AI speech recognition, reasoning, generation system and method

Through the deep learning algorithm and multi-scale feature extraction method of AI speech recognition system, the problem of inaccurate recognition of speech recognition at fast speech speed is solved, and higher recognition accuracy is achieved.

CN118447842BActive Publication Date: 2024-11-22SHENZHEN JINGTIAN NETWORK TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410638382.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-21
Publication Date
2024-11-22
Estimated Expiration
2044-05-21

AI Technical Summary

Technical Problem

Existing speech recognition technology will have inaccurate recognition problems when users speak too fast, mainly because the syllables and word boundaries in the speech signal become blurred.

Method used

Using AI speech recognition, inference and generation systems, deep learning algorithms and multi-scale feature extraction methods are used through speech input modules, adaptive noise suppression modules, speech recognition modules and generation modules, including short-term, medium-term and long-term feature extraction , feature fusion and model training to optimize the recognition accuracy of fast speech speed.

Benefits of technology

Through deep learning algorithms and multi-scale feature extraction methods, the system can more accurately identify words and syllables at fast speech speeds, comprehensively capture local and global information of speech signals, and improve the accuracy of recognition.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118447842B_ABST
    Figure CN118447842B_ABST
Patent Text Reader

Abstract

This application provides an AI speech recognition, reasoning, generation system and method, a speech input module for receiving user speech input, an adaptive noise suppression module for filtering the background noise of the current environment through a noise suppression algorithm and a noise filtering model, and a speech recognition module for recognizing speech through a deep learning algorithm to optimize the recognition accuracy of fast speech speed. Through deep learning algorithms and multi-scale feature extraction methods, this application enables the system to more accurately recognize words and syllables at fast speech speeds, thereby effectively solving the problem of inaccurate recognition caused by too fast speech speeds. By extracting short-term, medium-term, and long-term features, the system can fully capture the local and global information of speech signals, providing a rich feature basis for accurate speech recognition.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of speech recognition, and in particular to AI speech recognition, reasoning, generation systems and methods. Background Art

[0002] With the development of artificial intelligence technology, speech recognition technology has been widely used in various fields, such as smart homes, car systems, smart assistants, etc. However, existing speech recognition technology often has problems with inaccurate recognition when facing users who speak too fast. This is mainly because the boundaries between syllables and words in the speech signal become blurred due to the fast speaking speed.

[0003] Therefore, we made improvements to this and proposed AI speech recognition, reasoning, and generation systems and methods. Summary of the invention

[0004] The purpose of the present invention is to address the problem that the existing speech recognition technology often has inaccurate recognition when facing the situation that the user speaks too fast.

[0005] In order to achieve the above-mentioned purpose of the invention, the present invention provides the following AI speech recognition, reasoning, generation system and method to improve the above-mentioned problems.

[0006] The specific application is as follows:

[0007] AI speech recognition, reasoning, and generation systems, including:

[0008] A voice input module, used to receive a user's voice input;

[0009] Adaptive noise suppression module, which filters the background noise of the current environment through noise suppression algorithm and noise filtering model;

[0010] The speech recognition module recognizes speech through a deep learning algorithm and optimizes the recognition accuracy of fast-speech speeds;

[0011] A reasoning module, used for performing logical reasoning based on the recognized speech content;

[0012] The generation module generates corresponding outputs based on the reasoning results;

[0013] The speech recognition module recognizes speech through a deep learning algorithm, specifically including the following steps:

[0014] S1. Data preprocessing: preprocess the input speech signal, including noise reduction, standardization and segmentation, to ensure signal quality;

[0015] S2, multi-scale feature extraction: extract features of speech and extract features of corresponding scales, including short-term features, medium-term features and long-term features, to fully capture the local and global information of speech signals;

[0016] S3, feature fusion: The multi-scale features extracted in S2 are fused using a serial fusion strategy to form a comprehensive feature vector that is rich in multi-scale speech information;

[0017] S4, model training: use the fused feature vector to train a deep learning model;

[0018] S5. Recognition output: The speech signal to be recognized is recognized through the above-mentioned trained model, and the recognition result is output.

