JPH09134193A - Speech recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make it possible to recognize the place-names throughout Japan and to deal with a command response under high noise as well by switching speech recognition programs according to the environment to be used. SOLUTION: Speech characteristic patterns are extracted from speech signals inputted from a speech input section 10 in a speech recognition section 10 and the speeches are recognized in accordance with these speech characteristics. A first storage section 14 stores the vocabulary to be recognized and/or the standard patterns for recognition and/or the vocabulary groups to be recognized of different application objects and/or the standard pattern groups for recognition of different application objects. A second storage section 20 stores the plural speech recognition programs 121 to be loaded into the speech recognition section 12. The device includes a sensor for detecting external input signal 40 and an interface section 41. The plural speech recognition programs 121 are dynamically and selectively used according to the condition of the input signals 40 from the sensor.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device, and more particularly to a voice recognition device that inputs voice and outputs characters and other recognition results.

[0002]

2. Description of the Related Art Various researches and developments have hitherto been made on a voice recognition device for operating an electric device or other device by inputting a voice instead of operating it by a human hand. Ideally, the voice recognition technology is capable of recognizing a voice of arbitrary length, which a human utters from an arbitrary place at an arbitrary timing, with a probability of 100%. However, under actual usage conditions, there is noise, and if we try to capture speech that is uttered at any time, the beginning of the speech will be detected from the signal section that is also included in the speech during the speech input processing. It is inevitable that a large amount of calculation will be required because a complicated algorithm for detecting the noise and the end many times and excluding only the noise must always be executed. For this reason, the voice recognition technology requires a method and an algorithm for efficiently executing a complicated algorithm, and Japanese Patent Application No. 7-151698 and many other proposals have been made by the inventors of the present application. There is. A prior example of this voice recognition device will be described with reference to FIG.

In FIG. 3, the voice recognition device includes a voice input unit 1P including a microphone and other acoustoelectric converters for converting voices into voice signals, a waveform conversion unit 2P for converting voice waveform data into digital numerical values, and voice waveforms. From a voice feature extraction unit 3P that extracts a feature for voice recognition from a voice switch, a start switch unit 4P that gives a start edge detection start trigger at the time of voice segment detection for voice recognition, and a voice feature amount obtained from the voice feature extraction unit. A voice section detection unit 5P that determines only one start and end of the voice, and an input pattern storage unit 6P that takes in the voice feature amount from the voice start end to the end determined by the voice section detection unit and sets it as an unknown input pattern, Stored in a standard pattern storage unit 8P that stores a plurality of voice patterns for recognition with label names and an input pattern storage unit The unknown input voice pattern and the standard pattern stored in the standard pattern storage unit are calculated for the similarity, and the result is the distance value to the input voice pattern, for example, the Mahalanobis distance. A pattern matching unit 10P that outputs a distance value on the feature amount, and a distance comparison unit that determines the standard pattern having the smallest distance value among the distance values to the unknown input voice pattern output for each standard pattern. 1
1P, a result output unit 12P for transmitting the label name of the standard pattern determined to have the smallest distance value in the distance comparison unit to the upper host or the system bus.

In such a voice recognition device, a high-performance microprocessor or a CPU such as a digital signal processor is provided at the center of the device because it is necessary to give necessary and sufficient calculation ability to it. (See, for example, JP-A-7-140998). Using a high-performance CPU for other purposes is advantageous because it lowers the cost of using an application program according to the purpose, and therefore the form of the speech recognition apparatus as shown in FIG. 2 is adopted. Here, the message processing unit 11 causes the voice recognition unit 1 to
It is assumed that it is composed of a CPU forming a part of 2 or a CPU arranged independently. The voice recognition unit 12 of the voice recognition device 1
The input voice signal is recognized, and the message meaning of the result is interpreted by the message processing unit 11,
A program is transferred from the application program group 2 provided in the vicinity of the voice recognition device to the voice recognition device according to a message. At this time, it is extremely effective to equip the voice recognition device with an application program management table 13 that manages which application program the voice recognition device uses and can use. It is also possible to utilize a high-performance CPU forming the message processing unit 11 to execute a plurality of application programs 2 simultaneously as a multitask. While it is possible to operate the voice recognition device with such high performance, there is still room for improvement as described below in the performance of the voice recognition unit 12.

