TWI470387B - Wireless voice control system - Google Patents

Description

Wireless voice control system

The present invention relates to a control system, and more particularly to a wireless voice control system.

With the advancement of technology, users are becoming more and more dependent on electronic products, and the interaction with electronic products is becoming more frequent, which has led to the development of human-computer interaction interface.

Human-computer interaction technology can be roughly divided into two different aspects: contact and non-contact. Among them, there are touch technology known to the public in the contact technology part, and voice recognition and gesture recognition technology in the non-contact technology part. The contact and non-contact human-computer interaction technologies emphasize the human-centered approach, replacing traditional habits with keyboard and mouse communication.

However, voice has always been the most important means of communication between people and people. From the perspective of human-computer interaction, it is also the most convenient and practical input standard. The development time of the speech recognition field has gone through 20 to 30 years of evolution history, and the fields involved are also very extensive, and the research methods are also very diverse.

According to the development time of speech recognition technology, the earliest technology is Dynamic Time Warping (DTW), which compares speech with dynamic sample algorithm and sample speech to find the closest speech. The disadvantage is that the amount of calculation is large and the recognition time is slow. Later, the Artificial Neural Network (ANN) was proposed for speech recognition. Since the neural network architecture cannot be arbitrarily changed after training, the user's speaking speed will affect its speech recognition rate. In view of this, The shortcomings of the above traditional speech recognition technology are necessary to find a solution.

Accordingly, it is an object of the present invention to provide a wireless voice control system.

Thus, the wireless voice control system of the present invention comprises a voice control device and a passive control device.

The voice control device includes an input module, a voice recognition module, and a first communication module. The input module is configured to provide a user with voice input to generate a corresponding voice to be recognized. The voice recognition module is coupled to the input module for recognizing a voice keyword from the to-be-identified voice to generate a corresponding recognition result. The first communication module is coupled to the voice recognition module for transmitting the identification result in a wireless manner.

The passive control device includes a housing, a second communication module, a processing module, a mobile module, and a sensing module. The second communication module is disposed in the housing for receiving the identification result transmitted by the first communication module. The processing module is disposed in the housing and coupled to the second communication module for converting the identification result into a first control command. The mobile module is disposed in the housing and coupled to the processing module for driving the housing to move according to the first control command. The sensing module is disposed on the housing and coupled to the processing module to sense whether the housing collides with an obstacle in the surrounding area, and when the housing is sensed to collide with an obstacle nearby, The processing module is triggered to generate a second control command, and then the mobile module drives the housing to turn or stop according to the second control command.

The above and other technical contents, features and advantages of the present invention will be apparent from the following detailed description of the preferred embodiments.

Referring to FIG. 1, a preferred embodiment of the wireless voice control system of the present invention includes a voice control device 1, and a passive control device 2. In the preferred embodiment, the voice control device 1 controls the passive control device 2 by means of wireless remote control, and the passive control device 2 is a cleaning robot applied in a smart home, but is not limited thereto. Can be a remote control model, such as a remote control car, a remote control aircraft or a remote control boat.

The voice control device 1 includes an input module 11, a voice recognition module 12, a first communication module 13, and a display module 14.

The input module 11 is configured to allow a user to perform voice input to generate a corresponding voice to be recognized. In the preferred embodiment, the input module 11 can be, for example, a microphone that converts the analog voice signal input by the user into a digital voice signal, and then the voice to be recognized after being detected by the voice endpoint. . The voice endpoint detection method uses an energy detection method, and the calculation method is well known to those skilled in the art, and thus will not be described herein.

The voice recognition module 12 is coupled to the input module 11 for recognizing a voice keyword from the voice to be recognized to generate a corresponding recognition result. In the preferred embodiment, the voice recognition module 12 has a noise cancellation unit 121, a voice pre-processing unit 122, and a feature calculation unit 123.

In the preferred embodiment, the voice keyword is used to enable the passive control device 2 to perform corresponding actions, such as "forward", "backward", "stop", "left turn", "right turn", etc. The action, but not limited thereto, the voice keyword also increases the vocabulary as the action that the passive control device 2 can perform in the future, and the display module 14 is configured to display the voice keyword corresponding to the voice keyword. Identify the results.

