TW201015538A - Intelligent speech recognition control device - Google Patents

Abstract

An intelligent speech recognition control device is disclosed, which comprises a speech acquisition component, a host, a signal transmitting component and a control component. The host comprises an operator and a digital information storage unit, and the digital information storage unit is pre-stored with at least one speech command and a speech recognition processing software. The speech acquisition component receives a speech message. The operator of the host executes the speech recognition processing software to process the speech message by means of a normalized fuzzy logic Kalman filter process, an end-point detection and frame process, a cepstrum analysis, a Mel-scale process, a filter bank process, a Mel-frequency cepstral coefficient process and a dynamic time warping algorithm, so as to compare whether the speech message matches the speech command. After the speech message is confirmed to match the speech command, the signal transmitting component transmits an activation signal to the control component to activate the control component.

Description

201015538 IX. Description of the invention: [Technical field to which the invention pertains] A speech recognition control device, in particular, an intelligent speech recognition control device using a host and having a scalability. [Prior Art] The controller is used to control the startup interface of devices such as home appliances, machines, toys, etc., which is convenient for the user to use, and can be formed in various interface forms.

Come like buttons, remote control, joystick, etc. The user can operate the device such as a home appliance, a machine, a toy, etc. by pressing or pushing a button, a remote control n, a joystick, and the like. The control of money touch (4) relies on physical contact, so that the controller is controlled to control each person to have direct contact. The operation is used, and the use is already used. Therefore, it is inconvenient in use. A non-contact controller was developed. #Contact type controller currently has sound _ separately by sound, _ 4 voice control 7^, magnetron, etc., its ❺ first line force field to trigger 'according to the difference of principle, track and field controller feels 磬咅, stop # field When the pure magnetic field is touched, the sound, light, and magnetic technique "(4)" form a non-contact controller. The sound control circuit can be used by sound to know whether or not to activate the second ',,,, It is simply by feeling the presence or absence of the sound to determine if the ambient sound is too large, it is easy to start the control by mistake, but if the door is turned up, then the voice control is required to be known as Ray = 曰 / It causes troubles in use. And circuit .1 'is a hardware design, so its almost 5 201015538 is difficult to meet the use of space without expansion, so its application is quite limited. [Summary] In the absence of the present invention, the main purpose of the present invention is to provide a smart voice recognition control device, which can be identified by a sound software stored in the host to correctly activate the controller. Providing a type of intelligent voice _ (four) device, which includes a 70-piece sound, 70 (m) transmission and control components, a main: a ligature component and connects the control to the cow through the signal transmission component, the host includes - an arithmetic unit and a digital information storage unit, and the digital information storage unit prestores at least - a voice command twisted: the processing software 'where the sound pickup element is used to dig a voice message: by the operator of the second host The speech recognition processing software normalizes the speech message to the lion logic Kalman filter processing, the track measurement and the frame processing, the (4) spectrum analysis, the Meiguan degree processing, the filter period group processing, and the cepstrum parameter processing. And the dynamic time calibration & the voice message is processed by the phonetic recognition processing software to compare whether the voice message conforms to the voice command, and when the voice message conforms to the voice command, a motion signal is generated to The signal transmission component is transmitted to the control component to activate the control component. Accordingly, the present invention can process the signal after capturing the voice message and compare it to whether The pre-stored voice command is the same, and after the comparison is the same, the start signal is generated to activate the control component, so that the false threshold caused by the ambient noise can be solved without increasing the intensity threshold of the sound activation. The invention runs in software, so it has a large expansion space, which can meet the needs of the user. [Embodiment] The details and technical description of the present invention are as follows: Please refer to "Figure 2" and "Figure 3 shows that the present invention comprises a sound pickup 10 component 1 , a host 20 , a signal transmission component 30 and a control component 4 , wherein the sound pickup component 1 is used for capturing a voice message and the host 2, the host 20 includes an operator 21 and a digital information storage unit 22, and the digital information storage unit 22 prestores at least one voice command, and the operator 21 of the host 20 executes the voice recognition processing software pair. The voice message is subjected to normalized fuzzy logic Kalman filter processing, endpoint detection and sound box processing, cepstrum analysis, Meyer scale processing, and filter At φ processing, Mel Cepstrum parameters and the dynamic time alignment process method. The linear prediction coefficient algorithm is used in the Kalman filter. The formula for linear prediction is: y[n] = ^aky[nk] + e[n] The speech signal is used for noise elimination in the Kalman/wave filter. The voice enhancement process is performed to obtain the first filtered output signal. 7 201015538 , the Kalman filter output signal is obtained by the normalized least mean square algorithm to obtain the adaptive constant σ ", the formula is as follows: " 崎 (8) f at the same time with the card ® MANN filter wheel signal and the original voice signal minus The mean square processing of the error is the feedback error, and the error variation is the input of the fuzzy logic operation. The parameters of the Kalman filter are demodulated via the blur to obtain a better output speech signal. The output speech signal of the normalized fuzzy logic Kalman filter can be used as a linear prediction coefficient algorithm to estimate the spectrum model, and then the aforementioned linear prediction coefficient algorithm is repeated, in order to calculate the sample value in a sound frame. To find a set of linear prediction coefficients, so that the energy sum of the errors is minimized. 5=Σ 幻], **=1 y and endpoint detection and sound box processing, the steps include step 1: defining the voice length, step 2. cutting the sound box and step three: finding the voice with zero rate detection End point and end point. Generally, the discrete speech signals paid by sampling, the spectrum of the average person's speech is concentrated below 4 kHz, so according to the sampling principle, in order to avoid distortion, the sampling frequency should be set at more than twice the signal bandwidth. In order to reduce the impact of the amplitude of the voice signal on the system, the voice signal should be normalized, as follows. S眶=max〇S(«)), w = l,2,3,...,#

