utterance becomes large, and it may not be sufficient to register one phoneme string for each command.

  For example, recognition accuracy can be improved by registering phoneme strings extracted from a plurality of voice command samples in a dictionary for a speaker who is difficult to speak stably due to a disorder such as cerebral palsy.

  The phoneme strings to be registered can be obtained by manual analysis, but the recognition accuracy can also be improved by registering multiple phoneme strings automatically obtained using a continuous phoneme recognition engine such as a phoneme typewriter. Can be made. The continuous phoneme recognition engine is realized by the same technology as a normal continuous word recognition engine in which words are replaced with phonemes.

The following documents can be referred to as conventional documents related to this type of voice command recognition.
JP 2002-269146 A Japanese translation of PCT publication No. 2003-202897 JP 2000-84004 A Japanese Patent Laid-Open No. 2002-221984 Shikano et al .: Speech recognition system, Ohmsha, 2001. N. Christianiani and J.M. Shawe-Taylor: An Induction to Support Vector Machines, Cambridge University Press, 2002. Shaw-Taylor: An Introduction to Support Vector Machines, Cambridge University Press, 2002. H. Lodhi et al. : Text classification using string kernels, Advances in Neural Information Processing Systems 13, 2001. C. Cortes et al. : Rational kernels: theory and algorithms, Journal of Machine Learning Research, 5, 2004. C. Saunders et al. : Syllables and other string kernel extensions, Proceedings of International Conference on Machine Learning, 2002. T.A. Jebara et al. Batterchalya and Expected Likelihood Kernels, Proceedings of Computational Learning Theory, 2003. K. Fukunaga: Statistical pattern recognition (2nd ed.), Academic Press, 1990.

  By the way, in the case of the cerebral palsy patient mentioned above, stable voicing is difficult, so the number of commands to be recognized increases, or voices other than commands are uttered while using voice commands. If it is allowed and it is necessary to distinguish the voice command from the voice other than the command, the recognition accuracy is lowered.

  In this way, it is difficult for disabled people to make stable utterances, and the uttered voices are frequently inserted with unnecessary sounds or become partially different every time they utter, and different voices are This is because it is difficult to identify a voice command by accidentally being recognized as a partially similar phoneme string.

  A similar problem is not limited to the voice of a disabled person, but may also occur when recognizing voice in an elderly person or an environment with sudden noise.

  The present invention has been made in order to overcome such problems, and the object of the present invention is to provide unclear voices such as voices of elderly people and persons with disabilities, or voices in noisy environments. Alternatively, an object of the present invention is to provide an unclear voice command recognition device and an unclear voice command recognition processing method that can operate a device using a voice command even if the voice has a large fluctuation for each utterance.

  In order to achieve the above object, the present invention provides, as a first aspect, an unclear voice for causing an unclear voice command recognition apparatus according to the present invention to operate a device using an unclearly spoken voice as a voice command. A voice recognition means (100) for converting a voice signal of a voice command into an electrical signal and inputting it, and an analog / digital conversion means for using the electrical signal of the voice signal inputted by the voice input means as digital data. (101), feature extraction means (103) for extracting feature vectors from digital data of speech signals by cepstrum analysis, and recognizing speech subword units from the time series of feature vectors and outputting a subword sequence or graph Subword recognition means (106) and a subword unit column or graph for command identification Database update means (110) for generating training data and registering it in the database, support vector machine learning means (111) for configuring a command classifier by a support vector machine based on the training data, and the support vector machine learning means Voice identification means (108) for identifying a voice command by data processing by a configured command discriminator, and control signal output means for generating a control signal by the voice command identified by the voice identification means and sending it to a control target device (109).

  In this case, in the ambiguous voice command recognition apparatus, the subword recognition means recognizes a subword unit based on any unit of a phoneme, a syllable, or a phoneme of a voice signal. The support vector machine learning unit and the speech identification unit use a character string kernel for a subword unit column and a relational kernel for a subword unit graph. Furthermore, the voice input means is a microphone array in which a plurality of microphones are arranged, and a sound source separation means for separating a plurality of sound sources from digital data of a plurality of sound signals obtained from the microphone array and removing directivity noise ( 102), feature correction means (104) for removing stationary noise from the speech signal, and non-speech discrimination means (105) for distinguishing speech from non-speech by fundamental frequency estimation.