[0019] As a preferred technical solution of the present application, in S2, the feature extraction step specifically includes:

[0020] S200, pre-treatment:

[0021] Pre-emphasis: Use a high-pass filter to improve the signal-to-noise ratio of the signal in the high-frequency part. The pre-emphasis method is a first-order high-pass filter, and its formula is y[n]=x[n]-αx[n-1], where α is the filter coefficient, which ranges from 0.9 to 1.0, y[n] is the value of the current output signal, which is a point in the time series data, n represents the current sample point, x[n] is the value of the current input signal, x[n-1] is the input signal value of the previous sample point, and n-1 is the time point before the current time point;

[0022] Framing: The speech signal is divided into short-term stable frames. The framing operation is achieved by multiplying the Hamming window with the original signal. Where w(n) is the weight value of the window function at the nth sampling point, n is the sample, and N is the window length. The Hamming window is applied to the original speech signal and multiplied by each sample of the corresponding segment of the original signal to obtain a frame of windowed signal. The sliding of the Hamming window on the signal divides the entire speech signal into multiple overlapping frames.

[0023] S201, extraction: performing FFT on each frame signal to convert the time domain signal into a frequency domain signal;

[0024] S202, dynamic feature calculation: In order to capture the dynamic information in the speech signal, the first-order and second-order differential dynamic features are calculated.

[0025] As a preferred technical solution of the present application, in S201, the specific conversion steps include:

[0026] Input data preparation: Signal samples extracted from the preprocessed and framed speech signal are input into FFT;

[0027] Apply the Cooley-Tukey algorithm to obtain the frequency domain signal: Apply the formula Convert the time domain signal x[n] into the frequency domain signal X[k], where X[k] represents the complex value at the kth frequency component, and k is the index of the frequency domain ranging from 0 to N-1, where is the summation symbol, which means that for all N

[0028] To sum, is a complex exponential function, j is an imaginary unit, and j satisfies 2 =-1, is the phase difference, which determines the rotation speed and direction of each frequency component.

[0029] As a preferred technical solution of the present application, in S202, the specific calculation steps include:

[0030] S2020, first-order difference calculation:

[0031] Set an empty first-order difference sequence Δx[n];

[0032] Calculate the first-order difference: For each sample point x[n] in the time domain signal, its first-order difference Δx[n] is defined as the difference between the current sample point x[n] and the previous sample point x[n-1], that is, Δx[n] = x[n] - x[n-1];

[0033] Forming a first-order difference sequence: Repeat the above calculation process until all sample points in the time domain signal are processed, thereby obtaining a complete first-order difference sequence Δx[n];

[0034] S2021, second-order difference calculation:

[0035] Set an empty second-order difference sequence Δ 2 x[n];

[0036] Calculate the second-order difference: Second-order difference Δ 2 x[n] is the difference of the first-order differences, calculated as:

[0037] Δ 2 x[n]=Δx[n]-Δx[n-1], which is the difference between the first-order difference Δx[n] of the current sample point and the first-order difference Δx[n-1] of the previous sample point;

[0038] Forming a second-order difference sequence: Repeat the above calculation process until all sample points in the first-order difference sequence are processed, thereby obtaining a complete second-order difference sequence Δ 2 x[n].

[0039] As a preferred technical solution of the present application, in S2, the feature extraction step further includes:

[0040] S203, extracting short-term, medium-term and long-term features:

[0041] Short-time feature extraction: Calculate the energy of each frame signal. The short-time energy is the sum of the squares of all sample values ​​in the frame. The specific formula is: Where E is the short-time energy, x[n] is the sample in the frame, n is the length of the frame, and the number of times the sign of each frame signal changes is counted, which is recorded as the short-time zero-crossing rate;