Where k is the maximum amplitude, and after each sample point is divided by ^^, the amplitude of the entire speech signal is set between -1 and 1. In order to reduce the unnecessary information, and to improve the reaction time of the endpoint detection, the obtained voice signal is determined as the beginning and end of the voice. Because the voice is a time-varying signal, it can be observed that the voice signal changes slowly in a short time. Therefore, after obtaining the voice endpoint information, the system forms a frame with fixed sampling points for subsequent processing and analysis. The figure is shown in Figure 4-2. Regardless of whether it is recognized as a continuous sentence or a single word, endpoint detection relies on short time energy and ZER〇cr〇ssing rate as detection criteria. The description is as follows: (1) Short Time-distance energy: It is a small sound segment that cuts out a short period of time. The short-term energy of each frame will be defined as the time of the # 讯 ^ 二 能量 能量 能量 取 取 取 取 取 取Five (illusion, as shown in the following formula; if Qiu exceeds the preset energy threshold, it will be considered that the sound box contains voice information, and the real speaking paragraph will be higher than other silent parts, ° 9 201015538 Nl E(k) = Y^(n + m)\ OT=0 where #: indicates a length of the frame moxibustion: is the number of each frame «: sampling point time: sampling point in the sound box · · Μ Μ The short-time energy of the frame (2) is zero: the number of times the signal passes through the origin, that is, the amplitude of the adjacent signal sample. Under a positive-negative change, the cumulative value of the zero-crossing rate is increased by one, as shown in the following equation. The zero-rate detection is used to assist the short-term energy detection in the judgment of the deficiency, such as the friction sound when talking, nose Sounds, consonants, etc., are not enough to exceed the critical value of short-term energy detection in terms of energy performance, so there will be false positives.

1 Ν-Χ ^ m=0 1))1 sgn[5'(n)]= 1 "(8) 20 -1 if S{n) < 0 where #: indicates a length of the sound box 201015538 λ: is each Frame number w: sampling point time: sampling point zw in the sound box: the zero-zero rate value of the a-th sound box. In principle, the zero-zero rate measurement can distinguish between voiced and unvoiced voices, such as silent voices. Most of the energy is concentrated on more than 3 papers, so the higher the zero rate value will be higher; on the contrary, the higher the zero value of the voiced speech will be lower. Therefore, the endpoint detection of the voice signal is to use short-term energy detection to determine the beginning and end of the voiced speech; then, at the zero-rate detection, find the beginning and end of the voice. The search rules are as follows, if five (the magic is less than the low threshold of short-term energy, it is considered to be a non-speech part. If £ (the magic is greater than the low threshold of short-time energy, it is also higher than the high threshold of short-time energy, It is determined to be the starting point of the speech segment. If the fifth (the magic is less than the high critical value of the short-time energy, the zero-rate threshold should be added as an aid, and the threshold of the zero-crossing rate can be used to determine the starting point of the speech. To search, the fifth point is the end of the speech segment. Point. The search for the speech endpoint is 'from the end of the signal to the lower threshold than the short-term energy, then the spectrum analysis is because the speech recognition needs an effective The coefficients are compared 'so that in the speech signal analysis, the spectrum in the frequency domain will be observed.

