CN103065629A - Speech recognition system of humanoid robot - Google Patents

Abstract

The invention discloses a speech recognition system which comprises a speech input module, a preprocessing module, a feature extraction module, a training module, a recognition module, a recognition decision module and a threshold comparison module. The output end of the speech input module is connected with the input end of the preprocessing module, the output end of the preprocessing module is connected with the input end of the feather extraction module, the output end of the feather extraction module is connected with the input ends of the training module and the recognition module, and the training module is connected with the recognition module. The output end of the recognition module is connected with the input end of the recognition decision module, and the output end of the recognition decision module is connected with the input end of the threshold comparison module. The speech recognition system adopts a hidden Markov model (HMM) and a wavelet transform and a neural network technique, makes a decision by adopting threshold comparison and improves a recognition rate.

Description

A Speech Recognition System for Humanoid Robot

technical field

The present invention is a voice recognition system based on a humanoid robot, which is used for intelligent robots, and can also be used for intelligent systems or intelligent equipment, human-computer interaction devices and the like.

Background technique

Being able to use voice to communicate with machines so that machines can understand what you say is something that people have long dreamed of. Command to act, so as to realize the language communication between man and machine. With the continuous development of science and technology, speech recognition technology has emerged, which gradually realizes this ideal of human beings. But to fully realize this ideal requires the unremitting efforts of human beings. Speech recognition technology is a technology that allows machines to recognize and understand voice signals and convert them into corresponding text or commands.

Speech recognition is a very active research field in recent years. Its application fields are very wide, and the common ones are: voice input system, voice control system, voice dialing system, smart home appliances and so on. In the near future, speech recognition technology may be used as an important means of human-computer interaction, assisting or even replacing traditional keyboards, mice and other input devices for text entry and operation control on personal computers. In applications such as handheld PDAs, smart home appliances, and industrial field control, speech recognition technology has a broader development prospect. Especially in handheld embedded systems including PDA, mobile phone, etc., the existence of the keyboard has greatly hindered the miniaturization of the system. , It is also necessary to provide convenient text input capabilities. The traditional keyboard input method is no longer competent, and speech recognition technology is a very promising alternative. Moreover, the application of voice technology has become a competitive emerging technology industry. Therefore, the study of speech recognition technology has a wide range of application value and development prospects.

Speech recognition technology mainly includes three aspects: feature extraction technology, pattern matching criteria and model training technology. Voice recognition technology has also been fully used in the Internet of Vehicles. For example, in the Internet of Wing trucks, you only need to press the key to talk to the customer service personnel to dictate the destination and directly navigate, which is safe and convenient. However, speech recognition still faces the following five problems:

(1) Recognition and understanding of natural language. First, continuous speech must be decomposed into units such as words and phonemes, and secondly, a rule for understanding semantics must be established;

(2), the amount of voice information is large. Speech patterns are not only different for different speakers, but also for the same speaker. For example, the voice information of a speaker is different when he speaks casually and when he speaks seriously. the way a person speaks changes over time;

(3), the ambiguity of speech. Different words may sound similar when the speaker is speaking. This is common in English and Chinese;

(4) The phonetic characteristics of a single letter or word or character are affected by the context, resulting in changes in accent, pitch, volume and pronunciation speed, etc.;

(5) Environmental noise and interference have a serious impact on speech recognition, resulting in a low recognition rate.

In recent decades, many experts and scholars have continued to research and explore these problems, which has led to the development of speech recognition technology. And construct various speech recognition systems based on speech recognition technology. At present, the application fields of speech recognition system include: voice dialing of telephone communication, voice control of automobiles, industrial control and medical field, personal digital assistant (Personal Digital Assistant, PDA), smart toys, remote control of home appliances, etc. People continue to study speech recognition technology in the hope that one day it will be possible to achieve a free dialogue between humans and machines, just like the communication between humans, so as to realize the automation and intelligence of industrial production. With the development of science and technology and people's gradually in-depth research on speech recognition theory, the theoretical system is becoming more and more mature. With the development of digital signal processing technology, in the next 20 years, speech recognition technology will gradually enter the industry, home appliances, communications , automotive electronics, medical and various electronic equipment. It can be said with certainty that speech recognition technology will become a key technology in the future information industry. But it is also undeniable that it still has a long way to go. To be truly commercialized, it needs to make breakthroughs in many aspects and rely on the development of other related disciplines.