[0042] Mid-time feature extraction: Extract the amplitude and power spectrum as features from the spectrum obtained after FFT conversion, pass the spectrum obtained by FFT through the Mel filter bank, map the frequency to the Mel scale, take the logarithm of the output of each filter, and then perform discrete cosine transform to obtain MFCC;

[0043] Long-term feature extraction: The autocorrelation function (ACF) method is used to calculate the fundamental frequency of each frame. The specific formula is: Where m is the delay. Multiply the signal x[n] by its version x[nm] delayed by m units. Sum all the products from m to n-1 to get the autocorrelation function r[m]. For periodic signals, the ACF will have a peak at the delay corresponding to the signal period. The period is calculated based on the peak. The specific formula is: T is the period F s is the sampling rate, the fundamental frequency F0 is the inverse of the period, through the formula: Calculated;

[0044] Concatenate the fundamental frequencies of each frame to form a pitch contour.

[0045] As a preferred technical solution of the present application, in S3, the specific fusion steps include:

[0046] Feature concatenation: short-term features, medium-term features, long-term features, and first-order and second-order differential features are concatenated to form a long feature vector.

[0047] As a preferred technical solution of the present application, in S4, a long feature vector is used to train a deep learning model, which specifically includes the following steps:

[0048] Data preparation: Match the fused feature vectors to form a training data set;

[0049] Model selection: Choose a deep neural network;

[0050] Training process: Use the training data set to train the model and update the model parameters through the back-propagation algorithm and gradient descent optimizer.

[0051] The AI ​​speech recognition, reasoning, and generation method includes the following steps:

[0052] Step 1: Voice input and preprocessing: The user inputs voice through the voice input module, and the adaptive noise suppression module filters the background noise in real time, enhances the voice signal, and preprocesses the input voice signal, including noise reduction, standardization, and segmentation;

[0053] Step 2: Multi-scale feature extraction: extract features from the preprocessed speech signal;

[0054] Extract short-time features, including short-time energy and short-time zero-crossing rate, to capture the instantaneous characteristics of speech signals;

[0055] Extract mid-time features, including MFCC, to reflect the spectral characteristics of speech;

[0056] Extract long-term features, including fundamental frequency and pitch contour, to capture the intonation and rhythm information of speech;

[0057] Compute first-order and second-order differential dynamic features to capture the dynamic changes of speech signals over time;

[0058] Step 3: Feature fusion: The extracted short-time features, medium-time features, long-time features, and first-order and second-order differential features are fused in series to form a comprehensive feature vector that is rich in multi-scale speech information.

[0059] As the preferred technical solution of this application, the following steps are also included:

[0060] Step 4: Model training and recognition: Use the fused feature vector to train the deep neural network, update the model parameters through the back propagation algorithm and gradient descent optimizer to minimize the prediction error, input the speech signal to be recognized into the trained model for recognition, and output the recognition result.

[0061] As the preferred technical solution of this application, the following steps are also included:

[0062] Step 5: Reasoning and generation: The reasoning module performs logical reasoning based on the recognized speech content, and the generation module generates corresponding outputs based on the reasoning results.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] In the scheme of this application:

[0065] 1. Through deep learning algorithms and multi-scale feature extraction methods, the system can more accurately identify words and syllables at fast speech speeds, thereby effectively solving the problem of inaccurate recognition caused by too fast speech speeds. By extracting short-term, medium-term and long-term features, the system can fully capture the local and global information of speech signals, providing a rich feature basis for accurate speech recognition;

[0066] 2. To achieve feature fusion, the system combines features of different scales to form a comprehensive feature vector, so that the model can comprehensively consider multiple feature information when processing speech with too fast a speed, thereby improving recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 Schematic diagram of the AI ​​speech recognition, reasoning, and generation system provided for this application;

[0068] Figure 2 Schematic diagram of the AI ​​speech recognition, reasoning, and generation methods provided for this application. DETAILED DESCRIPTION

[0069] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0070] As described in the background art, existing speech recognition technologies often have the problem of inaccurate recognition when facing the situation where the user speaks too fast.