Set the cepstral coefficient, and then reverse it, and return to the time domain to get a new set of parameters, ie, the equation is expressed as follows. 201015538 j^kn 0<,η<Ν-^ where 'is a sequence of speech signals without (8), „ = 〇1,2,,#-l, obtained by discrete Fourier transform. The q(8) obtained by the above equation is The eigenvector of a set of speech frames. Mel is the unit of measure of pitch frequency. According to the experiment, at 1 kHz, people's perception of sound will be linear; but if it exceeds _ 1 kHz, logarithm will be presented. Relationship, the relationship between the two, ® is expressed as 2595 *1ο8 (1 + ^ζ/700) where ' is the value of the Meyer scale frequency, which is the scale used for the pitch measurement. & is the actual scale frequency value, is listening The relationship between the perceived pitch and the frequency. To represent the voice signal feature, the input voice signal can be processed by the filter bank, and the output will represent a set of parameters of the voice feature, and the processing mode will transmit the previous voice signal. Sending people - group of 24 bandpass filters, analog I signal.: ^ Filter, wave group is non-uniformly spaced in the frequency = low frequency signal part of the required filter group and the bandwidth will be more - The average person has a narrow and wide sense of pitch perception. Based on the real group - the same sample _ in the Chi _, the material record ^ under. 12 201015538 • (a) the frequency range below 1kHz: for the equal space, linear, every 100Hz set to a triangle filter center frequency. ' ( b) The frequency range is below 1 kHz: for the logarithmic grid, increase the center frequency of the triangular filter by approximately 1.2 times • (c) The center frequency intensity value of the triangular filter is 1. (d) The maximum frequency of the center of the triangular filter is low. In the signal bandwidth, which is H_z 2 2 , the current maximum frequency of the center of the triangular filter is φ 3. 9 kHz ° (e) The center frequency of the triangular filter is the center frequency of the critical band, and the cutoff frequencies on both sides are two adjacent critical bands. The center frequency. Mathematical expression, the filter bank output is represented by a symbol, the input is wood (10), and the shell β + ΔΙ» m Yt(ni) = + k=Bm-Am where & is the center of each bandpass filter Frequency, each value taken is the end point of the previous one, expressed as

Bm = + Am-1 and ~ represents the left and right bandwidth of each bandpass filter from the center frequency, where each bandpass filter has a bandwidth of 2Δ», so each bandpass filter here The bandwidth will be equal to 1.2 times of the previous one, expressed as 13 201015538 △ „== 1.2*4 The frequency response of the single bandpass filter can be obtained.

This will mean that each filter starts from the center point of the previous filter, each bandwidth is increased by 1.2 times, and the frequency response of the bandpass filter exhibits a triangular characteristic distribution. After processing by the filter bank, # output values will be generated on the logarithmic spectrum, where # is the number of filters in the filter bank. Multiply the frequency energy, lxWl2, by (4-7), and add up the energy of the filter.

Since the values on the logarithmic intensity spectrum are real and symmetric, the Fourier cepstrum can be obtained by performing a discrete cosine transform on the logarithmic energy of all μ filter outputs during Fourier transform. Its Mel cepstrum parameter formula is as follows