  In addition, as a second aspect of the present invention, the ambiguous voice command recognition processing method according to the present invention recognizes and processes an ambiguous voice command for operating a device by using an unclearly spoken voice as a voice command. An ambiguous voice command recognition processing method, in which a voice input step of converting a voice signal of a voice command into an electric signal by computer processing and inputting the electric signal, and an electric signal of the voice signal input by the voice input means are converted into digital data An analog-to-digital conversion step, a feature extraction step for extracting a feature vector from digital data of a speech signal by cepstrum analysis, and recognizing a speech subword unit from the time series of the feature vector and outputting a subword sequence or graph Subword recognition step and subword sequence or graph A database update step for generating training data for command identification and registering it in a database, a support vector learning step for configuring a command identifier by a support vector machine based on the training data, and the support vector machine learning means A voice identification step for identifying a voice command by data processing by the command classifier and a control signal output step for generating a control signal by the voice command identified by the voice identification means and sending it to the control target device are executed. It is characterized by doing.

  In this case, in the ambiguous voice command recognition processing method, the subword recognition step recognizes a subword unit based on any one of a phoneme, a syllable, or a phoneme of a voice signal. In the support vector machine learning step and the speech identification step, a character string kernel is used for a subword unit, and a relational kernel is used for a subword unit graph. Further, the voice input step inputs a plurality of voice signals from a microphone array in which a plurality of microphones are arranged, separates a plurality of sound sources from digital data of the plurality of voice signals obtained from the microphone array, and has directivity. A sound source separation step for removing noise, a feature correction step for removing stationary noise from the speech signal, and a non-speech identification step for distinguishing speech from non-speech by estimating the fundamental frequency are configured to be executed. .

  Further, as a third aspect of the present invention, an unclear voice command recognition processing program according to the present invention executes an unclear voice command recognition process for operating a device using an unclearly spoken voice as a voice command. A voice input means for converting a voice signal of a voice command into an electric signal and inputting the signal to the computer, and an analog-digital conversion means for using the electric signal of the voice signal input by the voice input means as digital data, Feature extraction means for extracting feature vectors from digital data of speech signals by cepstrum analysis, subword recognition means for recognizing speech subword units from the time series of feature vectors and outputting subword unit columns or graphs, subword unit units Generate training data for command identification with columns or graphs Database update means to be registered in the database, support vector machine learning means for configuring a command classifier by a support vector machine based on the training data, and voice by data processing by a command classifier constituted by the support vector machine learning means A voice identifying means for identifying a command and a control signal output means for generating a control signal based on the voice command identified by the voice identifying means and transmitting the control signal to a controlled device are characterized.

  In this case, in the ambiguous voice command recognition program, the subword recognition means recognizes a subword unit based on a unit of a phoneme, a syllable, or a phoneme of a voice signal. The support vector machine learning unit and the speech identification unit use a character string kernel for a subword unit column and a relational kernel for a subword unit graph. Furthermore, the voice input means inputs a plurality of voice signals from a microphone array in which a plurality of microphones are arranged, and directs the computer to separate a plurality of sound sources from digital data of the plurality of voice signals obtained from the microphone array. Sound source separation means that removes noise, feature correction means that removes stationary noise from the speech signal, and non-speech discrimination means that distinguishes speech and non-speech by fundamental frequency estimation.

  According to the ambiguous voice command recognition apparatus, the ambiguous voice command recognition processing method, and the ambiguous voice command recognition processing program according to the present invention, in the voice recognition processing, the speech is subword such as phoneme, syllable, or phoneme. Recognize as a sequence of units (Patent Document 4), identify a partial sequence that is characteristic of a specific command, remove a partial sequence that corresponds to an unnecessary sound, and perform a detailed and comprehensive analysis of the phoneme sequence, By identifying which command is an input subword unit string, it is possible to recognize unclear voices with large fluctuations for each utterance and voices in sudden noisy environments with high accuracy. It has become.

