JP2007206603A - Method of creating acoustic model - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an acoustic model capable of recognizing speech with good performance even when utterance distortion arising from boarding and so forth exist. <P>SOLUTION: A method of creating an acoustic model includes: a step of recording prescribed speech in an environment having substantially no acoustic noise and creating an undistorted speech corpus 1; a step of correcting the speech of the undistorted speech corpus created in the step using an utterance conversion filter means 2 for deforming it into the utterance corresponding to the moving speed of a mobile body and for creating a mobile environment speech corpus 3; and a step of learning the mobile environment speech corpus 3 created by the step as learning data to create the acoustic model 4. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a method for creating an acoustic model that may be used in a moving body represented by an automobile, a speech recognition device, and speech recognition.

  Automobile navigation devices that recognize the utterances of automobile drivers and passengers (hereinafter also referred to simply as occupants) and use them as operation instructions have been put into practical use. In the recognition device, it is known that a change in the environment of riding in a car affects an occupant, and utterance distortion caused thereby affects the recognition performance of the speech recognition system.

  For example, it is known that speech distortion (referred to as the Lombard effect) caused when a speaker on a car listens to running noise is a very robust phenomenon. In addition, according to the study of the present inventors, it has been confirmed that a passenger, particularly a driver, is tense when riding in an automobile, and the utterance distortion is caused by this tension (see FIG. 13). In many cases, such a phenomenon that utterance distortion occurs naturally occurs unconsciously.

  As shown in FIGS. 14 and 15, a conventional speech recognition device includes an element called an acoustic model having information for performing speech analysis, feature extraction, pattern matching, and the like.

  An acoustic model describes a phoneme or phoneme label and a signal obtained by converting a speech signal corresponding to the label into acoustic feature information, etc. In recent years, speech using a hidden Markov model (Hidden Markov Model) is used. Many recognition methods are used. A hidden Markov model is one of probabilistic models, and is a method for estimating an unknown parameter from observable information on the assumption that the system is a Markov process with an unknown parameter. In this acoustic model using the Hidden Markov Model, model learning of the Hidden Markov Model is performed by using a speech corpus including a speech signal having a speech waveform as a discrete signal and a phoneme label attached to the speech signal as learning data. Thus, the target acoustic model is created.

  By the way, since such an acoustic model depends on the content of the speech corpus as learning data, it is desirable that the content of the speech corpus is suitable for an actual environment in which speech recognition is used.

  That is, when creating an acoustic model for use in a speech recognition device, the closer the features of speech data in the speech corpus are to the features of the speaker's speech in the actual environment, the higher the speech recognition performance. . Therefore, in order to create an acoustic model of the in-vehicle speech recognition device, it is desirable to create using an audio corpus by utterances in which these utterance distortions occur.

  However, in order to record utterance data including utterance distortion, it is necessary to record while keeping the driver driving and not including ambient noise. It is technically very difficult to do. Moreover, even if this has become possible due to technological development, creating a voice corpus for each traveling environment of a vehicle is enormous and requires a lot of man-hours and costs.

On the other hand, although a method for correcting speech including utterance distortion input in an actual use environment has been proposed (for example, Non-Patent Document 1), the behavior of the utterance distortion phenomenon during traveling has not been clarified. Therefore, the processing process is limited only to correction of a partial region such as a cepstrum, and it cannot be said that the speech recognition performance is sufficient.
"Speech recognition under noisy environments using speech distortion model" (acoustics of the Acoustical Society of Japan, March 1995) Tadashi Suzuki, Yoshiharu Abe, Kunio Nakajima (Mitsubishi Electric & Information Technology Laboratories)

  An object of the present invention is to provide an acoustic model, a speech recognition apparatus, and a speech recognition method capable of recognizing speech with good performance even when there is speech distortion caused by riding or the like.

  In order to achieve the above object, the invention according to the first aspect creates a non-distorted speech corpus by recording predetermined speech in an environment substantially free of acoustic noise, and the speech of this undistorted speech corpus is Using the speech transformation filter means that transforms into speech corresponding to the moving speed of the moving object, the mobile environment speech corpus is corrected to create an acoustic model by learning based on the mobile environment speech corpus To do.

  In the invention according to the second aspect, the speech signal to be recognized is a speech signal having the same acoustic and statistical characteristics as the speech corpus used at the time of learning the acoustic model according to the detected moving speed of the moving body. The signal is corrected to a signal, and the corrected sound signal is converted into a label signal using an acoustic model.

  In the invention according to the first aspect, the utterance distortion due to movement is taken into account by correcting the deformation to the utterance corresponding to the moving speed of the moving body when creating the moving environment speech corpus that becomes the learning data of the acoustic model, In the invention according to the second aspect, when decoding using the acoustic model, the speech signal to be recognized is such that the speech signal has the same acoustic and statistical characteristics as the speech corpus that is the learning data of the acoustic model. Utterance distortion due to movement is taken into account. As a result, it is possible to recognize the speech with good performance even when there is an utterance distortion caused by riding or the like.

  The “environment substantially free of acoustic noise” as used in the claims, description and drawings means that there is a change in physiological phenomena such as noise caused by the driving environment or driving condition, or speaker tension. Means no environment or condition.

BEST MODE FOR CARRYING OUT THE INVENTION

  DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following description, a driver is used as an operation instruction for various in-vehicle devices such as a navigation device, an air conditioner, and an audio device mounted on a vehicle as a moving body. A speech recognition apparatus and speech recognition method for recognizing speech of passengers and passengers and a method of creating an acoustic model used therefor will be described as examples.

  However, the acoustic model creation method, speech recognition apparatus, and speech recognition method of the present invention can be applied to a mobile body other than an automobile, and not only a device that is always used in a mobile body, such as an in-vehicle device. For example, it is also within the scope of the present invention to apply to an operation instruction of a device that may be used in a mobile object such as a mobile phone.