[0005]

In order to satisfy the above-mentioned demand for the voice recognition device "to recognize any time, anybody's voice, any kind of content", calculation by a super-large general-purpose computer is required. Need. If the number of vocabularies is limited to tens of thousands or less, a computer having a workstation-level calculation function will suffice, but as is obvious from the size and weight of the workstation, this is not portable at all. And, in terms of cost, a cost of several hundred thousand yen to several million yen is currently required. This is the case when it is attempted to support all voice recognition by one excellent voice recognition device. Also,
For voice recognition under noisy environment, if a simple low-cost voice recognition device is constructed without using adaptive noise reduction technology, the number of target vocabulary is reduced to within a few words, which is a practical constraint. .

Such a problem of the voice recognition device is caused by the inability of the voice recognition program to cope with the environment in which the device is placed. This point will be specifically described. For example, a speech recognition program based on the well-known Hidden Markov Model HMM can target the recognition vocabulary of 1,000 to tens of thousands. For example, a surplus word such as "E ~" in "E ~ Sapporo" has an advantage of showing a high recognition rate even if it is attached before and after the recognition target vocabulary. On the other hand, there is also a point to be improved that it takes time and effort to create a dictionary model for recognition. And, for example, the SPLIT method (Kanmura, Furui:
Large Vocabulary Word Speech Recognition Using Pseudophonic Standard Patterns ", Theological Theory, J65-D, 8, pp. 1041-1048 (Sho 5
7)) can perform voice recognition with a smaller calculation amount and storage capacity than the HMM. On the other hand, it is difficult to deal with a large vocabulary as compared with the HMM because the standard pattern is characterized. In addition, as a simple voice recognition technique, an SA that compares the standard pattern with the autocorrelation coefficient after reducing the order, which is generally 10th order or more, to about 8th order
A program of the DP (Staggered Array DP) method is known. In this method, the amount of calculation and the amount of memory are about 1/10 of those of general speech recognition, and a quick response is possible, but the number of recognized vocabularies is limited to within 20 words. Further, the speech recognition technology incorporating adaptive noise reduction effective in a noisy environment is excellent in noise resistance, but has a drawback that the amount of calculation and the amount of storage increase by several tenths.

As described above, it can be said that there is no program for all situations as a speech recognition program, and it is emphasized from the viewpoint of the number of recognition vocabulary, noise resistance, small economy of computing ability and other requirements. It is the current situation that an algorithm or program suitable for the request is selected and installed in the voice recognition device. The present invention provides a voice recognition device that solves the above problems.

[0008]

A voice recognition unit having a voice input unit for inputting a voice signal, extracting a voice feature pattern from the input voice signal, and recognizing voice based on the voice feature pattern information. And a vocabulary group for recognition and / or a standard pattern for recognition, and / or a vocabulary group for recognition and / or a standard pattern group for recognition differently applied. The voice recognition apparatus is configured to include the second storage unit 20 that includes the storage unit 14 and stores the plurality of voice recognition programs 121 to be loaded in the voice recognition unit 12.

Then, in the above speech recognition apparatus, a speech recognition apparatus having a recognition target table 1211 attached to the speech recognition program 121 as a standard pattern is constructed. Further, in the above speech recognition apparatus, a sensor for detecting the external input signal 40 and the interface unit 41 are provided, and the speech recognition apparatus is configured to switch a plurality of speech recognition programs according to the detected external input signal.