  In addition, according to the present invention, it becomes possible to accurately recognize an unclear voice command in which unnecessary sounds are inserted and fluctuations for each utterance are large, and a voice command in an environment with sudden noise. Even in the case of people who have not been able to use voice recognition technology until now, or in situations where voice recognition technology has not been available so far, it becomes possible to control equipment using voice. In addition, even if the number of commands is increased or voices other than commands are uttered while using voice commands, the degradation of recognition accuracy can be suppressed to a level lower than that of the conventional technology, thus improving user convenience. Contribute.

  Hereinafter, an embodiment for carrying out the present invention will be described with reference to the drawings. FIG. 1 is a flowchart showing an example of a processing flow of speech recognition processing in an unclear speech command recognition apparatus, an unclear speech command recognition processing method, and an unclear speech command recognition processing program according to the present invention.

  In this voice recognition process, as shown in FIG. 1, a voice input process 100 for inputting voice as an analog electric signal by a microphone, an AD conversion process 101 for digitizing a voice analog electric signal into digital data, and a microphone A sound source separation process 102 that separates a plurality of sound sources from a plurality of sound signals obtained when using an array and removes directional noise; a feature extraction process 103 that performs a cepstrum analysis or the like to obtain a time series of feature vectors; A feature correction process 104 that removes stationary noise from the speech signal, a speech / non-speech discrimination process 105 that distinguishes speech and non-speech using fundamental frequency estimation, etc., and subwords such as phonemes and phonemes from the time series of feature vectors Sub-word recognition process 106 for recognizing units and outputting a sequence of sub-word units; A command sample database update process 110 for registering word units in a database, an SVM learning process 111 for learning a command classifier using a support vector machine (SVM), and a learned classifier is stored in a memory or a hard disk. A discriminator database update process 112, a SVM discrimination process 108 for identifying a command using a discriminator learned in SVM for a subword sequence at the time of command discrimination, and a control target device (electric wheelchair) based on the result Are provided with a plurality of processing modules for performing the respective processes of the control signal output process 109 for outputting a control signal for. By the data processing of these processing modules, an unclear voice command recognition process is performed and a control signal for the control target device is output.

  In order to perform the recognition processing of the unclear voice command here, as will be described later, the acoustic model database 121 storing the acoustic model for performing the subword recognition and the recognition used for command identification and classifier learning are performed. A command sample database 122 that stores a subword unit sequence or a subword unit graph and a classifier database 123 that stores a learned classifier are provided. The data of each database is used in data processing by each processing module, is referred to when performing recognition processing of unclear voice commands, and outputting a control signal to the control target device, and the data of these databases is also used. Update.

  The SVM learning process 111 will be described in more detail. A kernel function that efficiently calculates the similarity of character strings is required to identify a subword unit string, that is, a character string, using the SVM. Since the learning principle / algorithm of SVM is known, a detailed description thereof is omitted here, but Non-Patent Document 2 can be referred to for this.

  Conventionally, as shown in Non-Patent Document 3, a character string kernel is known as a kernel function that gives similarity of character strings. This kernel function calculates the similarity of character strings to p times the product of the lengths of character strings, based on how much any two character strings share any substring (which may be discontinuous). It is a function to calculate with the amount of calculation of the order. By applying this kernel function to SVM, detailed and exhaustive analysis of which subsequences contribute to command identification and which subsequences do not contribute in the subword sequence of voice commands Can do.

  Further, in the speech recognition processing here, the method disclosed in Non-Patent Document 5 is used, so that the phoneme “a” is not “k” and the phoneme “a” is “a” more than the phoneme “k”. It is possible to introduce similarity based on soft matching in units of subwords, such as “:”.

とする。ここで、ｐ_ｉとｑ_ｉは、それぞれサブワード単位ａとサブワード単位ｂとのＨＭＭのｉ番目の出力確率分布を表す。 As a method of giving such similarity in subword units, for example, the similarity between HMMs in subword units is calculated. Subword unit HMM is typically multiple states _s 1, ..., has a _{s n,} the transition state, the output probability distributions _p 1 signal, ..., _{p n} are defined. Therefore, the similarity _{s (p i, q} i) of each output probability distribution as the average of, defining the similarity of HMM. That is, the similarity A (a, b) between the subword unit a and the subword unit b is

And Here, p _i and q _i represent the i-th output probability distribution of the HMM of subword unit a and subword unit b, respectively.