Contents of the invention

The present invention is a speech recognition system, the main purpose of which is to provide a speech recognition system with high efficiency, stability, strong practicability and high recognition rate.

In order to achieve the above object, the present invention uses MATLAB as a realization tool, combined with a welcoming humanoid robot platform. After building a complete voice recognition system, the user uses the platform to command voice through the microphone, input the voice signal, process and recognize it, and get the result to act on the action of the welcome robot. Evaluate whether the system can achieve the expected indicators, strong recognition ability, high accuracy rate, and robust speech recognition system.

The present invention is achieved through the following technical solutions, a speech recognition system of a humanoid robot, comprising a speech input module, a preprocessing module, a feature extraction module, a training module, a recognition module, a recognition decision module, a threshold comparison module, and a speech input module The output end of the preprocessing module is connected with the input end of the preprocessing module, the output end of the preprocessing module is connected with the input end of the feature extraction module, the output end of the feature extraction module is respectively connected with the input end of the training module and the recognition module, and the training module and the recognition module connection; the output end of the identification module is connected with the input end of the identification decision module, and the output end of the identification decision module is connected with the input end of the threshold value comparison module.

The voice input module is used for inputting original voice signals.

The preprocessing module includes a pre-filtering unit, a sampling and quantization unit, a pre-emphasis unit, a windowing unit, and an endpoint detection unit connected in sequence;

The pre-filter unit is used to remove high-frequency noise of the original speech signal;

The sampling and quantization unit samples the Nyquist sampling theorem to sample and quantize the denoised analog signal to obtain a digital signal;

The pre-emphasis unit is used to enhance the high-frequency part, so that the frequency spectrum of the signal becomes flat for parameter analysis;

The windowing unit is used to limit the signal;

The endpoint detection unit is used to detect the start point and end point of the speech segment, remove unnecessary silent segments, and extract the speech signal segment of the reagent.

The endpoint detection unit adopts the method of combining double-threshold energy method and artificial neural network.

The feature extraction module adopts a wavelet transform-based hybrid feature parameter extraction algorithm, and the extraction algorithm is wavelet transform-based linear prediction cepstral parameters and wavelet transform-based Mel frequency cepstral coefficients.

The training module uses the Baum-Welch (expectation correction) algorithm as a training and learning method for the hidden Markov model.

The identification and decision-making module obtains the output probability through a Viterbi (Viterbi) algorithm.

The threshold comparison module is used to compare the obtained output probability value with the set threshold, if it is higher than the threshold, then output the recognition result, otherwise discard the recognition result.

The working process of the present invention: the voice signal is input from the microphone, that is, the voice input module, and is preprocessed by the preprocessing module. The preprocessing includes prefiltering, sampling and quantization, pre-emphasis, windowing and endpoint detection; after preprocessing, the signal is characterized Parameter extraction, the extracted parameter sequence is established and saved as a speech parameter template library, that is, the training template module; the speech recognition process is that the speech is input from the microphone, and after preprocessing and feature parameter extraction, the extracted feature parameters are combined with the established speech parameters. The template library performs probability calculation and matching, and the matching result is compared with the threshold through the threshold comparison module, and finally the recognition result is obtained.

In the present invention, a threshold value comparison is performed after calculating the probability. If it is higher than the threshold value, it is considered as the correct recognition result, otherwise, the recognition result is discarded, and the voice command of "please say it again" is prompted to re-input the voice command. The threshold is an empirical value, which is obtained after many experiments in a specific laboratory environment.