[0010]

BEST MODE FOR CARRYING OUT THE INVENTION According to the present invention, a plurality of voice recognition programs used for voice recognition are installed in a voice recognition device in the same manner as a stored program type computer, and the plurality of programs are provided according to the situation of an input signal from a sensor. This is a voice recognition device configured to perform recognition with a configuration that is not recognized in a computer of a stored program system that is dynamically used properly. As a means to satisfy the request imposed on the voice recognition device, "any time, anyone's voice, any content can be recognized",
The present invention addresses this demand by loading an appropriate recognition program according to the situation from the program storage unit into the voice recognition unit and reading it.

[0011]

DESCRIPTION OF THE PREFERRED EMBODIMENTS First, the outline of an embodiment of the present invention will be described with reference to FIG. In FIG. 1, the left part divided by a dotted arrow 1 indicates a voice recognition device. The voice input unit 10 of the voice recognition device 1 is for receiving voice and converting the voice into a voice signal, and is composed of, for example, an audio microphone, a digital signal input terminal for receiving voice waveform data, and the like. The voice input unit 10
The sound input to may be a mechanical sound, a synthetic sound of a computer, a cry of an animal, or other non-speech, but for convenience of explanation, these will be collectively described as voices.

The voice recognition unit 12 recognizes an input voice signal recorded via the voice input unit 10 and interprets a message indicated by the recognition result by a central processing unit (CPU) 111. The voice recognition unit 12 is generally composed of a DSP capable of executing a voice recognition algorithm at high speed, but may be a microprocessor similar to a CPU. Then, the CP of the voice recognition unit 12
A configuration using the CPU 111 as U can also be adopted. In this case, since it is necessary to simultaneously control the main body of the voice recognition device 1 and the arithmetic processing control of the voice recognition, the CPU
What has a multitasking function is used as 111.

The voice recognition device 1 and the CPU 111 are provided with a first storage section 14 in order to efficiently execute voice recognition or commands. The first storage unit 14 is a recognition dictionary / standard pattern 15 required for voice recognition, and an application program management table for managing which application programs the voice recognition device 1 uses and which ones can be used. 13 and a voice recognition program management table 16 for managing which voice recognition program the voice recognition device 1 uses or can use.

The voice recognition device 1 further includes a second storage unit 2.
0 is provided. The reason why the first storage unit 14 and the second storage unit 20 are both provided in this manner is to improve the utilization efficiency of the storage units. In general, D
A semiconductor memory such as a RAM or an SRAM has an advantage that it is small in size and responds at high speed as a storage unit. On the other hand, it has the drawback of being relatively expensive. A memory such as a CD-ROM or a digital video disk DVD can be used as a target memory for these semiconductor memories. These advantages and disadvantages are the disadvantages of the semiconductor memory,
The disadvantage is the advantage. For this reason, a device having a computer element generally uses a semiconductor memory together as a storage device. The present invention also includes, for example, a cache memory formed of a semiconductor memory as the first storage unit 14 and a large-capacity CD-ROM as the second storage unit 20.
A configuration including an external memory including a ROM is adopted.
At this time, in the second storage unit 20 having a large storage capacity, the voice recognition programs 121 to 121x and the recognition target table 121 attached to the voice recognition program 121 are stored.
1 is stored. The second storage unit 20 also stores a plurality of application programs 2 so that the voice recognition device can immediately use the requested program by understanding the command instructed by voice. It

In the voice recognition apparatus 1, tables or programs stored in the first storage unit 14 and the second storage unit 20 in a distributed manner are exchanged, or the voice recognition unit 12 and the CPU 111 are replaced. Loading is always done. This operation is performed via the memory management unit MMU30. As shown in FIG. 1, the MMU 30 may be configured as a module different from the CPU 111, or may be configured by utilizing a part of the functions of the CPU 111 or the DSP which is the core of the voice recognition unit 12.

The voice recognition apparatus 1 uses a bus which is a high-speed transmission path of data information in order to call the stored contents, the recognition result or the application program at high speed between the above-mentioned storage unit, voice recognition unit and CPU. To have. Reference numeral 31 is a first bus provided in the first storage unit 14, and 32 is a second bus provided in the second storage unit 20. Additional buses can be provided if desired. It is also possible to connect the first bus 31 and the second bus 32. These buses include a voice recognition unit 12, a CPU 111, and an MMU3.
0, the first storage unit 14 or the second storage unit 20 is provided with a route for passing data or command information, and each is used efficiently so as not to cause a failure in the operation of the voice recognition device. The above MMU 30 creates a state in which the above can be performed.