<< First Embodiment >>
FIG. 1 is a block diagram showing a first embodiment of a method for creating an acoustic model of the present invention.

  The acoustic model creation method of the present embodiment includes a first step of creating a non-distorted speech corpus 1 by recording predetermined speech in an environment substantially free of acoustic noise, and a non-distortion created in the first step. A second step of correcting the voice of the voice corpus 1 by using the utterance conversion filter means 2 that transforms the voice into the utterance corresponding to the moving speed of the moving body to create the mobile environment voice corpus 3 and the second step. And learning the mobile environment speech corpus 3 as learning data to create a target acoustic model 4 at least.

  The acoustic model 4 is a dictionary used when analyzing the characteristics of an input speech signal and converting it into phoneme or phoneme information, and many speech signals included in a speech corpus (speech signal with a text label). In order to express the acoustic features of the above, it is a signal model that probabilistically generates time series of various lengths.

  In general, a signal model using a hidden Markov model is often used in the acoustic model 4, and learning of the hidden Markov model can be realized by using, for example, a Baum-Weltch learning algorithm. (See, for example, Japanese Patent Laid-Open No. 9-152886, “Voice Recognition System” Ohm Corporation). These acoustic models are considered to improve the recognition performance of the speech recognition system as the acoustic features of the speech set included in the speech corpus to be learned are closer to the acoustic features of the speech set input in the real environment. Therefore, in order to improve the performance of the speech recognition system, it is necessary to collect utterances that are close to utterances in the real environment and to make a speech corpus.

  That is, in order to create an acoustic model that can cope with utterance distortion during travel, the speech corpus used as learning data needs to include speech related to utterance distortion during travel.

  Therefore, in the present embodiment, an undistorted speech corpus 1 is created in advance by recording in a clean environment such as a soundproof room where the influence of noise and reverberation characteristics is small, and the clean speech included in the undistorted speech corpus 1 is created. A signal obtained by transforming the signal into a speech signal with utterance distortion during traveling using the speech conversion filter means 2 prepared for each vehicle speed is used as a mobile environment speech corpus 3, and the mobile environment speech corpus 3 after the transformation is used as learning data. The acoustic model 4 is learned using the Baum-Welch learning algorithm described above.

  In this way, the content of the mobile environment speech corpus 3 which is the learning data of the target acoustic model 4 includes a traveling speech signal, and when speech recognition is performed by the acoustic model 4 thus learned. In addition, it is possible to obtain a robust performance that does not deteriorate the recognition accuracy even for the utterance in the driving environment.

  The undistorted speech corpus 1 records an enormous speech utterance on a personal computer or the like in a stationary environment that is less affected by noise and reverberation characteristics and is free from the tension of the speaker in a running environment, for example, the above-described soundproof room. Then, the undistorted speech corpus 1 is obtained by giving label information such as from where to what in the recorded speech utterance signal (computer sound file).

  When the undistorted speech corpus 1 is created, the speech is converted by using the speech conversion filter means 2 for each traveling speed of the automobile, and the mobile environment speech corpus 3 is created.

  The speech conversion filter means 2 (parameters) according to the present embodiment can include the following means, but the contents of these speech conversion filter means 2 are based on the results of the following experiment conducted by the inventors. To do.

  First, the utterances of 6 test subjects who can speak are recorded in the driving environment and the laboratory environment. In the driving environment, utterances were recorded in each of the driving environments maintaining the four speeds of idling, 30 km / h driving, 60 km / h driving, and 100 km / h driving. The subject records while sitting in the driver's seat or passenger's seat, uses a close-up microphone for recording, and the content of the paper with one word written on one page in front of the subject is recorded under multiple circumstances. Recorded for each utterance.

  On the other hand, in a laboratory environment, the subject listens to the subject in the semi-anechoic room while listening to binaural recordings of noise in the subject's binaural position when idling, traveling at 30 km / h, traveling at 60 km / h, and traveling at 100 km / h. The contents of a sheet with one word written on one page placed in front were recorded for each utterance under multiple environments. This noise was adjusted to the sound pressure level at the time of recording.

  Collect voice utterances of 6 male and female speakers recorded in each of these environments, and change the parameters related to power, fundamental frequency, spectral tilt, formant frequency, and speech speed to high-quality speech analysis conversion synthesis method STRAIGHT (“Analysis of auditory scene analysis and high-quality speech analysis transformation synthesis method STRAIGHT” Hidenori Kawahara, Proceedings of the Acoustical Society of Japan 1-2-1, pp.189-192, Sep.1997) was used for analysis.

(1) Audio power (energy)
According to the experiments by the present inventors, it has been clarified that the sound power increases as the traveling speed increases in the analysis result of 90% or more. Specifically, it is clear that it is possible to correct from the laboratory environment to the driving environment by increasing the voice power of 3.5 dB on average and 7 dB at maximum with respect to the vehicle speed changes of 0 km / h and 100 km / h. Became. In addition, when the undistorted audio corpus 1 is created in an acoustically clean environment (such as a soundproof room), recording is performed in a clean environment by increasing the audio power by 6.8 dB on average and 14 dB at maximum. It was also confirmed that a mobile environment speech corpus including speech distortion when traveling at a speed of 100 km / h with respect to the undistorted speech corpus 1 can be created.

  Therefore, in this embodiment, as one form of the utterance conversion filter means 2, as the running speed increases, the power average after extracting the unit frame of the audio signal (observed for a predetermined time using a window function or the like). The average of the power of the audio signal) or the power sum of the audio section after the cutout is increased. As a whole, the sound power may be increased by a maximum of 7 dB compared to the idling time and by a maximum of 14 dB compared with the sound recorded in the clean environment.

  This parameter is changed when applied to a speech recognition system that absorbs the effects of power fluctuations by a method such as normalization, or a speech recognition system that does not use power fluctuations in individual phonemes and phonemes as recognition parameters. You don't have to.