[0071] In order to solve this technical problem, the present invention provides an AI speech recognition, reasoning, generation system and method, which are applied to the field of speech recognition.

[0072] Specifically, please refer to Figure 1 , the AI ​​speech recognition, reasoning, and generation system specifically includes:

[0073] A voice input module, used to receive a user's voice input;

[0074] Adaptive noise suppression module, which filters the background noise of the current environment through noise suppression algorithm and noise filtering model;

[0075] The speech recognition module recognizes speech through a deep learning algorithm and optimizes the recognition accuracy of fast-speech speeds;

[0076] A reasoning module, used for performing logical reasoning based on the recognized speech content;

[0077] The generation module generates corresponding outputs based on the reasoning results;

[0078] The speech recognition module recognizes speech through a deep learning algorithm, specifically including the following steps:

[0079] S1. Data preprocessing: preprocessing the input speech signal, including noise reduction, standardization and segmentation, to ensure signal quality (this process is a well-known technology and will not be described in detail);

[0080] S2, multi-scale feature extraction: extract features of speech and extract features of corresponding scales, including short-term features, medium-term features and long-term features, to fully capture the local and global information of speech signals;

[0081] S3, feature fusion: The multi-scale features extracted in S2 are fused using a serial fusion strategy to form a comprehensive feature vector that is rich in multi-scale speech information;

[0082] S4, model training: use the fused feature vector to train a deep learning model;

[0083] S5. Recognition output: The speech signal to be recognized is recognized through the above-mentioned trained model, and the recognition result is output.

[0084] The AI ​​speech recognition, reasoning, and generation system provided by the present invention, through deep learning algorithms and multi-scale feature extraction methods, can more accurately recognize words and syllables at fast speaking speeds, thereby effectively solving the problem of inaccurate recognition caused by too fast speaking speeds. By extracting short-time, medium-time, and long-time features, the system can comprehensively capture the local and global information of speech signals, providing a rich feature basis for accurate speech recognition.

[0085] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings.

[0086] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features and technical solutions in the embodiments may be combined with each other.

[0087] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0088] Example 1, please refer to Figure 1 ,AI speech recognition, reasoning, and generation system,In S2, the feature extraction steps specifically include:

[0089] S200, pre-treatment:

[0090] Pre-emphasis: Use a high-pass filter to improve the signal-to-noise ratio of the signal in the high-frequency part. The pre-emphasis method is a first-order high-pass filter, and its formula is y[n] = x[n]-αx[n-1], where α is the filter coefficient, which ranges from 0.9 to 1.0. It determines the degree of influence of the previous sample point on the current output value. By adjusting the value of α, the intensity or degree of smoothing of the filter can be controlled. y[n] is the value of the current output signal, which is a point in the time series data. n represents the current sample point, x[n] is the value of the current input signal, x[n-1] is the input signal value of the previous sample point, and n-1 is the time point before the current time point. It can be used for signal smoothing, denoising, and prediction. When α is close to 1, the filter gives greater weight to past values, resulting in a smoother output; when α is close to 0, the output of the filter will be closer to the original signal;

[0091] Framing: The speech signal is divided into short stable frames (since the speech signal is non-stationary, it needs to be divided into short stable segments, i.e. frames. The commonly used frame length is 20-30ms, and the frame shift is usually 10ms). The framing operation is achieved by multiplying the Hamming window with the original signal. Among them, w(n) is the weight value of the window function at the nth sampling point, n is the sample, N is the window length (the total number of sampling points), 0.54 and 0.46 are the specific coefficients of the Hamming window, which determine the specific characteristics of the window shape, including the height of the side lobe and the width of the main lobe. These coefficients are selected to optimize the performance of the window in signal processing. This process slides over the entire signal, usually with a certain frame shift (i.e., the number of samples each time the window moves) to analyze the local characteristics of the signal. The Hamming window is applied to the original speech signal and multiplied by each sample of the corresponding segment of the original signal to obtain a frame of windowed signal. The sliding of the Hamming window on the signal (by changing the starting position of the window) divides the entire speech signal into multiple overlapping frames. After the framing operation, we obtain a series of short-time windowed signal frames, which can be regarded as approximately stable signal segments, so as to facilitate subsequent spectrum analysis, feature extraction and other processing;