Description of drawings

Fig. 1 speech recognition system block diagram;

Fig. 2 voice signal preprocessing block diagram;

Figure 3. Block diagram of DWTM calculation process.

Detailed ways

In order to better understand the present invention, the embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings: this embodiment is implemented on the premise of the technical solution of the present invention, and detailed implementation methods and specific operation processes are provided. The scope of protection of the invention is not limited to the following examples.

As shown in Figure 1, it is a system block diagram of the present invention, a kind of speech recognition system, comprises speech input module, preprocessing module, feature extraction module, training module, identification module, recognition decision-making module, threshold comparison module, the speech input module The output end is connected to the input end of the preprocessing module, the output end of the preprocessing module is connected to the input end of the feature extraction module, the output end of the feature extraction module is respectively connected to the input end of the training module and the identification module, and the training module is connected to the identification module ; The output end of the identification module is connected to the input end of the identification decision module, and the output end of the identification decision module is connected to the input end of the threshold comparison module.

The voice input module is used for inputting original voice signals.

As shown in FIG. 2 , the preprocessing module includes a pre-filtering unit, a sampling and quantization unit, a pre-emphasis unit, a windowing unit, and an endpoint detection unit connected in sequence.

In the preprocessing, the voice signal is pre-filtered first, and its purpose is to prevent aliasing interference. The pre-filter is actually a band-pass filter, and its upper and lower cut-off frequencies are f _H and f _L respectively; then A/D conversion is performed on the signal , the analog voice signal is a continuous signal and cannot be processed by a computer, so the first step in voice signal processing is to convert the analog signal into a digital signal. Therefore, it must go through two steps of sampling and quantization, so that after the voice signal is recorded from the microphone, the analog signal is converted into a digital signal by A/D conversion, and then it is sampled and quantized. The sampling and quantization are performed by a computer or other digital recording equipment. The collected voice signals have already been digitized, and generally do not need to be digitized by the user. According to the Nyquist sampling theorem: f _s > 2*f _max , sampling at a frequency of 8000Hz, divided into 200 sampling frames, and adjacent frames overlap by 50%; then pre-emphasize the quantized signal, due to the average of the speech signal The power spectrum is affected by glottal excitation and mouth and nose radiation, and the high-frequency end drops by 6dB/octave above 800 Hz. For this reason, the voice signal must be pre-emphasized. The purpose of pre-emphasis is to enhance the high-frequency part and make the spectrum of the signal flat, so as to facilitate spectrum analysis or channel parameter analysis. The pre-emphasis digital filter H(z)=1-uz ^-1 , u is 0.97; finally Windowing is performed on the signal. Due to the movement of the human's own vocal organs, the speech signal is a typical non-stationary signal, and its characteristics change with time. However, this physical movement is much slower than the speed of sound wave vibration. Therefore, the speech signal can often be assumed to be a stationary signal within a time period of 10-20 ms, and its spectral characteristics and some physical characteristic parameters can be approximated. regarded as invariant; what is adopted in the present invention is the Hamming window.

The pre-filter unit is used to remove high-frequency noise of the original speech signal; remove unnecessary components, prepare for subsequent signal processing, and ensure signal quality and speed.

The sampling and quantization unit samples the Nyquist sampling theorem to sample and quantize the denoised analog signal to obtain a digital signal; since the original voice signal is an analog signal mixed with high-frequency noise, the original voice signal is de-high-frequency noise And digital processing, according to the Nyquist sampling theorem, use 8000Hz frequency sampling and quantization to obtain digital voice signals.

The pre-emphasis unit is used to enhance the high-frequency part to flatten the spectrum of the digital signal for parameter analysis; the average power of the voice signal is affected by glottal excitation and mouth and nose radiation, and the high-frequency end will drop. Perform pre-emphasis processing for the speech signal, enhance the high-frequency part, and flatten the spectrum of the signal for parameter analysis. The pre-emphasis digital filter is:

H(z)＝1-μz ^-1

Among them, the value of u is close to 1, and 0.97 is taken here.