(2) Fundamental frequency of speech The lowest frequency of the harmonic components of speech frequency is called the fundamental frequency of speech, but it is known to match the fundamental frequency of human vocal cord vibration. It is said to be a physical characteristic of height.

  According to the experiments by the present inventors, it has been found that the fundamental frequency increases as the traveling speed increases in the analysis result of 70% or more. Specifically, the vehicle speed increased by 21 Hz at maximum with respect to changes in vehicle speed of 0 km / h and 100 km / h. In comparison between the undistorted speech corpus recorded in the clean environment and the mobile environment speech corpus, the maximum was increased by 36 Hz. However, this parameter was observed to be highly dispersed and not change depending on the speaker.

  Therefore, in the present embodiment, as one form of the speech conversion filter means 2, the fundamental frequency is increased as the traveling speed increases. More specifically, since the amount of change is 36 Hz at the maximum, for example, when the vehicle speed changes from 0 km / h to 100 km / h, it is increased by 9 Hz every 25 km / h.

  As a general rule, the speech conversion filter means 2 is configured to increase the fundamental frequency as the traveling speed increases. However, as described above, since this parameter has a large variance, the speech corpus corrected by the speech conversion filter means 2 is not corrected. A speech corpus may coexist, and when creating an acoustic model, a search route composed of a corrected speech corpus (mobile environment speech corpus 3) and a search composed of an uncorrected speech corpus (undistorted speech corpus 1) A route may coexist.

  Moreover, since it has been confirmed by experiments of the present inventors that the variation varies depending on the type of vowel, it may be set for each detected vowel. For example, from the fluctuation results for each vowel in the survey results, / a / (A) is increased by about 25 Hz, and / i / (I) is inconsistent in fluctuation, so 0 Hz and / u / (U) are increased by about 29 Hz. And / e / (e) may be increased by approximately 36 Hz, and / o / (o) may be increased by approximately 26 Hz. This will be described later.

(3) Speech spectrum regression line The speech spectrum regression line is an element obtained by approximating the frequency spectrum of speech spectrum envelope from 0 Hz to 4 kHz with a linear line, and an example thereof is indicated by a straight line X in FIG.

  According to the experiments by the present inventors, it has been found that the slope of the spectral regression line increases with an increase in traveling speed in an analysis result of 80% or more. For example, with respect to the undistorted audio signal recorded in a clean environment, the spectral inclination of the audio signal recorded in a moving environment when traveling at 100 km / h increased by 0.0081 at the maximum. On the other hand, between the audio signals in the moving environment at the time of traveling, for example, in the relationship between traveling at 0 km / h and traveling at 100 km / h, the slope of the spectral regression line increases by a maximum of 0.0067 only in the analysis result of about 56%. Turned out to be.

  Therefore, in the present embodiment, as one form of the utterance conversion filter means 2, when the speech corpus used at the time of learning the acoustic model is an undistorted speech signal in a clean environment, the travel speed increases. Correction is made to increase the slope of the spectral regression line.

  For example, the high range is increased to increase 0.5 dB per 1 kHz up to 4 kHz, and 0.5 → 1. Every time the traveling speed increases from idling (0 km / h) → 30 km / h → 60 km / h → 100 km / h. Processing such as increasing the slope such that 0 → 1.5 → 2.0 dB (increase by about 1.5 dB for the entire waveform) is performed. In addition, parameter modification common to all phonemes may be performed, or individual increase settings may be performed for each phoneme such as a vowel, as in the case of the basic frequency of speech described above.

(4) Voice formant frequency As shown in FIG. 10, the formant frequency refers to the median or maximum amplitude frequency of a peak generated by energy concentration in a specific frequency region on the spectrum envelope of the voice. Formant frequencies are mainly observed in stationary vowels and are called first formant and second formant in order from the lowest frequency, and the combination of each formant frequency differs for each vowel. Formant is a characteristic caused by resonance of the human vocal tract during speech generation.

  Figure 11 (Source: Tatsumi Ifukube “Design of Voice Typewriter” CQ Publishing 1984) takes the first formant frequency F1 on the horizontal axis and the second formant frequency F2 on the vertical axis. U, e, o) distribution is displayed, the solid circle connected ○ mark indicates male voice, and the dotted line ● mark indicates female voice. However, the voice utterance in the mobile environment changes as shown in FIG. 12, for example, as shown in the figure, the voice of / u / (u) may be recognized as / e / (e) in some cases. .

  As described above, the formant frequency parameter shows a correlation between the clean environment and the driving environment, and different results are obtained depending on gender and the like.

(4-1) First formant frequency In an investigation in the laboratory environment, an increase in the first formant frequency was observed in the analysis result of 80% or more in the relationship between the clean voice signal and the running voice signal at 100 km / h. It was done. Specifically, the first formant frequency increased by about 300 Hz at the maximum and about 50 Hz on the average with respect to the clean sound signal and the environmental change at 100 km / h.

  On the other hand, with respect to the traveling voice signals, for example, in the relationship between traveling at 0 km / h and traveling at 100 km / h, an increasing tendency of the first formant frequency was observed in the analysis result of 75% or more. Specifically, with respect to environmental changes during travel at 0 km / h and at 100 km / h, the first formant was observed to increase at a maximum of about 300 Hz and an average of about 10 Hz.

  On the other hand, in the survey in the driving environment, the tendency changed depending on the sex of the subject. That is, in the female subject, the first formant decreased for the result of 70% or more in the relationship between the clean voice signal and the running voice signal at 100 km / h. Specifically, the first formant decreased by about 100 Hz at the maximum and about 20 Hz on the average with respect to the clean sound signal and the environmental change at 100 km / h.

  On the other hand, also in the traveling voice signals, for example, in the relationship between traveling at 0 km / h and traveling at 100 km / h, a decreasing tendency of the first formant frequency was observed in the analysis result of 75% or more. Specifically, with respect to environmental changes during travel at 0 km / h and 100 km / h, the first formant was observed to decrease at a maximum of about 170 Hz and an average of about 50 Hz.