As described above, the performance of the voice recognition device can be comprehensively improved by changing the voice recognition program according to the usage environment of the voice recognition device. The configuration of the voice recognition device of the present invention is slightly more complicated than that of the conventional voice recognition device. Further, since it is necessary to develop a plurality of voice recognition programs, this causes an additional cost increase. However, since a large number of voice recognition devices according to the present invention are provided to the world, the development cost is proportionally distributed, and the cost burden of each voice recognition device becomes insignificant. To increase the capacity of the storage unit that stores multiple programs, a low-cost CD-ROM unit is used to store the programs all the time and a semiconductor memory that can be accessed at high speed during recognition is temporarily used (cache memory and It is possible to suppress an effective increase in cost by using this as a "name".

The present invention can improve the efficiency of voice recognition by incorporating into a voice recognition device a general sound pressure including voice, a signal resulting from vibration and acceleration, a GPS signal and other external input signals. That is, these general external input signals 40 are detected by the sensor and interface unit 41 included in the voice recognition device, and are supplied to the bus 32 via the interface unit. The signal obtained through the sensor and the interface unit 41 is transmitted to the CPU 111 and used for judging the preset environmental condition.
Hereinafter, the case where the invention is applied to a car navigation device will be described as an example.

As is known, a car navigation system first sets a destination. If the name of a place in Japan is to the level of a town or village, it is necessary to recognize one place out of about 200,000. Whether or not the setting is before departure is determined by determining whether or not it is the first command after the engine key is inserted, and whether or not the G is included in the car navigation device.
It can be determined by detecting the time change of the position information signal from the PS sensor. If the position information does not change due to the GPS signal, it means that the vehicle is stationary. The movement of the vehicle can also be determined from the signal of an acceleration sensor that detects acceleration. When setting the first destination of the car navigation system, the only problem in the operating environment is the engine sound of the car, and road noise or wind noise, which accounts for most of the car's noise, is not a problem. . In such an environment, it is a good idea to load a program that has a higher recognition rate than noise resistance or can recognize a large number of vocabulary words into a voice recognition device and operate it.

Next, assuming that the vehicle is traveling, whether or not the vehicle is traveling is detected by detecting an acceleration signal obtained from the above-described acceleration sensor and a vehicle speed signal obtained by converting the rotational speed of the tire generated by the vehicle into a pulse signal. Can be determined by
In such a situation, instead of the difficult operation / command to select one vocabulary from a huge recognition target such as place name recognition, instead of "where is now?", "What is next?", "Map expansion", etc. It is sufficient that the limited commands can be recognized, and the number of vocabularies to be recognized at one time is limited to the number of vocabularies of several tens or less. In an environment where road noise or wind noise is noticeable during running, it is a good idea to load a recognition program for a small number of vocabulary objects, which emphasizes noise resistance, into a voice recognition device to operate.

The speech recognition unit 1 in the embodiment shown in FIG.
2 can be configured by, for example, a voice recognition device described in Japanese Patent Application No. 7-151698. This will be described with reference to FIG. 3. In FIG. 3, (1) the waveform conversion unit 2P for converting the voice data obtained from the voice input unit 1P into a digital numerical value is provided. This waveform conversion unit is, for example, a process for converting an analog voice waveform into digital data, and ADPC for voice.
A process of receiving as compressed data such as M and converting it into linear data is also included.

Then, (2) a voice feature extraction unit 3P for detecting a voice section from the voice waveform data obtained by the waveform conversion unit 2P and extracting a feature amount used for voice recognition is provided. As the analysis method of the voice feature extraction unit, a short-time logarithmic power analysis, a cepstrum analysis, and other analysis methods well known in the voice recognition technology are adopted. The details of the voice recognition program, the algorithm, and the protocol with other modules are stored in the first storage unit 14 or the second storage unit 20, and are loaded into the voice recognition unit 12 as necessary to be used. To fulfill the function of.