The windowing unit is used to limit the digital signal; since the voice signal is an unbalanced signal, its characteristics change with time. However, the voice signal can often be assumed to be a stationary signal within the time period of 10 ms to 20 ms, and its spectral characteristics are also approximately unchanged. Therefore, the speech signal is windowed and divided into short segments, each of which is called an analysis frame. The digitized speech signal is divided into 200-sample frames, and adjacent frames have 50% overlap. Use the Hamming window to add a window to the speech signal, and the Hamming window function is:

Among them, N is the number of sampling points per frame, and N=200 in this system.

Speech endpoint detection, the purpose is to determine the start and end of the speech from a signal containing speech, which is an extremely critical step in the speech recognition system. Only by accurately determining the endpoint of the speech signal can the correct speech processing be performed. Effective endpoint detection can not only minimize the processing time, but also eliminate the noise interference of the silent segment, so that the processing quality is guaranteed. A good endpoint detection method can change the problems of speech recognition software such as unsatisfactory detection effect and low recognition rate, and can provide a reliable basis for speech recognition. It should have good robustness and can well distinguish background noise, Non-speech sounds and the voices of non-interlocutors and normal conversational speech, reducing endpoint errors and resulting false interruptions caused by these sounds. The high precision of endpoint detection can ensure that the signal input to the recognizer is an effective and complete speech signal, making the recognition effect more accurate and faster. The commonly used method is the energy method, but in practical applications, it cannot achieve a high signal-to-noise ratio, so the endpoint detection is not accurate, and the detected speech will be incomplete, affecting the recognition effect. In the present invention, the windowed signal first calculates its short-term energy and short-term average zero-crossing rate, and initially detects its unvoiced sound, voiced sound and each speech segment; then through a multi-layer perceptron neural network, further detection, and then Smoothing is performed using a median filter, smoothing each spectrum. Experiments have proved that this hybrid method achieves very good endpoint detection results.

The endpoint detection unit is used to detect the starting point and end point of the speech segment, remove unnecessary silent segments, and extract the speech signal segment of the reagent; it is an extremely critical step in the speech recognition system. Only by accurately determining the endpoint of the speech signal can it be correct. for voice processing.

The endpoint detection unit adopts the method of combining double-threshold energy method and artificial neural network. The energy method, that is, the combination of the short-term energy method and the short-term average zero-crossing rate, can only achieve high accuracy in the case of low signal-to-noise ratio (SNR).

The short-term energy detection algorithm is suitable for detecting voiced sounds, and its calculation formula is:

${\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; \&infin;</span>}{\overset{<span\; class="notranslate">\; \&infin;</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

The short-term average zero-crossing rate is suitable for detecting unvoiced sounds, and its calculation formula is:

${\mathrm{<span\; class="notranslate">\; Z</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; \&infin;</span>}{\overset{<span\; class="notranslate">\; \&infin;</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; sgn</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; sgn</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span>$

$<span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; sgn</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; sgn</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span>$

in, $\mathrm{sgn}\left[x\left(\mathrm{no}\right)\right]=\left\{\begin{array}{c}1,x\left(\mathrm{no}\right)\&Greater\; Equal;0\\ -1,x\left(\mathrm{no}\right)<0\end{array}\right..$

The double-threshold energy method is based on the energy method to set two thresholds, one is relatively small, used to detect the boundary between the silent segment and the speech signal segment, and the other is relatively large, used to detect the strength of the speech signal. Here, the two thresholds are set looser to ensure the integrity of the voice signal. Here, the signal is then input to the artificial neural network detector.