(4-2) Second formant frequency In either the laboratory environment or the traveling environment, there was no tendency to increase or decrease. However, there was a slight increase in male subjects in any environment. Specifically, the second formant increases at a maximum of about 150 Hz and an average of about 7 Hz with respect to the clean sound signal of the driving environment and the environmental change during driving at 100 km / h, and when driving at 0 km / h and 100 km / h. With respect to the environmental changes, the second formant was observed to increase at a maximum of about 130 Hz and an average of about 50 Hz.

  The second formant frequency has a different tendency for each phoneme. For example, the phoneme of / a / (A) can be used regardless of gender for the clean voice signal of the driving environment and the environmental change during driving at 100 km / h. As a whole, there was a downward trend, and the maximum was reduced by about 90 Hz.

(4-3) 3rd formant Although there is no tendency to increase or decrease in the laboratory environment, the relationship between the clean audio signal in the driving environment and the driving audio signal at 100 km / h is about 70% or more. Third formant increased. In addition, male subjects tended to increase as a whole.

  Specifically, the third formant increases at a maximum of about 390 Hz and an average of about 100 Hz with respect to the clean sound signal of the driving environment and the environmental change during driving at 100 km / h, and when driving at 0 km / h and at 100 km / h. With respect to the environmental change, the third formant was observed to increase at a maximum of about 540 Hz and an average of about 30 Hz.

  The third formant frequency also has a different tendency for each phoneme. For example, the / o / (o) phoneme is the same regardless of gender for clean voice signals in the driving environment and environmental changes during driving at 100 km / h. Although it was increasing, it increased by about 20 Hz at maximum, but the phoneme of / e / (E) was decreasing regardless of gender, and decreased by about 70 Hz at maximum.

Therefore, in the present embodiment, as one form of the utterance conversion filter means 2, it is preferable to perform parameter modification for the formant frequency for each vowel, for each speech corpus environment, and for each gender. An example of performing formant frequency parameter modification for each vowel is shown in Table 1 below based on the survey results.
(Table 1)
/ a /: Every time the speed increases by 30km / h, the first formant is 10Hz.
Second formant: 20Hz
Third formant: Increase by 10Hz
/ i /: Every time the speed increases by 30km / h, the first formant: 20Hz
Second formant: 5Hz
Third formant: Increase by 0Hz
/ u /: First formant: 20Hz every time the speed increases by 30km / h
Second formant: 0Hz
Third formant: Increase by 10Hz
/ e /: Every time the speed increases by 30km / h, the first formant is 20Hz.
Second formant: 5Hz
Third formant: Increase by 0Hz
/ o /: Every time the speed increases by 30km / h, the first formant: 15Hz
Second formant: 10Hz
Third formant: Increase by 0 Hz However, the amount of change of these parameters is an example, and may be corrected as necessary depending on the recording environment of the undistorted speech corpus and the target driving environment.

(5) Extension of the beginning of the voice utterance vocabulary (leading mora) As a result of the present inventors' extensive research, the beginning of the word, that is, the beginning mora (mora: relative length of sound that is a unit of stress or inflection in accents) It has been confirmed that the sustaining length of the vehicle becomes longer depending on the driving environment. Therefore, in the present embodiment, as one form of the utterance conversion filter unit 2, the mora length of the beginning is extended by about 40% at maximum as compared with the non-running environment.

(6) Extension of the utterance (final mora) of the speech utterance vocabulary As a result of our earnest research, when the auditory feedback is hindered by noise accompanying the running speed, the speaker tends to speak carefully and clearly. It has been confirmed that there is. Therefore, in the present embodiment, as one form of the utterance conversion filter unit 2, the ending mora length is extended by about 80% at maximum as compared with the non-running environment.

  As described above, as parameters of the speech conversion filter unit 2 according to the present embodiment, the power of speech, the fundamental frequency of speech, the slope of the speech spectral regression line, the speech formant frequency (first to third formant frequencies), the speech The head of the vocabulary and the ending of the utterance vocabulary are illustrated, but these may be used alone or in combination of two or more parameters. The concrete means to convert these parameters is a method that is generally proposed as a morphing technique ("Hearing scene analysis and high quality speech analysis conversion synthesis method STRAIGHT" Hideki Kawahara, Acoustical Society of Japan Vol. 1-2-1, pp.189-192, Sep.1997) etc. can be used for deformation. However, the algorithm is not particularly limited to the conversion method described in the document, and any algorithm that can transform the utterance parameters can be applied.

  Returning to FIG. 1, when the undistorted speech corpus 1 is converted into the mobile environment speech corpus 3 for each travel speed by the speech conversion filter means 2 described above, the mobile environment speech corpus 3 created for each travel speed is used as learning data. By learning, the acoustic model 4 for each target traveling speed can be created. As the learning method in this case, the above-described Balm-Welch learning algorithm or the like can be used.

  Since the acoustic model 4 of the present embodiment is created taking into account the utterance distortion corresponding to the traveling speed, it can exhibit high recognition performance for the utterance input during actual traveling.

  Next, a speech recognition method and speech recognition apparatus using the above-described acoustic model 4 will be described. FIG. 2 is a block diagram showing a first embodiment of the speech recognition apparatus of the present invention, and FIG. 3 is a flowchart showing a control procedure of the first embodiment of the speech recognition apparatus of the present invention.

  As shown in FIG. 2, the speech recognition apparatus according to the present embodiment generates noise from a speech input device 5 including a microphone for inputting speech to be recognized, and a speech signal input to the speech input device 5. A noise removing unit 6 including a noise filter for removing the noise, a decoder 7 for converting the speech signal from which noise has been removed to a text signal, and a dictionary referred to when the decoder 7 converts the audio signal An acoustic model 4 and a language model 8 are included.