[0092] S201, extraction: performing FFT (Fast Fourier Transform) on each frame signal to convert the time domain signal into a frequency domain signal;

[0093] S202, dynamic feature calculation: In order to capture the dynamic information in the speech signal, the first-order and second-order differential dynamic features are calculated. In order to capture the dynamic changes of the signal over time, these differential features can provide information on the changes of the signal over time.

[0094] Example 2 further optimizes the AI ​​speech recognition, reasoning, and generation system provided in Example 1. Specifically, Figure 1 As shown, in S201, the specific conversion steps include:

[0095] Input data preparation: Signal samples extracted from the preprocessed and framed speech signal are input into FFT;

[0096] Apply the Cooley-Tukey algorithm to obtain the frequency domain signal: Apply the formula Convert the time domain signal x[n] into the frequency domain signal X[k], where X[k] represents the complex value at the kth frequency component, and k is the index of the frequency domain ranging from 0 to N-1, where is the summation symbol, which means summing all N (from 0 to N-1). is a complex exponential function also called a rotation factor, j is an imaginary unit, and satisfies j 2 =-1, is the phase difference, which determines the rotation speed and direction of each frequency component. The entire formula represents the discrete Fourier transform of the input signal x[n], converting it from the time domain to the frequency domain. For each frequency component k, a complex number X[k] is calculated. This complex number contains the amplitude and phase information of the frequency component.

[0097] like Figure 1 As shown, in S202, the specific calculation steps include:

[0098] S2020, first-order difference calculation:

[0099] Set an empty first-order difference sequence Δx[n];

[0100] Calculate the first-order difference: For each sample point x[n] in the time domain signal, its first-order difference Δ[n] is defined as the difference between the current sample point x[n] and the previous sample point x[n-1], that is, Δx[n] = x[n] - x[n-1]. For the first sample point of the sequence, since there is no previous sample point, the first-order difference is usually set to an initial value of 0:

[0101] Forming a first-order difference sequence: Repeat the above calculation process until all sample points in the time domain signal are processed, thereby obtaining a complete first-order difference sequence Δx[n];

[0102] S2021, second-order difference calculation:

[0103] Set an empty second-order difference sequence Δ 2 x[n];

[0104] Calculate the second-order difference: Second-order difference Δ 2x[n] is the difference of the first-order differences, calculated as:

[0105] Δ 2 x[n] = Δx[n] - Δx[n-1], which is the difference between the first-order difference Δx[n] of the current sample point and the first-order difference Δx[n-1] of the previous sample point. For the beginning of the sequence (especially the first and second sample points), since there are not enough preceding sample points for calculation, the first two second-order differences are set to 0;

[0106] Forming a second-order difference sequence: Repeat the above calculation process until all sample points in the first-order difference sequence are processed, thereby obtaining a complete second-order difference sequence Δ 2 x[n], through this process, the first-order and second-order differential dynamic characteristics of the time domain signal can be obtained, which are helpful for analyzing the dynamic changes and trends of the signal.

[0107] In S2, the feature extraction step also includes:

[0108] S203, extracting short-term, medium-term and long-term features:

[0109] Short-time feature extraction: Calculate the energy of each frame signal. The short-time energy is the sum of the squares of all sample values ​​in the frame. The specific formula is: Where E is the short-time energy, x[n] is the sample in the frame, n is the length of the frame, and the number of times the sign of each frame signal changes is counted, which is recorded as the short-time zero-crossing rate (the zero-crossing rate reflects the frequency of the signal);

[0110] Mid-time feature extraction: Extract the amplitude and power spectrum as features from the spectrum obtained after FFT conversion. Pass the spectrum obtained by FFT through the Mel filter bank, map the frequency to the Mel scale, take the logarithm of the output of each filter, and then perform discrete cosine transform (DCT) to obtain MFCC (Mel frequency cepstral coefficient). MFCC is a commonly used mid-time feature in speech recognition.