The artificial neural network detector uses a multi-layer perceptual artificial neural network. The advantage of using it for endpoint detection is that the correlation between speech frames and the probability of error are minimized. But the main disadvantage is that it is difficult to find speech features that can significantly distinguish voiced and unvoiced sounds by using this algorithm, which just complements the advantages of the dual-threshold energy method. The artificial neural network is composed of many processing units (neurons), and the interconnection pattern reflects the structure of the neural network, which determines the ability of the network. Each neuron (such as neuron j) accepts information transmission from other neurons (such as neuron i), and the relationship between the total input is:

${\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}$

In the above formula, w _ij is the connection weight from neuron i to neuron j; x _i is the output of neuron i; θ _j is the threshold of neuron j. The output relationship of neuron j is: o _j = f(I _j ), where the function, f( ) is called the activation function, and the S-type function is selected as:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; x</span>}}}$

Multilayer perception networks (MLPN) neural network, which is a forward neural network, is a neural network with a multi-layer neuron, feedforward, and error backpropagation structure. It usually consists of three parts: a set of perceptual units form the input layer; one or more hidden layers of computing nodes; and an output layer of one layer of computing nodes. The input layer is used to perceive external information, the output layer gives classification results, and the hidden layer processes the input information. Each neuron in the hidden layer or output layer is used to perform two calculations: calculate the function signal and gradient vector estimation calculation that appear at the output of a neuron, it needs to pass through the network in reverse, so the backpropagation method ( BP algorithm) as the standard training algorithm of MLPN network, the BP algorithm learns in a way with a teacher teaching, and the learning process consists of a forward propagation process and a back propagation process. First, the teacher sets an expected output value for each input mode, and then inputs the actual learning and memory mode (that is, training samples) to the net sheet, and propagates from the input layer to the output layer through the intermediate layer. The difference between the actual output and the expected value is It is the error, and then the connection weight is corrected layer by layer from the output layer to the middle layer according to the error. Through such a process, the actual output gradually approaches the corresponding expected value output, and finally the connection weight ω _ij between each layer is determined.

Compared with the background noise, the amplitude of the speech signal has a large dynamic range. It can be considered that there are many random events in the speech segment, and the average amount of information is large, that is, the entropy value is large; moreover, the energy distribution of the background noise segment is relatively stable at each frequency. Reflected on the amount of information, it is considered that the average amount of information contained in it, that is, the spectral entropy, is relatively large. Therefore, the speech signal processed by the double-threshold energy method is used to calculate its amplitude entropy and spectral entropy and take it as the input of the MLPN neural network, propagate through each layer of the neural network, and finally obtain the detection result.

The feature extraction module adopts a wavelet transform-based hybrid feature parameter extraction algorithm, and the extraction algorithm is wavelet transform-based linear prediction cepstral parameters and wavelet transform-based Mel frequency cepstral coefficients.

Speech recognition is a matching process. First, a model is established according to the characteristics of the speech, the input speech signal is analyzed, the required features are extracted, and the template required for speech recognition is established on this basis. In the recognition process, according to the overall model of speech recognition, the speech template stored in the computer is compared with the characteristics of the input speech signal, and a series of optimal matching with the input speech is found according to a certain search and matching strategy. Template to obtain the recognition result, which is the principle of speech recognition. Commonly used methods for extracting feature parameters of speech signals include: LPCC (Linear Prediction Cepstral Coefficient) parameters, MFCC (Mel Frequency Cepstral Coefficient) parameters, parameters based on LPCC or MFCC first-order difference, and so on. The disadvantage of the LPCC parameter is that it does not use the hearing of the ear. Although the MFCC parameter uses the auditory characteristics of the ear, these two parameters only reflect the static characteristics of the speech and cannot reflect its dynamic characteristics. Therefore, a first-order difference parameter is proposed. Describe the dynamics of speech. Speech signal is non-stationary signal, and Fourier transform is a kind of analysis method to the globality of signal, can't reflect the locality of speech signal, the present invention has introduced discrete wavelet transform, when extracting LPCC parameter and MFCC parameter, replace Fourier transform with discrete wavelet transform Transform, taking advantage of wavelet transform, can improve the recognition performance of the speech recognition system. In order to improve the recognition rate of the system, calculate its first-order difference parameters, combine the two, and use ΔDWTL+DWTL, ΔDWTM+DWTM as the characteristic parameters of the speech signal. Among them, DWTL and DWTM are LPCC and MFCC parameters using wavelet transform respectively, and ΔDWTL and ΔDWTM are difference parameters of LPCC and MFCC parameters respectively.