  The decoder 7 extracts parameters that express the characteristics from the audio signal input to the audio input device 5, and parameters associated with the text information recorded in the acoustic model 4 and the language model 8 stored in a memory or the like. Compare and output the most appropriate text information. At this time, for example, a technique using statistical parameters represented by a hidden Markov model can be used.

  In particular, in this embodiment, as described above, the acoustic model 4 corresponding to each traveling speed (four types of 0, 30, 60, and 100 km / h in the embodiment) is stored in a memory or the like.

  Further, in the voice recognition device according to the present embodiment, the vehicle speed detection device 9 that detects the traveling speed of the vehicle, and the four types of acoustic models 4 stored in the memory according to the vehicle speed detected by the vehicle speed detection device 9. A selection device 10 for selecting the corresponding acoustic model 4 from the inside is provided. The selection device 10 has a map in which the correspondence relationship between the vehicle speed range and the acoustic model 4 within the range is determined in advance, and the most suitable acoustic model 4 is selected based on the actual vehicle speed detected by the vehicle speed detection device 9. select. For example, the acoustic model 4 of 0 km / h is selected when the speed is 0 to 15 km / h, and the acoustic model 4 of 30 km / h is selected when the speed is 15 to 45 km / h.

  Next, the operation of the speech recognition apparatus of this embodiment will be described with reference to FIG.

  First, initialization processing is performed in step S100. At this time, all the speech recognition processes are initialized. The voice recognition apparatus may be in an input signal waiting state for voice recognition processing, or may be activated at a timing when the user indicates an intention to input (for example, a PTT switch is turned on) to enter an input signal waiting state.

  In step S110, a change in vehicle speed is detected using CAN (Controller Area Network vehicle-mounted LAN standard) or the like. If the vehicle speed has changed, the process proceeds to step S115, and if not, the detection process in step S110 is repeated. The variation range of the vehicle speed conforms to the corresponding vehicle speed of the acoustic model 4 prepared for each vehicle speed. For example, when acoustic models of 30 km / h and 60 km / h are prepared, it is assumed that a change in vehicle speed is detected at a stage exceeding 45 km / h.

  Next, in step S115, the vehicle speed when the sound signal is input by the user is detected, and the corresponding acoustic model 4 is selected from the acoustic model 4 shown in FIG.

  Next, if a voice input is detected in step S120, the process proceeds to step S130. If a voice input is not detected, the process returns to step S110 and the above processing is repeated.

  In step S130, the input speech signal is recognized. At this time, when the acoustic model 4 is selected in step S115, speech recognition processing is performed using the selected acoustic model 4.

  Finally, in step S140, the recognized voice recognition processing result, that is, text information is sent to the other operation device.

  In the speech recognition apparatus according to the present embodiment, the element to be mounted is the one shown in FIG. 2 including a plurality of acoustic models 4, and although the capacity of the acoustic model 4 is increased, the speech input to the speech input device 5 is Since the text information can be output to the outside by simply converting it with the decoder 7 as it is, not only the speech recognition performance but also the recognition speed can be increased.

<< Second Embodiment >>
FIG. 4 is a block diagram showing a second embodiment of a method for creating an acoustic model according to the present invention.

  In the first embodiment described above, a plurality of acoustic models 4 that take into account the utterance distortion under a moving environment are created for each traveling speed, and this is applied to the in-vehicle voice recognition device, thereby improving the speech recognition performance including the utterance distortion. Although it is configured to increase, there is a demerit that a capacity necessary for storing the plurality of acoustic models 4 is increased.

  Therefore, in this embodiment, the storage capacity of the voice recognition device mounted on the vehicle is made small. That is, as shown in FIG. 4, a first step of creating a non-distorted speech corpus 1 by recording predetermined speech in an environment substantially free of acoustic noise, and a non-distorted speech corpus created in the first step A second step of learning 1 as learning data to create a distortion-free acoustic model 11 and correction using speech conversion filter means 2 that transforms the speech of the distortion-free speech corpus 1 into speech corresponding to the speed of the moving object A third step of creating the mobile environment speech corpus 3, and a fourth step of creating the acoustic model 4 by adapting the undistorted acoustic model 11 using the mobile environment speech corpus 3 created in the third step; , At least.

  Here, since the configurations of the undistorted speech corpus 1, the speech conversion filter means 2 and the acoustic model 4 are the same as those in the first embodiment described above, the detailed description thereof is omitted, but the undistorted sound according to the present embodiment. The model 11 is an acoustic model obtained by using a large-scale undistorted speech corpus 1 as learning data, for example, a Balm-Welch learning algorithm. The undistorted acoustic model 11 has a large capacity and is mounted on a vehicle. Are stored in the voice recognition device, but a plurality of them are not present for each traveling speed, but are configured as one acoustic model.

  Instead, the capacity of the mobile environment speech corpus 12 according to the present embodiment is configured by a small-scale corpus, and a plurality of small-scale mobile environment speech corpora 12 corresponding to the traveling speed are created and mounted on the vehicle. Store in the device.

  Then, using this small mobile environment speech corpus 12, using, for example, an environment adaptation algorithm such as MLLR (Maximum Likelihood Linear Regression algorithm for converting an acoustic model for speaker-independent speaker recognition to a speaker) The target acoustic model 4 is created by adaptation.

  Since the acoustic model 4 of the present embodiment is created taking into account the utterance distortion corresponding to the traveling speed, it can exhibit high recognition performance for the utterance input during actual traveling. In addition, since the mobile environment speech corpus 12 can be configured with a small corpus, it is preferable to be applied to a speech recognition apparatus mounted on a vehicle.

  Next, a speech recognition method and speech recognition apparatus using the above-described acoustic model 4 will be described. FIG. 5 is a block diagram showing a second embodiment of the speech recognition apparatus of the present invention, and FIG. 6 is a flowchart showing a control procedure of the second embodiment of the speech recognition apparatus of the present invention.