The noise cancellation unit 121 is configured to decompose the to-be-identified speech having the noise portion and the user voice portion into the decomposed speech to be recognized according to an Empirical Mode Decomposition (EMD), wherein the decomposition is performed. The latter speech to be recognized has a plurality of sets of Intrinsic Mode Functions (IMF), and the detailed implementation method is well known to those skilled in the art, and therefore will not be described here.

It is worth mentioning that in the preferred embodiment, the optimal combination parameters of the first five essential modal functions (IMF1~IMF5) are calculated by the Real Genetic Algorithm (GA), and The optimal combination parameters add up the IMF1~IMF5 weights to generate a recombined voice signal, thereby achieving the effect of separating the excess noise components in the speech to be recognized.

Among them, the real-type gene algorithm parameters are set as follows: 1. The number of chromosomes per generation is 16; 2. Each chromosome has 5 real numbers, corresponding to IMF1~IMF5; 3. The chromosome survival rate is 0.5 and the mutation rate is 0.05; and 4. The mating method is random pairing, and the evolutionary algebra is 500 generations.

Similarly, the detailed implementation is well known to those skilled in the art and will not be described herein.

The voice pre-processing unit 122 performs the following voice pre-processing procedures, such as pre-emphasis, Hamming Window, Fast Fourier Transform (FFT), etc. Etc. and further utilizing Mel-Frequency Cepstral Coefficients (MFCCs) to perform feature extraction from a plurality of sound boxes of the reconstructed speech signal to generate a feature vector group corresponding to each of the sound frames. To perform a subsequent speech recognition process.

It is to be noted that, in other preferred embodiments of the present invention, for example, in a quiet and noise-free environment, the voice recognition module 12 may only have the voice pre-processing unit 122 and the feature calculation unit 123. . The speech pre-processing unit 122 performs the following speech pre-processing procedures, such as pre-emphasis, Hamming window, fast Fourier transform, etc., and further utilizes the Mel cepstral coefficient, from the speech to be recognized. Feature capture is performed in each of the frames to generate a set of feature vectors corresponding to each of the frames, and then the same subsequent speech recognition process is performed with the speech recognition module 12 having the noise cancellation unit 121. The feature calculation unit 123 uses Discrete Hidden Markov Models (DHMM) as a statistical model, and recognizes the voice keyword from the feature vector groups.

In addition, before performing the speech recognition process, the discrete hidden Markov model must be trained in advance, and in the preferred embodiment, 8 males and 2 females with different speaking speeds are directed to the following speech keywords, "forward" ," Backward, Stop, Left Turn, and Right Turn are recorded 10 times, and the 500 unvoiced voice files generated are used as training corpus.

The training step is similar to the speech pre-processing program executed by the speech pre-processing unit 122; firstly, the training corpus after the endpoint detection is used to perform pre-emphasis, Hamming window, fast Fourier transform, etc.; The Mel Cepstral Coefficient obtains the feature vector set of each of the frames in the training corpus; and then, by Fuzzy Vector Quantization (FVQ) combined with the Viterbi Algorithm, from the feature vector groups The optimized observation sequence is obtained, and the optimized observation sequence is substituted into the discrete hidden Markov model for training until the discrete hidden Markov model parameters converge, and then the whole speech training process is completed.

After the discrete hidden Markov model training is completed, in the preferred embodiment, the feature computing unit 123 performs fuzzy vector quantization on the set of feature vectors, and brings the quantized feature vector groups into the trained A discrete hidden Markov model is used to identify the speech keyword.

Therefore, after adding different background noises such as TV environment, air-conditioning environment, kitchen, vocals, etc., 8 males and 2 females each having the same speaking speed are sequentially directed to the same voice keyword of the microphone of the input module 11 Input 10 times, so that for each voice keyword, there are 10×10, a total of 100 sets of test data. Finally, the average speech recognition rate after the test data and its comparison are shown in Tables 1, 2 and 3, Table 1 The experimental result of the voice recognition module 12 does not have the function of the noise cancellation unit 121; the second is that the voice recognition module 12 has the experimental result of the noise cancellation unit 121; and the third table is the second phase of the table. Compared with Table 1, the process of improving the speech recognition rate degree.

It can be clearly seen from Table 3 that the noise cancellation unit 121 with the noise cancellation function presented in Table 2 clearly wins the speech recognition rate without the noise cancellation function; In a quiet background, such as in an air-conditioned environment, the speech recognition rate presented in Table 2 is less obvious than that in Table 1.