Feature parameter extraction refers to the process of obtaining a set of characteristic parameters that can describe the speech signal from the speech signal. The proposed algorithm is based on linear predictive cepstral parameters (LPCC) and Mel-frequency cepstral coefficients (MFCC), replacing discrete Fourier transform with discrete wavelet transform. The window can be adaptively adjusted according to different frequencies, so that it can accurately reflect the instantaneous changes of non-stationary signals; wavelet analysis has different resolutions at different positions on the time-frequency plane, and is a multi-resolution analysis method. That is, it has higher frequency resolution and lower time resolution in the low frequency part, and has higher time resolution and lower frequency resolution in the high frequency part. characteristic capabilities.

The linear predictive cepstral parameter (LPCC) is a characteristic parameter that characterizes the personality of the speaker, and its model system function:

$\mathrm{<span\; class="notranslate">\; lm\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{<span\; class="notranslate">\; \&infin;</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; z</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; no</span>}}$

Combining the above two formulas, and then deriving both sides with respect to z, we have

$<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; p</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; z</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{<span\; class="notranslate">\; \&infin;</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; no</span>}\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; z</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; p</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; k</span>}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; z</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}$

也即得出LPCC参数。Let the constant terms around the above formula and the coefficients of each power of z ^-1 be equal to obtain the recursive relationship between h(n) and a _k , and then get the cepstrum from the prediction coefficient a _k

That is to say, the LPCC parameters are obtained.

Calculation steps of linear predictive cepstrum parameters (referred to as DWTL parameters) based on wavelet transform:

(1), the original voice signal must first be preprocessed;

(2), carry out wavelet packet decomposition to each frame signal, calculate wavelet packet coefficient;

(3), obtain the cepstrum D _n to the wavelet packet coefficient obtained in the previous step;

(4) Merge D _n into a new vector [D ₁ D _{2 .} . . D _n ] as a DWTL parameter.

The most important feature of the Mel frequency cepstrum coefficient (MFCC) is the use of the auditory principle of the human ear and the decorrelation characteristics of the cepstrum. The Mel frequency and the Hz frequency have a nonlinear correspondence, which is based on the relationship between them. Transform the frequency spectrum of a speech signal into the perceptual frequency domain. Its relationship is:

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2595</span>}\mathrm{<span\; class="notranslate">\; lg</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>\frac{{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; Hz</span>}}}{\mathrm{<span\; class="notranslate">\; 700</span>}}<span\; class="notranslate">\; )</span>$

Among them, f _m is the perceived frequency domain in Mel, and f _Hz is the actual frequency domain in Hz.

The calculation process of Mel frequency cepstrum parameters (abbreviated as DWTM parameters) based on wavelet transform is shown in Figure 3.

DWTM parameter calculation steps:

(1), the original voice signal must first be preprocessed;

(2), each frame signal after preprocessing is all processed through discrete wavelet transform, and the signal after processing is carried out wavelet packet decomposition, calculates wavelet packet coefficient;

(3), filter the wavelet coefficients decomposed on each frequency band;

(4), obtain the energy value S(m) of the coefficient again with the obtained in the previous step;

(5) Transform the logarithmic spectrum S(m) into the cepstrum frequency domain through discrete cosine transform DCT.