[0111] Long-term feature extraction: The autocorrelation function (ACF) method is used to calculate the fundamental frequency of each frame. The specific formula is: Where m is the delay. Multiply the signal x[n] by its version x[nm] delayed by m units. Sum all the products from m to n-1 to get the autocorrelation function r[m]. When m=0, r[0] represents the total energy of the signal. For periodic signals, the ACF will have a peak at the delay corresponding to the signal period. The shape and peak position of the ACF can provide information about the periodic component in the signal. The period is calculated based on the peak value that appears. The specific formula is: T is the period (seconds) F s is the sampling rate (samples / second), the fundamental frequency F0 is the inverse of the period, through the formula: Calculated;

[0112] The fundamental frequencies of each frame are connected to form a pitch contour, which reflects the intonation changes of speech, measures duration, and estimates speech rate by counting the number of syllables and words uttered per unit time.

[0113] Embodiment 3 further optimizes the AI ​​speech recognition, reasoning, and generation system provided in Embodiment 1 or 2. Specifically, Figure 1 As shown, in S3, the specific fusion steps include:

[0114] Feature concatenation: short-time features, medium-time features, long-time features, and first-order and second-order differential features are concatenated to form a long feature vector. The short-time feature is a D1-dimensional vector, the medium-time feature is a D2-dimensional vector, and the long-time feature is a D3-dimensional vector. The concatenated features will be a (D1+D2+D3)-dimensional vector. After concatenation, some post-processing may be required on the fused features, such as standardization and normalization, to ensure that the feature values ​​are on the same scale, which helps the machine learning model to learn and predict better.

[0115] like Figure 1 As shown in the figure, using long feature vectors to train a deep learning model includes the following steps:

[0116] Data preparation: Match the fused feature vectors to form a training data set;

[0117] Model selection: Choose a deep neural network (DNN);

[0118] Training process: Use the training data set to train the model and update the model parameters through the back-propagation algorithm and gradient descent optimizer.

[0119] Example 4, please refer to Figure 2 ,AI speech recognition, reasoning, and generation method, including the following steps:

[0120] Step 1: Voice input and preprocessing: The user inputs voice through the voice input module, and the adaptive noise suppression module filters the background noise in real time, enhances the voice signal, and preprocesses the input voice signal, including noise reduction, standardization, and segmentation;

[0121] Step 2: Multi-scale feature extraction: extract features from the preprocessed speech signal;

[0122] Extract short-time features, including short-time energy and short-time zero-crossing rate, to capture the instantaneous characteristics of speech signals;

[0123] Extract mid-time features, including MFCC, to reflect the spectral characteristics of speech;

[0124] Extract long-term features, including fundamental frequency and pitch contour, to capture the intonation and rhythm information of speech;

[0125] Compute first-order and second-order differential dynamic features to capture the dynamic changes of speech signals over time;

[0126] Step 3: Feature fusion: The extracted short-time features, medium-time features, long-time features, and first-order and second-order differential features are fused in series to form a comprehensive feature vector, which is rich in multi-scale speech information. The fused features are post-processed, including standardization and normalization.

[0127] The following steps are also included:

[0128] Step 4: Model training and recognition: Use the fused feature vector to train the deep neural network, update the model parameters through the back propagation algorithm and gradient descent optimizer to minimize the prediction error, input the speech signal to be recognized into the trained model for recognition, and output the recognition result.

[0129] The following steps are also included:

[0130] Step 5: Reasoning and generation: The reasoning module performs logical reasoning based on the recognized speech content, and the generation module generates corresponding outputs based on the reasoning results, including text, speech synthesis or other forms of response.