  As shown in FIG. 5, the voice recognition apparatus according to the present embodiment generates noise from a voice input device 5 including a microphone for inputting voice to be recognized, and a voice signal input to the voice input device 5. A noise removing unit 6 including a noise filter for removing the noise, a decoder 7 for converting the speech signal from which noise has been removed to a text signal, and a dictionary referred to when the decoder 7 converts the audio signal An acoustic model 4 and a language model 8 are included.

  Here, since the voice input device 5, the noise removing unit 6, and the decoder 7 are the same as those in the first embodiment described above, detailed description thereof is omitted.

  In particular, in this embodiment, as described above, the mobile environment voice corpus 12 corresponding to each traveling speed (four types of 0, 30, 60, and 100 km / h in the embodiment) is stored in a memory or the like.

  In the speech recognition apparatus according to the present embodiment, the vehicle speed detection device 9 that detects the traveling speed of the vehicle, and four types of moving environment speech corpora stored in the memory according to the vehicle speed detected by the vehicle speed detection device 9. A selection device 10 for selecting a corresponding mobile environment speech corpus 12 from among the 12 is provided. The selection device 10 has a map in which the correspondence relationship between the range of the vehicle speed and the movement environment voice corpus 12 in the range is determined in advance, and the most suitable movement environment based on the actual vehicle speed detected by the vehicle speed detection device 9. The voice corpus 12 is selected. For example, the mobile environment speech corpus 12 of 0 km / h is selected when it is 0 to 15 km / h, and the mobile environment speech corpus 12 of 30 km / h is selected when it is 15 to 45 km / h. Has been.

  Next, the operation of the speech recognition apparatus of this embodiment will be described with reference to FIG.

  First, initialization processing is performed in step S100. At this time, all the speech recognition processes are initialized. The voice recognition apparatus may be in an input signal waiting state for voice recognition processing, or may be activated at a timing when the user indicates an intention to input (for example, a PTT switch is turned on) to enter an input signal waiting state.

  Next, in step S110, a change in vehicle speed is detected using CAN or the like. If the vehicle speed has changed, the process proceeds to step S116. If not, the detection process in step S110 is repeated. The change width of the vehicle speed conforms to the corresponding vehicle speed of the moving environment voice corpus 12 prepared for each vehicle speed. For example, when the mobile environment voice corpus 12 of 30 km / h and 60 km / h is prepared, a change in the vehicle speed is detected when the speed exceeds 45 km / h.

  Next, in step S116, the vehicle speed when the sound signal is input by the user is detected, and the corresponding mobile environment speech corpus 12 is selected from the mobile environment speech corpus 12 shown in FIG. Adaptation processing is executed by the applicator 13 using the mobile environment speech corpus 12.

  Next, if a voice input is detected in step S120, the process proceeds to step S130. If a voice input is not detected, the process returns to step S110 and the above processing is repeated.

  In step S130, the input speech signal is recognized. At this time, speech recognition processing is performed using the acoustic model 4 adapted in step S116.

  Finally, in step S140, the recognized voice recognition processing result, that is, text information is sent to the other operation device.

  In the speech recognition apparatus according to the present embodiment, the mobile environment speech corpus 12 to be mounted on the vehicle can be configured with a small-scale corpus, so that adaptation processing is added, but not only speech recognition performance but also speech with a small storage capacity. A recognition device can be constructed.

<< Third Embodiment >>
FIG. 7 is a block diagram showing a third embodiment of the acoustic model creation method of the present invention. In creating an acoustic model that takes into account utterance distortion in a mobile environment, the speech conversion filter means 2 is applied in the first embodiment described above considering the input speech signal as one type (category). Depending on a person's individual difference and gender and other categories, the parameter fluctuation mode may vary depending on the traveling speed. For this reason, in the present embodiment, the utterance conversion filter means 2 is prepared according to phonemes such as vowels or depending on whether the speaker is male or female, and the utterance conversion filter means 2 is prepared by the phoneme / speaker selection device 14. A mobile environment speech corpus 3 is created based on the selected utterance conversion filter means 2.

  For example, a phoneme of a vowel of / a /-/ o / is converted using different utterance conversion filter means 2. More specifically, when the phoneme of / a / (A) is input, conversion of the speech conversion filter means 2 is performed. In addition, when both male and female voices are input, a corpus of a voice signal having a reduced first formant frequency and a voice signal having a first formant frequency increased is transmitted. Further, when a female voice is input, the first formant frequency is decreased, and when a male voice is input, a voice signal having the first formant frequency increased is transmitted to the mobile environment voice corpus 3.

  Since the acoustic model 4 of the present embodiment is created in consideration of the utterance distortion corresponding to the traveling speed, in addition to being able to demonstrate high recognition performance for the utterance input during actual traveling, Since the acoustic model 4 is created by performing conversion according to the characteristics of the speaker, the speech recognition performance is further improved.

<< 4th Embodiment >>
FIG. 8 is a block diagram showing a fourth embodiment of the speech recognition apparatus of the present invention, and FIG. 9 is a flowchart showing a control procedure of the fourth embodiment of the speech recognition apparatus of the present invention.

  In the first to third embodiments described above, utterance distortion in a moving environment is incorporated in the acoustic model 4 used in the speech recognition apparatus, but in this embodiment, the input voice signal does not consider the moving environment. Before conversion using a general acoustic model and language model, pre-processing correction considering the moving environment is executed.

  That is, the speech recognition apparatus according to the present embodiment includes a speech input device 5 including a microphone for inputting speech to be recognized as shown in FIG. 8, and noise from the speech signal input to the speech input device 5. A noise removing unit 6 including a noise filter for removing noise, an utterance correction filter unit 15 that corrects an input voice signal to a voice signal obtained by subtracting utterance distortion in a moving environment, and a corrected noise. Has a decoder 7 for converting the speech signal from which the sound is removed into a text signal, and an acoustic model 4 and a language model 8 which are dictionaries referred to when the decoder 7 converts the speech signal.