Further, as the similarity of the output probability distribution, the Bhattacharya kernel (Non-patent Document 6) can be used as follows.

と等価であることが知られている。 When the probability distribution p and the probability distribution q are normal distributions, they can be integrated analytically, and the logarithm of the reciprocal is the Bhattacharya distance (Non-patent Document 7),

Is known to be equivalent to

ただし、ここでのｐ_ｉとｑ_ｊは、それぞれ混合正規分布のコンポーネントを表しており、ｍ_ｐ（ｉ）とｍ_ｑ（ｊ）は、その混合比を表す。 However, when the output probability distribution is a mixed normal distribution, an analytical solution is not known, so the following approximate expression is used.

Here, p _i and q _j represent components of a mixed normal distribution, and m _p (i) and m _q (j) represent the mixing ratio.

  Based on the similarity of the subword unit as described above, the similarity of the subword unit column is calculated using the character string kernel, and this is applied to the support vector machine, whereby the subword unit column discriminator is calculated. It is configured by learning, and consequently, a voice command discriminator can be configured by learning. The classifier constituted by these becomes highly discriminative by learning.

  Further, in the subword recognition process 106, not only the subword unit string with the highest likelihood is output, but also when the plurality of recognition candidates with the highest likelihood are output, the above voice command recognition process can be expanded.

  In that case, a graph having phonemes as nodes can be used as a format for expressing a plurality of recognition candidates in a compact manner. For example, a graph in which a recognition score is given to an edge can be output to the speech recognition engine Julius (Non-Patent Document 1). In order to apply the speech recognition processing based on SVM identification that is characteristic in the present invention to such an output, a kernel function of a Rational kernel (Non-patent Document 4) that calculates similarity between weighted automata is used. . In this case, by using this kernel function instead of the above-described character string kernel, it is possible to configure a voice command discriminator in consideration of recognition errors in units of subwords.

  Next, the SVM identification process 108 will be described in detail. In the subword recognition process 108, the recognized subword unit sequence or subword unit graph is used to identify which voice command using the classifier loaded from the classifier database into the memory. Or not a voice command.

  Basically, the discriminator that can be learned by SVM is a two-class discriminator that discriminates whether or not it is A. N two-class classifiers are used to identify whether or not.

  In this way, in the SVM identification process 108, a subword unit sequence or a subword unit graph is input to N classifiers, and N confidence levels representing how probable each command is obtained. Then, the command with the highest certainty factor is adopted as the identification result. However, when the highest certainty factor is smaller than a predetermined threshold, the identification result is rejected, and it is determined that the input voice is a voice other than the command. As will be described later, when an identification result command is obtained, a control signal for controlling the electric wheelchair corresponding thereto is output.

  FIG. 2 is a block diagram showing the main configuration of the apparatus when the voice recognition processing according to the present invention is used for controlling an electric wheelchair. In FIG. 2, reference numeral 210 denotes an electric wheelchair as a control target device. The electric wheelchair 210 is directly connected to a drive motor 205 for driving wheels on an axle, and a control switch 206 for manually operating the electric wheelchair is arranged at the speaker's hand. The microphone array 207 is provided at an appropriate position such as an armrest portion or a backrest portion of the electric wheelchair 210, for example. A voice signal input from the microphone array 207 is input to the voice recognition device 200, a voice command of the input voice signal is recognized, a corresponding control signal is output, and is input to the control controller 204. In the controller 204, the drive motor 205 is controlled by a control signal output from the speech recognition apparatus 200 or a control signal by manual operation output from the control switch 206.

  In addition, as shown in FIG. 2, in one embodiment of the electric wheelchair using the voice recognition processing according to the present invention, the electric wheelchair 210 includes a voice recognition device 200, and the voice recognition device 200 includes an analog voice of a microphone array. Data processing apparatus for performing input sub-word recognition, learning a discriminator using SVM, identifying sub-word unit sequence or sub-word unit graph, input unit 207, AD conversion unit 201 for digitizing analog speech (CPU) 202 and memory 203 are provided.