[0131] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection, an electrical connection, or communication with each other; it can be a direct connection, or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements, unless otherwise clearly defined. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0132] Obviously, the embodiments described above are only some embodiments of the present invention, rather than all embodiments. The preferred embodiments of the present invention are given in the accompanying drawings, but they do not limit the patent scope of the present invention. The present invention can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive. Although the present invention has been described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific embodiments, or to perform equivalent replacements for some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of the present invention, directly or indirectly used in other related technical fields, is also within the scope of patent protection of the present invention.

Claims

1. AI speech recognition, reasoning, and generation system, characterized by: include: A voice input module, used to receive a user's voice input; Adaptive noise suppression module, which filters the background noise of the current environment through noise suppression algorithm and noise filtering model; The speech recognition module recognizes speech through a deep learning algorithm and optimizes the recognition accuracy of fast-speech speeds; A reasoning module, used for performing logical reasoning based on the recognized speech content; The generation module generates corresponding outputs based on the reasoning results; The speech recognition module recognizes speech through a deep learning algorithm, specifically including the following steps: S1. Data preprocessing: preprocess the input speech signal, including noise reduction, standardization and segmentation, to ensure signal quality; S2, multi-scale feature extraction: extract features of speech and extract features of corresponding scales, including short-term features, medium-term features and long-term features, to fully capture the local and global information of speech signals; S3, feature fusion: The multi-scale features extracted in S2 are fused using a serial fusion strategy to form a comprehensive feature vector that is rich in multi-scale speech information; S4, model training: use the fused feature vector to train a deep learning model; S5, recognition output: the speech signal to be recognized is recognized by the above-mentioned trained model, and the recognition result is output; In S2, the feature extraction step specifically includes: S200, pre-treatment: Pre-emphasis: Use a high-pass filter to improve the signal-to-noise ratio of the signal in the high-frequency part. The pre-emphasis method is a first-order high-pass filter, and its formula is y[n]=x[n]-αx[n-1], where α is the filter coefficient, which ranges from 0.9 to 1.0, y[n] is the value of the current output signal, which is a point in the time series data, n represents the current sample point, x[n] is the value of the current input signal, x[n-1] is the input signal value of the previous sample point, and n-1 is the time point before the current time point; Framing: The speech signal is divided into short-term stable frames. The framing operation is achieved by multiplying the Hamming window with the original signal. Where w(n) is the weight value of the window function at the nth sampling point, n is the sample, and N is the window length. The Hamming window is applied to the original speech signal and multiplied by each sample of the corresponding segment of the original signal to obtain a frame of windowed signal. The sliding of the Hamming window on the signal divides the entire speech signal into multiple overlapping frames. S201, extraction: performing FFT on each frame signal to convert the time domain signal into a frequency domain signal; S202, dynamic feature calculation: in order to capture the dynamic information in the speech signal, calculate the first-order and second-order differential dynamic features; In S201, the specific conversion steps include: Input data preparation: Signal samples extracted from the preprocessed and framed speech signal are input into FFT; Apply the Cooley-Tukey algorithm to obtain the frequency domain signal: Apply the formula Convert the time domain signal x[n] into the frequency domain signal X[k], where X[k] represents the complex value at the kth frequency component, and k is the index of the frequency domain ranging from 0 to N-1, where is the summation symbol, which means summing all N. is a complex exponential function, j is an imaginary unit, and j satisfies 2 =-1, is the phase difference, which determines the rotation speed and direction of each frequency component; In S202, the specific calculation steps include: S2020, first-order difference calculation: Set an empty first-order difference sequence Δx[n]; Calculate the first-order difference: For each sample point x[n] in the time domain signal, its first-order difference Δx[n] is defined as the difference between the current sample point x[n] and the previous sample point x[n-1], that is, Δx[n] = x[n] - x[n-1]; Forming a first-order difference sequence: Repeat the above calculation process until all sample points in the time domain signal are processed, thereby obtaining a complete first-order difference sequence Δx[n]; S2021, second-order difference calculation: Set an empty second-order difference sequence Δ 2 x[n]; Calculate the second-order difference: Second-order difference Δ 2 x[n] is the difference of the first-order differences, calculated as: Δ 2 x[n] = Δx[n] - Δx[n-1], which is the difference between the first-order difference Δx[n] of the current sample point and the first-order difference Δx[n-1] of the previous sample point; Forming a second-order difference sequence: Repeat the above calculation process until all sample points in the first-order difference sequence are processed, thereby obtaining a complete second-order difference sequence Δ 2 x[n]; In S2, the feature extraction step further includes: S203, extracting short-term, medium-term and long-term features: Short-time feature extraction: Calculate the energy of each frame signal. The short-time energy is the sum of the squares of all sample values ​​in the frame. The specific formula is: Where E is the short-time energy, x[n] is the sample in the frame, n is the length of the frame, and the number of times the sign of each frame signal changes is counted, which is recorded as the short-time zero-crossing rate; Mid-time feature extraction: Extract the amplitude and power spectrum as features from the spectrum obtained after FFT conversion, pass the spectrum obtained by FFT through the Mel filter bank, map the frequency to the Mel scale, take the logarithm of the output of each filter, and then perform discrete cosine transform to obtain MFCC; Long-term feature extraction: Use the autocorrelation function method to calculate the fundamental frequency of each frame. The specific formula is: Where m is the delay. Multiply the signal x[n] by its version x[nm] delayed by m units. Sum all the products from m to n-1 to get the autocorrelation function T[m]. For periodic signals, the ACF will have a peak at the delay corresponding to the signal period. The period is calculated based on the peak. The specific formula is: T is the period F s is the sampling rate, the fundamental frequency F0 is the inverse of the period, through the formula: Calculated; Concatenate the fundamental frequencies of each frame to form a pitch contour.