  Here, since the configurations of the voice input device 5, the noise removal unit 6, the decoder 7, the acoustic model 4, and the language model 8 are the same as those in the second embodiment described above, detailed description thereof is omitted.

  In the present embodiment, the speech signal input by the speech correction filter means 15 is preprocessed and corrected. This preprocessing correction is the same as the speech corpus used when learning the acoustic model 4 for the input speech signal. It corrects to an audio signal having acoustic and statistical characteristics. That is, the speech signal input to the speech input device 5 includes utterance distortion in an actual mobile environment, while the speech corpus that is the learning data of the acoustic model 4 is a mobile environment such as an undistorted speech corpus. This is a speech corpus that does not include speech distortion.

  Therefore, in the present embodiment, the following preprocessing correction is performed by the utterance correction filter means 15, but the utterance correction filter means 15 of the present embodiment is the utterance conversion filter means 2 described in detail in the first embodiment described above. It has the reverse characteristics of

  That is, as parameters of the speech correction filter means 15, speech power, speech fundamental frequency, speech spectral regression line slope, speech formant frequency (first to third formant frequencies), speech vocabulary prefix, speech vocabulary The ending can be illustrated, and the specific correction value is that the parameter “increase” in the utterance conversion filter means 2 of the first embodiment is “decrease” in the utterance correction filter means 15 of this example, Similarly, a parameter that is “extended” in the utterance conversion filter unit 2 of the first embodiment is “reduced” in the utterance correction filter unit 15 of this example. The absolute value of reduction or reduction is the same value as that of the utterance conversion filter means 2 of the first embodiment.

  Next, the operation of the speech recognition apparatus of this embodiment will be described with reference to FIG.

  First, initialization processing is performed in step S100. At this time, all the speech recognition processes are initialized. The voice recognition apparatus may be in an input signal waiting state for voice recognition processing, or may be activated at a timing when the user indicates an intention to input (for example, a PTT switch is turned on) to enter an input signal waiting state.

  Next, in step S110, a change in vehicle speed is detected using CAN or the like. If the vehicle speed has changed, the process proceeds to step S117. If not, the detection process in step S110 is repeated. The change width of the vehicle speed conforms to the corresponding vehicle speed of the speech correction filter means 15 prepared for each vehicle speed. For example, when the speech correction filter means 15 of 30 km / h and 60 km / h are prepared, a change in the vehicle speed is detected at a stage exceeding 45 km / h.

  Next, in step S117, the vehicle speed when the sound signal is input by the user is detected, and the corresponding utterance correction filter means 15 is selected from the utterance correction filter means 15.

  Next, when a voice input is detected in step S120, the process proceeds to step S125, and the input voice signal is corrected using the speech correction filter means 15 selected in step S117. If no voice input is detected, the process returns to step S110 and the above processing is repeated.

  In step S130, the input speech signal is recognized. Finally, in step S140, the recognized voice recognition processing result, that is, text information is sent to the other operation device.

  In the speech recognition apparatus according to the present embodiment, the speech signal to be converted is preprocessed and corrected to a speech signal that does not include speech distortion and has the same characteristics as the speech corpus that is the learning data of the acoustic model. Performance can be increased. In particular, the present embodiment can be applied to a mobile phone call.

  The embodiment described above is described in order to facilitate understanding of the present invention, and is not described in order to limit the present invention. Therefore, each element disclosed in the above embodiment is intended to include all design changes and equivalents belonging to the technical scope of the present invention.

It is a block diagram which shows 1st Embodiment of the preparation method of the acoustic model of this invention. It is a block diagram which shows 1st Embodiment of the speech recognition apparatus of this invention. It is a flowchart which shows the control procedure of 1st Embodiment of the speech recognition apparatus of this invention. It is a block diagram which shows 2nd Embodiment of the preparation method of the acoustic model of this invention. It is a block diagram which shows 2nd Embodiment of the speech recognition apparatus of this invention. It is a flowchart which shows the control procedure of 2nd Embodiment of the speech recognition apparatus of this invention. It is a block diagram which shows 3rd Embodiment of the preparation method of the acoustic model of this invention. It is a block diagram which shows 4th Embodiment of the speech recognition apparatus of this invention. It is a flowchart which shows the control procedure of 4th Embodiment of the speech recognition apparatus of this invention. It is a graph which shows the spectrum envelope of the audio | voice for demonstrating a formant frequency. It is a graph which shows the relationship between a Japanese vowel and a formant frequency. It is a figure which shows the relationship between the speech distortion in driving | running | working environment, and speech recognition performance. It is a graph for demonstrating the influence of the utterance distortion resulting from the tension | tensile_strength etc. under a driving environment. It is a block diagram which shows an example of the conventional speech recognition apparatus. It is a block diagram which shows an example of the creation method of the conventional acoustic model.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Undistorted speech corpus 2 ... Speech conversion filter means 3 ... Mobile environment speech corpus 4 ... Acoustic model 5 ... Speech input device 7 ... Decoder 8 ... Language model 9 ... Vehicle speed detection device 10 ... Selection device

Claims (26)