  The speech recognition apparatus 200 further includes an acoustic model database 121 that stores an acoustic model for performing subword recognition, and a recognized subword unit sequence or subword unit graph for use in command identification and classifier learning. A command sample database 122 for saving and a discriminator database 123 for storing learned discriminators are provided.

  Further, as described above, a joystick control switch 206 for a normal manual operation is provided in addition to an operation by a voice command, and a control signal from the control switch 206 and a control signal from the voice recognition device 200 are input. The control controller 204 selects an appropriate control signal and finally controls the drive motor 205.

  Here, the acoustic model database 121, the command sample database 122, and the discriminator database 123 can be accessed at high speed, and further, a nonvolatile memory or a database that does not lose its contents even when the power is turned off. It is desirable to mount using a hard disk drive.

  Moreover, since it is desirable to prepare these databases (121, 122, 123) for each speaker, a plurality of databases are mounted when a plurality of speakers use an electric wheelchair. However, it has been confirmed by experiments that the acoustic model database 121 does not deteriorate so much even if it is shared by a plurality of speakers.

  FIG. 3 is a diagram showing a result of voice command identification experiment of a cerebral palsy patient using voice recognition processing according to the present invention. In the figure, “STD-FULL” indicated by a two-dot chain line is an experimental result obtained by registering a phoneme sequence based on a normal utterance of a healthy person in the command dictionary, and “ALL-FULL” indicated by a broken line. Is an experimental result obtained by registering a phoneme string obtained from an actual voice command of a subject using a phoneme typewriter in a dictionary, and “SVM-FULL” indicated by a solid line is a voice recognition process according to the present invention. The experimental result using is shown.

  In each case, an experiment for identifying five commands corresponding to front, rear, right, left, and stop and voices other than the command was performed, and the command reproduction rate and matching rate were plotted while changing the threshold value. As for the five commands, taking into account the ease of speech of the subject, etc., front: / mae /, rear: / koutai /, right: / migi /, left: / hidari / or / dari /, stop: / A- / was used.

  From the results of this experiment, it can be seen that the speech recognition processing technique according to the present invention can achieve a higher precision than the existing technique for most reproduction rate ranges.

  According to the indistinct voice command recognition device of the present invention, the voice recognition device for the disabled or the elderly who cannot use the conventional speech recognition device because of the indistinct voice, or because of the noise, Even in a situation where the voice recognition device cannot be used, it is possible to control the device using voice commands.

It is a flowchart which shows an example of the processing flow of the speech recognition process in the unclear voice command recognition apparatus of this invention, the unclear voice command recognition processing method, and an unclear voice command recognition processing program. It is a block diagram which shows the main structures of the apparatus in the case of using the speech recognition process by this invention for control of an electric wheelchair. It is a figure which shows the result of the voice command identification experiment of the cerebral palsy patient using the voice recognition process by this invention.

Explanation of symbols

121 Acoustic Model Database 122 Command Sample Database 123 Discriminator Database 200 Speech Recognition Device 201 AD Converter 202 Data Processing Device (CPU)
203 Memory 204 Control Controller 205 Drive Motor 206 Control Switch 207 Microphone Array 210 Electric Wheelchair

Claims (12)