2. The AI ​​speech recognition, reasoning, and generation system according to claim 1, characterized in that: In S3, the specific fusion steps include: Feature concatenation: short-term features, medium-term features, long-term features, and first-order and second-order differential features are concatenated to form a long feature vector.

3. The AI ​​speech recognition, reasoning, and generation system according to claim 2, characterized in that: In S4, the long feature vector is used to train the deep learning model, which specifically includes the following steps: Data preparation: Match the fused feature vectors to form a training data set; Model selection: Choose a deep neural network; Training process: Use the training data set to train the model and update the model parameters through the back-propagation algorithm and gradient descent optimizer.

4. An AI speech recognition, reasoning, and generation method, using the AI ​​speech recognition, reasoning, and generation system as claimed in claim 3, characterized in that: The following steps are involved: Step 1: Voice input and preprocessing: The user inputs voice through the voice input module, and the adaptive noise suppression module filters the background noise in real time, enhances the voice signal, and preprocesses the input voice signal, including noise reduction, standardization, and segmentation; Step 2: Multi-scale feature extraction: extract features from the preprocessed speech signal; Extract short-time features, including short-time energy and short-time zero-crossing rate, to capture the instantaneous characteristics of speech signals; Extract mid-time features, including MFCC, to reflect the spectral characteristics of speech; Extract long-term features, including fundamental frequency and pitch contour, to capture the intonation and rhythm information of speech; Compute first-order and second-order differential dynamic features to capture the dynamic changes of speech signals over time; Step 3: Feature fusion: The extracted short-time features, medium-time features, long-time features, and first-order and second-order differential features are fused in series to form a comprehensive feature vector that is rich in multi-scale speech information.

5. The AI ​​speech recognition, reasoning and generation method according to claim 4, characterized in that: The following steps are also included: Step 4: Model training and recognition: Use the fused feature vector to train the deep neural network, update the model parameters through the back propagation algorithm and gradient descent optimizer to minimize the prediction error, input the speech signal to be recognized into the trained model for recognition, and output the recognition result.

6. The AI ​​speech recognition, reasoning, and generation method according to claim 5, characterized in that: The following steps are also included: Step 5: Reasoning and generation: The reasoning module performs logical reasoning based on the recognized speech content, and the generation module generates corresponding outputs based on the reasoning results.