A method for creating an acoustic model for converting a voice signal into a label signal used in a voice recognition device that may be used in a mobile body,
Recording a predetermined sound in an environment substantially free of acoustic noise to create an undistorted sound corpus;
Correcting the speech of the undistorted speech corpus created in the step using speech conversion filter means that transforms the speech into speech corresponding to the moving speed of the moving object, and creating a mobile environment speech corpus;
Learning the mobile environment speech corpus created in the step as learning data to create an acoustic model, and creating an acoustic model. A method for creating an acoustic model for converting a voice signal into a label signal used in a voice recognition device that may be used in a mobile body,
Recording a predetermined sound in an environment substantially free of acoustic noise to create an undistorted sound corpus;
Learning the undistorted speech corpus created in the step as learning data to create an undistorted acoustic model;
Correcting the speech of the undistorted speech corpus using speech conversion filter means that transforms the speech into speech corresponding to the speed of the mobile body, and creating a mobile environment speech corpus;
Adapting the undistorted acoustic model using the mobile environment speech corpus created in the step, and creating an acoustic model. The mobile environment speech corpus is selectively corrected from the speech of the undistorted speech corpus so that the data capacity of the mobile environment speech corpus is smaller than the data capacity of the undistorted speech corpus. 2. A method for creating the acoustic model according to 2. Screening the sound of the undistorted speech corpus according to a predetermined category;
The acoustic model according to any one of claims 1 to 3, wherein the speech conversion filter unit corrects the speech corresponding to the category selected in the step and corresponding to the moving speed of the moving body. How to make. 5. The acoustic model according to claim 1, wherein the utterance conversion filter unit corrects the voice so as to increase the power of the vowel of the voice as the moving speed of the moving body increases. How to make. The acoustic model according to any one of claims 1 to 4, wherein the utterance conversion filter unit corrects the voice so as to increase a pitch frequency of a vowel of the voice as the moving speed of the moving body increases. How to create 5. The speech conversion filter means corrects the power of the voice band frequency so as to increase the slope of the spectrum regression line of the voice vowel as the moving speed of the moving body increases. A method for creating an acoustic model according to any one of the above. The speech conversion filter means corrects the sound so as to increase at least one of the first to third formants of the sound as the moving speed of the moving body increases. A method for creating the acoustic model described in the above. 5. The utterance conversion filter means corrects the voice so as to increase the duration of the vowel at the beginning of the vocabulary spoken as the moving speed of the moving body increases. A method for creating the acoustic model described in 1. The utterance conversion filter means corrects the voice so as to increase the duration of the vowel at the end of the uttered vocabulary as the moving speed of the moving body increases. A method for creating the acoustic model described in 1. A speech recognition method that may be used in a mobile body,
Detecting a moving speed of the moving body;
Inputting the speech to be recognized;
Converting the audio signal input in the step into a label signal using the acoustic model created by the method according to any one of claims 1 to 10 according to the detected moving speed of the moving body; A speech recognition method comprising: A speech recognition device that may be used in a mobile body,
Speed detecting means for detecting the moving speed of the moving body;
Voice input means for inputting voice to be recognized;
Storage means for storing an acoustic model created by the method according to any one of claims 1 to 10,
Conversion means for converting an audio signal input to the audio input means into a label signal using an acoustic model stored in the storage means in accordance with the moving speed of the moving body detected by the speed detection means. A speech recognition apparatus characterized by that. A speech recognition device that may be used in a mobile body,
Speed detecting means for detecting the moving speed of the moving body;
Voice input means for inputting voice to be recognized;
Storage means for storing an acoustic model in which an audio signal and a label signal are associated;
The audio signal input to the input unit is acoustically the same as the audio corpus used when learning the acoustic model stored in the storage unit according to the moving speed of the moving body detected by the speed detecting unit. Speech correction filter means for correcting the speech signal having statistical characteristics;
A speech recognition apparatus comprising: a conversion means for converting the speech signal corrected by the speech correction filter means into a label signal using an acoustic model stored in the storage means. 14. The speech recognition apparatus according to claim 13, wherein the utterance correction filter unit corrects the speech signal so as to decrease the power of the vowel of speech as the moving speed of the moving body increases. The speech recognition apparatus according to claim 13, wherein the speech correction filter unit corrects the speech signal so as to decrease the pitch frequency of the vowel of the speech as the moving speed of the moving body increases. The utterance correction filter unit corrects the power of the voice band frequency so as to decrease the slope of the spectrum regression line of the vowel voice as the moving speed of the moving body increases. Voice recognition device. 14. The speech correction filter unit corrects the speech signal so as to decrease at least one of the first to third formants of the speech as the moving speed of the moving body increases. Voice recognition device. The speech according to claim 13, wherein the speech correction filter means corrects the speech signal so as to reduce the duration of the vowel at the beginning of the spoken vocabulary as the moving speed of the moving body increases. Recognition device. The speech according to claim 13, wherein the speech correction filter means corrects the speech signal so as to reduce the duration of the vowel at the end of the spoken vocabulary as the moving speed of the moving body increases. Recognition device. A speech recognition method that may be used in a mobile body,
Detecting a moving speed of the moving body;
Inputting the speech to be recognized;
Storage means in which an acoustic model is stored;
The same acoustic and statistical characteristics as the speech corpus used when learning the acoustic model in which the speech signal and the label signal are associated with the speech signal input in the step according to the detected moving speed of the moving body Correcting to an audio signal having
Converting the voice signal corrected in the step into a label signal using the acoustic model. 21. The speech recognition method according to claim 20, wherein in the correcting step, the input speech signal is corrected so that the power of the vowel of the speech is decreased as the moving speed of the moving body increases. 21. The speech recognition method according to claim 20, wherein in the correcting step, the input speech signal is corrected so as to decrease the pitch frequency of the vowel of the speech as the moving speed of the moving body increases. . In the correcting step, the power of the voice band frequency of the input voice signal is corrected so as to reduce the slope of the spectrum regression line of the voice vowel as the moving speed of the moving body increases. The speech recognition method according to claim 20. 21. In the correcting step, the input audio signal is corrected so as to decrease at least one of the first to third formants of the audio as the moving speed of the moving body increases. The speech recognition method described in 1. The input speech signal is corrected in the correcting step so that the duration of the vowel at the beginning of the spoken vocabulary is decreased as the moving speed of the moving body increases. The speech recognition method described. The input speech signal is corrected in the correcting step so that the duration of the vowel at the end of the spoken vocabulary is decreased as the moving speed of the moving body increases. The speech recognition method described.

Images