The first communication module 13 is coupled to the voice recognition module 12 for transmitting the identification result in a wireless manner.

In the preferred embodiment, the first communication module 13 is a TWS-BS wireless transmission module having a frequency of 315 MHz, an operating voltage of 3 to 12 V, and a transmission distance of 80 m to 120 m, and is modulated by amplitude ( The Amplitude-Shift Keying, ASK) method transmits the identification result. The TWS-BS wireless transmitting module can also be combined with the HT-12E encoding IC (not shown), and the identification result can be encoded before the identification result is transmitted.

Referring to FIG. 1 and FIG. 2 , the passive control device 2 includes a housing 20 , a second communication module 21 , a processing module 22 , a mobile module 23 , a sensing module 24 , and a charging module . 25.

The second communication module 21 is disposed in the housing 20 for receiving the identification result transmitted by the first communication module 13.

In the preferred embodiment, the second communication module 21 is a RWS-371 wireless receiving module, and has a receiving sensitivity of -117 dBm, and an operating voltage of 3.5 V to 5.5 V, and can be received and transmitted by amplitude modulation. To identify the results. The RWS-371 wireless receiving module can also be combined with the HT-12D decoding IC (not shown), and can decode the encoded identification result after receiving the encoded identification result.

The processing module 22 is disposed in the housing 20 and coupled to the second communication module 21 for converting the identification result into a first control command.

In the preferred embodiment, the processing module 22 is a microcontroller of the AT89S5X series manufactured by ATMEL, such as AT89S51, but is not limited thereto.

The mobile module 23 is disposed on the housing 20 and coupled to the processing module 22 for driving the housing 20 to move according to the first control command.

In the preferred embodiment, the mobile module 23 has a motor 231, but not limited thereto, the number of the motors 231 can be changed in quantity according to the actual operation of the passive control device 2, so the number of motors 231 Can also be more than one.

After the mobile module 23 receives the first control command, the motor 231 can perform actions such as "forward", "reverse", "stop", "left turn", "right turn", and the like.

The sensing module 24 is disposed on the housing 20 and coupled to the processing module 22 for sensing whether the housing 20 collides with a surrounding obstacle and senses that the housing 20 collides with the surrounding When the obstacle is triggered, the processing module 22 is triggered to generate a second control command, and then the mobile module 23 is in accordance with the second control The command drives the housing 20 to turn or stop.

In the preferred embodiment, the sensing module 24 can be an infrared sensor disposed around the housing 20, and when the sensing module 24 senses that the housing 20 collides with a surrounding obstacle, The sensing module 24 transmits a collision message to the first communication module 13 via the second communication module 21, so that the collision message is displayed on the display module 14.

The charging module 25 has a detecting unit 251 and a rechargeable battery unit 252. The detecting unit 251 is configured to detect the battery status of the rechargeable battery unit 252, and the battery status is controlled by the second communication module 21. The first communication module 13 is transmitted to the voice control device 1, and the battery state is displayed on the display module 14.

In summary, the present invention can control the passive control device 2 by means of the wireless voice remote control by the voice control device 1, but since the control effect depends on the recognition result of the voice, the present invention has a small amount of computational complexity. The hidden Markov model is used as the speech recognition model, while the discrete hidden Markov model is used for statistical recognition to solve speech problems. It can solve the problem of user's speaking speed and can also identify semantics. In addition, the present invention is supplemented by the noise canceling unit 121 before performing voice pre-processing and voice recognition, and can improve the voice recognition rate in different noise backgrounds, so that the object of the present invention can be achieved.

The above is only the preferred embodiment of the present invention, and the scope of the invention is not limited thereto, that is, the simple equivalent changes and modifications made by the scope of the invention and the description of the invention are All remain within the scope of the invention patent.

1‧‧‧Voice control device

11‧‧‧Input module

12‧‧‧Voice recognition module

121‧‧‧ Noise Elimination Unit

122‧‧‧Voice pre-processing unit

123‧‧‧Characteristic calculation unit

13‧‧‧First communication module

14‧‧‧Display module

2‧‧‧Passive control device

20‧‧‧shell

21‧‧‧Second communication module

22‧‧‧Processing module

23‧‧‧Mobile Module

231‧‧‧Motor

24‧‧‧Sensing module

25‧‧‧Charging module

251‧‧‧Detection unit

252‧‧‧Rechargeable battery unit

1 is a block diagram showing a preferred embodiment of a wireless voice control system of the present invention; and FIG. 2 is a schematic diagram showing a housing of a passive control device in the preferred embodiment.