The order of DWTM and DWTL parameters is 13, the order of 12 is taken after removing the DC component c(0), and the filter group is 24 groups. Take ΔDWTL+DWTL, ΔDWTM+DWTM as the characteristic parameters of the speech signal. The effective combination of dynamic features and static features can improve the recognition rate of the system. Among them, DWTL and DWTM are the LPCC and MFCC parameters using wavelet transform respectively, and ΔDWTL and ΔDWTM are the difference parameters of DWTL and DWTM parameters respectively. The mixing parameters are a 24*24 matrix. The matrix is as follows:

$\mathrm{<span\; class="notranslate">\; ML</span>}<span\; class="notranslate">\; =</span>\left[\begin{array}{cc}\mathrm{<span\; class="notranslate">\; \&Delta;DWTL</span>}& \mathrm{<span\; class="notranslate">\; \&Delta;DWTM</span>}\\ \mathrm{<span\; class="notranslate">\; DWTL</span>}& \mathrm{<span\; class="notranslate">\; DWTM</span>}\end{array}\right]$

Commonly used identification decisions include dynamic time bending algorithm, Viterbi algorithm, etc. Among the probability statistics methods, the most representative method is the HMM method, and the algorithm most closely combined with HMM is the Viterbi algorithm. The Viterbi algorithm is a dynamic programming algorithm used to find the most likely hidden state sequence (Viterbi path) generated by the observed information (Observed Event). The forward algorithm is a similar algorithm used to calculate the probability of a series of observed events. . During identification, the probability is calculated for each observed event, and the highest probability is the identification result. Sometimes, the highest probability is not necessarily the correct identification result. Therefore, in the present invention, a threshold comparison is performed after calculating the probability. If it exceeds the threshold, it is considered to be the correct recognition result; otherwise, the recognition result is discarded, and the voice command "Please say it again" is prompted and the voice command is re-entered. The threshold is an empirical value, which is obtained after many experiments in a specific laboratory environment.

The training module uses the Baum-Welch (expectation correction) algorithm as a training and learning method for the hidden Markov model.

The identification and decision-making module obtains the output probability through a Viterbi (Viterbi) algorithm.

The threshold comparison module is used to compare the obtained output probability value with a set threshold, and if it is higher than the threshold, the recognition result is output, otherwise, the recognition result is discarded.

The speech recognition system includes preprocessing, endpoint detection, feature parameter extraction, recognition decision-making and threshold comparison, etc. When a complete speech recognition system is set up, then a template library will be set up. The template library will be selected and established according to the needs of the actual application. template library. The voice commands of the template library include: direction commands, such as "left", "right", "forward", "stop" and so on; greeting commands, such as "hello", "thank you"; question and answer commands, such as "you What's your name?", "Where are you from?", "What functions do you have?", etc. Each command includes the following processes: input template speech, preprocessing, detection of speech endpoints, extraction of speech and feature parameters, recognition decision, threshold comparison.

1) Voice recording: Digitize the analog voice signal, and then sample and quantify the voice with a sampling frequency of 8000HZ;

2), pre-processing includes: pre-filtering, pre-filtering is actually a band-pass filter, its purpose is to prevent aliasing interference; pre-emphasis, the purpose of pre-emphasis is to enhance the high frequency part, so that the spectrum of the signal becomes flat , in order to carry out spectrum analysis or channel parameter analysis, the pre-emphasis digital filter H(z)=1-uz ^-1 , u is 0.97; add window, the speech signal is a non-stationary signal, and the speech signal can usually be assumed to be in 10～20ms In such a time period, the speech signal is a stationary signal, and its spectral characteristics and some physical characteristic parameters can be approximately regarded as constant;

3), endpoint detection, purpose is to determine the starting point and the end point of speech from a section signal that comprises speech, in the present embodiment, first use energy method to initially detect its endpoint, then do further detection through multi-layer perceptron neural network;

4), extract characteristic parameters, calculate DWTL parameter and DWTM parameter respectively, then calculate first-order difference parameter ΔDWTL and ΔDWTM respectively; Then they are combined into a matrix of 24*24 as speech characteristic parameter;

5) Establish a template library, and establish a parameter sequence with the extracted feature parameters.