An obscure voice command recognition device for operating a device using voice utterly uttered as a voice command,
Voice input means for converting voice signals of voice commands into electric signals and inputting the voice signals;
Analog-digital conversion means for converting the electrical signal of the voice signal input by the voice input means into digital data;
Feature extraction means for extracting feature vectors from digital data of speech signals by cepstrum analysis;
Subword recognition means for recognizing a subword unit of speech from a time series of feature vectors and outputting a sequence or graph of the subword unit;
Database update means for generating training data for command identification by a subword column or graph and registering it in the database;
A support vector machine learning means for configuring a command classifier by a support vector machine based on the training data;
A voice identification means for identifying a voice command by data processing by a command identifier constituted by the support vector machine learning means;
A control signal output means for generating a control signal according to the voice command identified by the voice identification means and sending the control signal to the control target device;
An indistinct voice command recognition device comprising: In the unclear voice command recognition device according to claim 1,
An unclear voice command recognition apparatus, wherein the subword recognition means recognizes a subword unit based on a unit of a phoneme, a syllable, or a phoneme of a voice signal. In the unclear voice command recognition device according to claim 1,
The support vector machine learning unit and the voice identification unit use a character string kernel for a subword unit column and a relational kernel for a subword unit graph. In the unclear voice command recognition device according to claim 1,
The voice input means is a microphone array in which a plurality of microphones are arranged.
Sound source separation means for separating a plurality of sound sources from digital data of a plurality of sound signals obtained from the microphone array and removing directional noise;
Feature correction means for removing stationary noise from the audio signal;
An unclear voice command recognizing device comprising: a non-speech identifying means for distinguishing between speech and non-speech by fundamental frequency estimation. An ambiguous voice command recognition processing method for recognizing and processing an ambiguous voice command for operating a device by using an unclear voice as a voice command.
A voice input step of converting a voice signal of a voice command into an electric signal and inputting the voice signal;
An analog-digital conversion step in which the electrical signal of the audio signal input by the audio input means is converted into digital data;
A feature extraction step of extracting a feature vector from the digital data of the speech signal by cepstrum analysis;
A subword recognition step of recognizing a subword unit of speech from a time series of feature vectors and outputting a sequence or graph of the subword unit;
A database update step for generating training data for command identification by a subword column or graph and registering it in the database;
A support vector machine learning step of configuring a command classifier by a support vector machine based on the training data;
A voice identification step of identifying a voice command by data processing by a command classifier configured by the support vector machine learning step;
A control signal output step of generating a control signal according to the voice command identified by the voice identification means and sending it to the device to be controlled;
A process for recognizing an unclear voice command characterized by executing the following process. The unclear voice command recognition processing method according to claim 5,
In the subword recognition step, the subword unit is recognized by any one unit of a phoneme, a syllable, or a phoneme of a speech signal. The unclear voice command recognition processing method according to claim 5,
In the support vector machine learning step and the voice identification step, a character string kernel is used for a subword unit column, and a relational kernel is used for a subword unit graph. The unclear voice command recognition processing method according to claim 5,
The audio input step inputs a plurality of audio signals from a microphone array in which a plurality of microphones are arranged,
A sound source separation step of separating a plurality of sound sources from digital data of a plurality of sound signals obtained from the microphone array and removing directional noise;
A feature correction step for removing stationary noise from the audio signal;
An unclear voice command recognition processing method characterized by further executing a process of a non-voice identification step of distinguishing between voice and non-voice by fundamental frequency estimation. A program for executing an unclear voice command recognition process for operating a device by using an unclearly spoken voice as a voice command, which is a computer that converts a voice command voice signal into an electric signal and inputs it. Input means;
Analog-digital conversion means for converting the electrical signal of the voice signal input by the voice input means into digital data;
Feature extraction means for extracting feature vectors from digital data of speech signals by cepstrum analysis;
Subword recognition means for recognizing a subword unit of speech from a time series of feature vectors and outputting a sequence or graph of the subword unit;
Database update means for generating training data for command identification by a subword column or graph and registering it in the database;
A support vector machine learning means for configuring a command classifier by a support vector machine based on the training data;
Voice identification means for identifying a voice command by data processing by a command classifier constituted by the support vector machine learning means, and control for generating a control signal by the voice command identified by the voice identification means and sending it to a control target device A program for recognizing an unclear voice command characterized by functioning as signal output means. The unclear voice command recognition program according to claim 9,
The subword recognition means recognizes a subword unit based on any one of a phoneme, a syllable, or a phoneme of a speech signal. The unclear voice command recognition program according to claim 9,
The support vector machine learning unit and the voice identification unit use a character string kernel for a subword unit column and a relational kernel for a subword unit graph, and an unclear voice command recognition program. The unclear voice command recognition program according to claim 9,
The audio input means functions to input a plurality of audio signals from a microphone array in which a plurality of microphones are arranged.
Sound source separation means for separating a plurality of sound sources from digital data of a plurality of sound signals obtained from the microphone array and removing directional noise;
Feature correction means for removing stationary noise from the audio signal;
An unclear voice command recognition program which functions as a non-speech discrimination means for distinguishing between speech and non-speech based on fundamental frequency estimation.

Images