1‧‧‧Voice control device

11‧‧‧Input module

12‧‧‧Voice recognition module

121‧‧‧ Noise Elimination Unit

122‧‧‧Voice pre-processing unit

123‧‧‧Characteristic calculation unit

13‧‧‧First communication module

14‧‧‧Display module

2‧‧‧Passive control device

21‧‧‧Second communication module

22‧‧‧Processing module

23‧‧‧Mobile Module

231‧‧‧Motor

24‧‧‧Sensing module

25‧‧‧Charging module

251‧‧‧Detection unit

252‧‧‧Rechargeable battery unit

Claims (9)

A wireless voice control system includes: a voice control device, including an input module, a voice recognition module, and a first communication module, wherein the input module is configured to provide a user with voice input to generate a And corresponding to the voice to be recognized, the voice recognition module is coupled to the input module, and configured to identify a voice keyword from the to-be-identified voice to generate a corresponding identification result, the first communication mode The group is coupled to the voice recognition module for transmitting the identification result in a wireless manner; and a passive control device includes a casing, a second communication module, a processing module, and a mobile module And a sensing module, the second communication module is disposed in the housing for receiving the identification result transmitted by the first communication module, the processing module is disposed in the housing and coupled to the The second communication module is configured to convert the identification result into a first control command. The mobile module is disposed in the housing and coupled to the processing module, and is configured to drive the housing to move according to the first control instruction. The sensing module is disposed on the shell And being coupled to the processing module, configured to sense whether the housing collides with an obstacle in the surrounding area, and trigger the processing module to generate a second control when sensing that the housing collides with an obstacle in the surrounding area Commanding, and then the mobile module drives the housing to turn or stop according to the second control command; wherein the voice recognition module has a voice pre-processing unit, and a feature computing unit, the voice pre-processing unit utilizes Meyer a cepstrum coefficient, performing feature extraction from a plurality of sound boxes of the to-be-identified speech to generate a feature vector group corresponding to each of the sound frames, and the feature calculation unit The speech keyword is identified from the set of feature vectors by using a discrete hidden Markov model. The wireless voice control system according to claim 1, wherein the feature computing unit performs fuzzy vector quantization on the set of feature vectors, and brings the quantized feature vector groups into the discrete hidden after training. Markov model to identify the speech keyword. The wireless voice control system according to claim 2, wherein the discrete hidden Markov model is trained by using fuzzy vector quantization combined with a Viterbi algorithm. The wireless voice control system according to claim 1, wherein the voice recognition module has a noise cancellation unit, a voice pre-processing unit, and a feature calculation unit, and the noise cancellation unit is used according to experience. The modal decomposition method decomposes the to-be-recognized speech, and recombines the decomposed speech to be recognized by a genetic algorithm to generate a recombined speech signal, and the speech pre-processing unit utilizes the Mel cepstral coefficient. Feature capturing is performed in a plurality of sound boxes of the recombined voice signal to generate a feature vector group corresponding to each sound box, and the feature computing unit uses the discrete hidden Markov model to identify from the feature vector groups The voice keyword is output. The wireless voice control system according to claim 4, wherein the feature calculation unit performs fuzzy vector quantization on the set of feature vectors, and brings the quantized feature vector groups into the discrete hidden after training. Markov model to identify the speech keyword. According to the wireless voice control system described in claim 5, The discrete hidden Markov model is trained by using fuzzy vector quantization combined with Viterbi algorithm. The wireless voice control system of claim 1, wherein the voice control device further includes a display module, and the passive control device further includes a charging module, the charging module has a detecting unit and a charging battery unit, the detecting unit is configured to detect a battery state of the rechargeable battery unit, and transmit the battery state to the first communication module of the voice control device by using the second communication module, thereby The battery status is displayed on the display module. The wireless voice control system of claim 1, wherein the voice control device further comprises a display module, and the display module of the voice control device is configured to display the identification result. The wireless voice control system of claim 1, wherein the mobile module of the passive control device has at least one motor.