Further, (3) a voice section detecting unit 5P for identifying a voice start end and a voice end from a voice feature amount obtained from the voice feature extraction unit 3P is provided. As a method to detect the voice section, measure the noise level before voice utterance,
It is possible to use a method in which a section in which a signal component having a logarithmic power value equal to or more than a certain threshold value changes within a certain time period as a speech section is compared with the noise level. The noise level can be detected by using the microphone itself included in the voice recognition device, but it is more practical to input the noise level separately due to the frequency characteristic or directivity of the microphone. Generally, a unidirectional microphone is suitable for the input section, and an omnidirectional microphone is suitable for the noise detection section.

Further, (4) there is provided an activation switch section 4P for giving a trigger for starting the start edge detection at the time of detecting the voice section in the voice recognition processing. As the activation switch, a voice switch may be used to automatically capture the time when the user utters, or the user may press the press talk button when uttering.

Here, (5) there is provided an input pattern storage unit 6P which takes in the voice feature amount from the voice start end to the voice end determined by the voice section detection unit and stores it as an unknown input pattern. Then, (6) a standard pattern storage unit 8P that stores a plurality of voice standard patterns that have been analyzed and stored by the same procedure as the procedure for storing the unknown input pattern in the input pattern storage unit, and have label names. It is equipped with. This standard pattern information also includes voice section information corresponding to that detected by the voice section detection unit.

(7) A pattern matching unit 10P is provided for calculating the degree of similarity between the unknown input voice pattern stored in the input pattern storage unit and each standard pattern stored in the standard pattern storage unit. . The calculation of the similarity is performed using DP matching, for example. Furthermore,
(8) The distance comparison unit 11P that accumulates the calculation results of the respective degrees of similarity and determines the standard pattern that minimizes the difference between any standard pattern and the unknown input voice pattern is provided.

Further, (9) the result comparing section 12P is provided for outputting the label name of the standard pattern determined to be the most similar in the distance comparing section to the CPU 111 which is the upper host of the voice recognizing section 12. In the above-mentioned embodiment, it is usual that a standard pattern that has been analyzed and prepared in advance is already registered as the standard pattern. That is, the registered standard pattern is the recognition dictionary / standard pattern 15 in FIG. 1, and as a small-scale standard pattern, the recognition target table 1 attached to the voice recognition program 121.
It can also be used in the form of 211. In the present invention, in order to adapt the voice recognition device to its usage environment, or to use the voice recognition device to make an efficient response in response to a user's request, a recognition target vocabulary, a standard pattern for recognition, Alternatively, both of them, a recognition target vocabulary group having a different application target, a recognition standard pattern group having a different application target, or both of them are stored in a storage unit as necessary. There are situations such as the SPLIT method or the voice recognition method using the HMM, and other situations suitable for each program. If necessary, the voice recognition unit 1 can quickly obtain program information via a bus.
A mechanism to transfer to 2 is required.

Next, in the present invention, a voice recognition program by DP matching and a simple voice recognition program for performing voice recognition only by an 8th order autocorrelation coefficient are installed in a voice recognition device, and a recognition experiment is carried out depending on the situation. The result of performing is explained. The vocabulary to be recognized is based on the literature “Acoustic Society Proceedings,
Common speech data for voice recognition "(author Itabashi, published in 1985), the top 60 words and commands in 100 words of Japanese city names (large: enlarge map, small: reduce map,
End: The end of the command, and the other 10 words) were uttered by four male speakers in an environment with noise levels of 65 dB and 75 dB. The sound is 300Hz to 3.4.
Converted at 8 kHz through a kHz filter, and 128 for a voice recognition program using DP matching.
A short LPC cepstrum analysis was performed every msec. The voice section was detected with logarithmic power for a short time. In the method of detecting the voice start point, when the signal power value continues from a state in which there is no voice at a large value of a certain threshold value or more for a predetermined time, the rising position of the signal power value is set as the start point. After that, the voice section detection unit detects the attenuation point of the signal power value of the voice and sets it as the end of the voice. For the simple voice recognition program, the pre-processing stage of the voice signal has the same condition as that of the DP matching program, and the comparison of the similarities is performed according to the procedure of the program itself. Calculation of similarity in speech recognition program by DP matching is fixed at the beginning
It is a termination-free Staged Array DP.