14 201015538 where is the total number of bands, w ^ café of the Melhon spectrum coefficient (whip c). The frequency band ' w is a voice signal and the dynamic time calibration method is a lifting state time calibration method. The short form is a dynamic time calibration template matching method for the invention. A good nonlinear ❹ dynamic time calibration method is a similarity calculation method. i sat through feature extraction and feature job compression, and will produce t or several templates for each side. The recognition process is to calculate the similarity between the feature vector of the recognition mode to be tested and each template. one type. The method uses to set the reference template feature vector sequence j = {ai, a2,, α}, and input the speech feature vector sequence to... Therefore, the purpose of the (4)^ quasi-method is to find an optimal time normal function, so that the time axis of the speech template to be tested: • non-linear mapping to the time axis of the reference template; Use the equation to denote ❹ C = {C(l), C(2), .."C(i\〇} where Y is the path length and C(8)=((4),;(Λ)) represents the first "match point, This is made up of the ()th feature vector of the reference template and the (8)th feature vector of the template to be tested. The distance (or distortion value) between the two is (10), which is the local matching distance. * The algorithm of the dynamic time calibration method is to achieve the minimum of the weighted distance sum by the local optimization algorithm, expressed as 15 201015538 ' [[啦(8)七(8)),] ^min =mjn^-^- £w« _ n= l 'where w" is a weighting function, and the following two factors are selected: (1) According to the section "Selecting the local path of the previous step of the matching point, and limiting the local path in the 45° direction to adapt to the case of 〃·/. (2) Different parts of the speech will be considered to give different weights to enhance certain distinguishing features. After the voice message is processed by the voice recognition processing software, it can be compared with whether the voice message conforms to the voice command, and the host 20 can be coupled to a sound output component 50, which is captured in the sound pickup component 10 After the voice message, the voice message can be re-exported for the user to confirm, and the voice recognition processing software can generate a motion signal when the voice message conforms to the voice command. The signal transmission component 30 is a signal transmission line 'and the signal transmission@ component 30 is coupled to the host 20 to receive the activation signal from the host 20, and the control component 40 is coupled to the signal transmission component 30 and receives the signal. The start signal is activated to activate the control element 40. The signal transmission component 31 may also include a wireless transmitter 311 and a wireless receiver 312. The wireless transmitter 311 is coupled to the host 20, and the wireless receiver 312 is shown in FIG. It is coupled to the control element 40. Referring to FIG. 5 and FIG. 6 , the wireless transmitter 311 and the wireless receiver 312 can be an RF transmitter 16 201015538 'and an RF receiver. Accordingly, it forms wireless sound control, which increases the convenience of use. As described above, the present invention performs the voice recognition processing software by the arithmetic unit 21 of the host 20 to perform signal processing after the borrowing of the audio component 1〇_takes the voice message, and is compared with whether it is the same as the pre-stored voice command. The start signal is generated after the same to start the control element 4, so there is no need to raise the threshold of the voice-activated start, so under the high ambient noise, the false start can be prevented, and the interference of the environmental noise can be avoided, and the correct operation can still be performed. The voice control starts the controller, and it runs in software mode, with future expansion space to meet the needs of users. The above are only the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. That is, the equivalent changes and modifications made by the scope of the patent application of the present invention are covered by the scope of the invention.

17 201015538 * [Simple diagram of the diagram] Figure 1, a conventional voice controller circuit. Figure 2 is a diagram showing the system architecture of the present invention. • Fig. 3 is a circuit diagram showing the implementation of the sound pickup element of the present invention. 4 is a system architecture diagram of another embodiment of the present invention. Figure 5 is a circuit diagram of a wireless transmitting device of the present invention. Figure 6 is a circuit diagram of a wireless receiving device of the present invention. [Main component symbol description] ❹ 知知1: Voice control circuit 10: Sound pickup component 20: Host 21: Operator 22: Digital information storage unit ^ 30, 31: Signal transmission component 311: Wireless transmitter 312: Wireless Receiver 40: Control Element 50: Sound Output Element 18

Claims (1)

201015538 X. Patent application scope: A kind of intelligent voice recognition control device, which comprises: a sound material, a Guanyin component is used for recording a voice message; a machine includes an operator and a digital information storage unit, and the The touch-storage unit pre-county at least - the voice command and the one-to-one ^ = body and the host's operator performs the voice recognition process. = body mountain to this; H line normalization _ material Karl (four) wave device, knowing point system and sound box processing, _ 镨 analysis, plum _ degree processing, wave wave group New Zealand, Mei _ county parameter New and dynamic _ The calibration method, the speech tfi is processed by the speech recognition processing software=the information is (4) the Wei tone command, and the listening sound_new soft message matches the voice command, and generates a motion signal; the signal transmission component, the The transmission transmission component is coupled to the charmer and receives the activation signal from the host; and the control component 'controls the connection with the signal transmission component and receives the activation signal to start. The intelligent voice recognition control device of the first item of the m patent scope, wherein the signal transmission component is a signal transmission line. 3. The intelligent voice recognition control device as claimed in claim i, wherein the + signal transmission component comprises a wireless transmitter (four) and a flawless receiver, the wireless transmission n and the voice processing software. Connected, the wireless receiver is coupled to the control element. 4. The intelligent voice recognition control device of claim 3, wherein the wireless transmitter and the wireless receiver are a W transmitter 19 201015538 'and an RF receiver. 5. The intelligent speech recognition 'control device according to claim 1, wherein the host is coupled to a sound output component to allow the sound output component to be recovered after the sound pickup component captures the voice message颂 Output the voice message. ❿ 20