For each voice command, go through steps 1) to 5) and repeat the cycle to finally establish a complete voice template library. After the template library is built, it will be tested, and the process is as follows:

Step 1), input test voice, input voice command to microphone, as " to the right ", through A/D, analog voice signal is converted into digital voice signal, namely digitization; Then it is sampled, quantized.

Step 2), preprocessing, the process is as above step 2).

Step 3), endpoint detection, the process is as above step 3).

Step 4), extracting characteristic parameters, the process is as above step 4).

Step 5), identify the decision, calculate the probability for each extracted parameter, and the one with the highest probability is the result to be identified.

Step 6), the maximum probability calculated in step 5) is compared with the threshold value, the maximum probability greater than or equal to the threshold value is used as the recognition result, and if it is less than the threshold value, the result to be recognized is discarded, and the user is prompted to input the voice command again.

Do the same process for other test voice commands until all test voices are tested. And discover the deficiencies of the system, then adjust the parameters, and then test. Experiments have proved that the present invention can recognize voice commands very well, for example, say "turn left" to the microphone, after processing and recognition, the robot turns left. Other commands or questions and answers can be effectively recognized.

Claims (8)

1. A voice recognition system comprises a voice input module (1), a preprocessing module (2), a feature extraction module (3), a training module (4) and a recognition module (5), wherein the output end of the voice input module (1) is connected with the input end of the preprocessing module (2), the output end of the preprocessing module (2) is connected with the input end of the feature extraction module (3), the output end of the feature extraction module (3) is respectively connected with the input ends of the training module (4) and the recognition module (5), and the training module (4) is connected with the recognition module (5); the device is characterized by further comprising an identification decision module (6) and a threshold comparison module (7), wherein the output end of the identification module (5) is connected with the input end of the identification decision module (6), and the output end of the identification decision module (6) is connected with the input end of the threshold comparison module (7).

2. Speech recognition system according to claim 1, characterized in that the speech input module (1) is adapted to input an original speech signal.

3. The speech recognition system according to claim 1, characterized in that the pre-processing module (2) comprises a pre-filtering unit, a sampling and quantization unit, a pre-emphasis unit, a windowing unit, an end-point detection unit, connected in series;

the pre-filtering unit is used for removing high-frequency noise of an original voice signal;

the sampling and quantizing unit samples Nyquist sampling theorem sampling and quantizing de-noised analog signals to obtain digital signals;

the pre-emphasis unit is used for improving a high-frequency part and flattening the frequency spectrum of the signal so as to analyze parameters;

the windowing unit is used for limiting the signal;

the end point detection unit is used for detecting the starting point and the end point of the voice section, removing the unnecessary mute section and extracting the voice signal section of the reagent.

4. The speech recognition system of claim 3 wherein the endpoint detection unit employs a combination of a dual threshold energy method and an artificial neural network.

5. The speech recognition system according to claim 1, wherein the feature extraction module (3) employs a wavelet transform-based hybrid feature parameter extraction algorithm, which is wavelet transform-based linear prediction cepstrum parameters and wavelet transform-based Mel-frequency cepstrum coefficients.

6. The speech recognition system according to claim 1, characterized in that the training module (4) is a training learning method as a hidden markov model by means of the Baum-Welch algorithm.

7. The speech recognition system of claim 1, wherein the recognition decision block (6) derives the output probability by a Viterbi algorithm.

8. The speech recognition system according to claim 1, characterized in that the threshold comparison module (7) is adapted to compare the obtained output probability value with a set threshold, and to output the recognition result if the obtained output probability value is higher than the set threshold, and to discard the recognition result if the obtained output probability value is not higher than the set threshold.

Images