Here, in the case of voice recognition by the DP matching program, the erroneous recognition rate was 5% when the noise level was 65 dB, whereas in the environment of 75 dB where the noise level was increased by 10 dB, The misrecognition rate increased to about 40%, which was expected to be a problem in practical use. It is expected that similar results will be obtained in the case of voice recognition by the SPLIT method or HMM.

On the other hand, in the case of voice recognition by the simple voice recognition program, it is possible to use 6 commands for the 10 defined commands.
The misrecognition rate remained within 2% under the noise of 5 dB, and was within 10% even under the noise of 75 dB. The relatively high recognition rate under the noise of the voice recognition by the simple voice recognition program shows that the recognition dictionary / standard pattern 15 and the recognition target table 1 shown in FIG.
When creating 211, the effect of having a vocabulary structure that is easily distinguishable from each other is also included. This is also one effect of the present invention in which these pieces of information are dynamically switched and used.

Based on these results, a prototype car navigation device with a voice input function, in which the voice recognition device performs the following functions when the engine key is inserted and the accessories of the vehicle are energized, is produced with good results. I was able to get 1. Initialize car navigation system with voice input function and load SPLIT voice recognition program.

2. Purpose setting before departure by voice input 3. Start navigation by detecting the moving speed of the car with the acceleration sensor At the same time, switch the main program of the voice recognition unit from the previous one to the simple voice recognition program. 4. 4. Wait for input of response command, enlargement, reduction, ... 5. Support for loading application programs and screen operations according to user commands. Automatic termination of car navigation system with voice input function by removing engine key

[0033]

As described above, the present invention employs a configuration in which a voice recognition program is switched according to the environment in which the voice recognition device is used, and the voice recognition is operated only by a single program. Depending on the device, it is difficult to realize, for example, a dual-purpose voice recognition device capable of recognizing a place name nationwide in Japan and also capable of responding to a command response under high noise. By using the acoustic sensor of the voice recognition device for switching the program, a natural voice recognition service can be provided without burdening the user.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating an embodiment.

FIG. 2 is a block diagram illustrating a conventional example of a voice recognition device.

FIG. 3 is a block diagram illustrating a preceding example of a voice recognition unit.

[Explanation of symbols]

 1 voice recognition device 10 voice input unit 12 voice recognition unit 111 CPU 13 application program management table 14 first storage unit 15 recognition dictionary / standard pattern 16 voice recognition program management table 20 second storage unit 30 MMU 31 first bus 32 second bus 40 external input signal 41 sensor and interface section 121 speech recognition program 1211 recognition target table

Claims (3)

[Claims] 1. A voice recognition unit, comprising: a voice input unit for inputting a voice signal; extracting a voice characteristic pattern from the input voice signal; and recognizing a voice based on the voice characteristic pattern information. Target vocabulary and / or standard pattern for recognition, and / or recognition target vocabulary group and / or different application target
Alternatively, it is provided with a first storage unit for storing a recognition standard pattern group having different application targets, and a second storage unit for storing a plurality of voice recognition programs to be loaded into the voice recognition unit. And a voice recognition device. 2. The voice recognition device according to claim 1, further comprising a recognition target table attached to a voice recognition program as a standard pattern. 3. The voice recognition device according to claim 1, further comprising a sensor for detecting an external input signal and an interface unit, wherein a plurality of voices are detected according to the detected external input signal. A voice recognition device comprising a configuration for switching a recognition program.