Speech recognition method, device, equipment and storage medium

Abstract

The present invention provides a speech recognition method, device, equipment and storage medium, comprising collecting audio signals; preprocessing the audio signals to obtain target audio signals; extracting target acoustic features of the target audio signals; using a dynamic time warping algorithm to perform similarity matching between the target acoustic features and pre-stored reference acoustic features, and outputting the target acoustic features with the highest similarity as results; wherein the step of preprocessing the speech signals comprises pre-emphasizing the audio signals, and performing non-overlapping framing on the pre-emphasized audio signals to obtain a plurality of non-overlapping signal frames, adding a connection frame to the front end or the back end of each signal frame, and windowing each signal frame after the connection frame is added, so that the frames are non-overlapping and coherent, thereby ensuring the validity of the acoustic features and greatly reducing the computational complexity.

Description

Technical Field

[0001] The present invention relates to the field of speech recognition technology, and in particular to a speech recognition method, device, equipment and storage medium. Background Art

[0002] Speech recognition technology has made significant progress in recent years and is widely used in fields such as intelligent voice assistants, voice-controlled devices, and human-computer interaction systems; these applications not only improve user experience, but also promote technological development and innovation.

[0003] However, the traditional method is usually to use Hidden Markov Model (HMM) and Deep Neural Networks (DNN) for speech recognition. This method is computationally intensive and consumes a lot of performance.

[0004] In embedded systems, there are usually limitations in computing resources, storage space, and energy consumption. This application uses dynamic time warping (DTW) for speech recognition. The dynamic time warping algorithm has the advantages of easy deployment, no need for supervision, and low resource consumption. By improving the dynamic time warping algorithm, it becomes possible to implement speech recognition in resource-constrained embedded systems. Summary of the invention

[0005] The purpose of the present invention is to provide a speech recognition method, apparatus, device and storage medium to solve the problem mentioned in the above background technology that the traditional method in speech recognition technology has a large amount of calculation and consumes a lot of performance.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a speech recognition method, applied to an embedded system, the method steps comprising: collecting audio signals; preprocessing the audio signals to obtain target audio signals; extracting target acoustic features of the target audio signals; using a dynamic time warping algorithm to perform similarity matching between the target acoustic features and pre-stored reference acoustic features; outputting the target acoustic features with the highest similarity as the results; wherein the step of preprocessing the audio signals comprises pre-emphasizing the audio signals, and performing non-overlapping framing on the pre-emphasized audio signals to obtain a number of non-overlapping signal frames, adding a connection frame at the front end or the back end of each of the signal frames, and windowing each of the signal frames after the connection frame is added.

[0007] Optionally, the step of dividing the pre-emphasized audio signal into frames without overlap includes: dividing a segment of the pre-emphasized audio signal into a plurality of frames in a non-overlapping segmentation manner, and sampling points between two adjacent frames do not overlap with each other.

[0008] Optionally, the step of adding a connection frame at the front end or the back end of each signal frame includes: adding 8%-12% of the number of sampling points in the signal frame at the front end or the back end of each signal frame, and taking an integer number of sampling points as the connection frame.

[0009] Optionally, the step of preprocessing the audio signal to obtain the target audio signal also includes: presetting the endpoint threshold of the feature estimation as the speech activation threshold and the speech exit threshold; calculating the short-time energy and the short-time zero-crossing rate of each windowed signal frame; substituting the short-time energy and the short-time zero-crossing rate into a preset function, and adjusting the weight of the short-time energy and the weight of the short-time zero-crossing rate to obtain the feature estimation of the signal frame; and taking the signal frames corresponding to several consecutive feature estimations within the endpoint threshold range as a target audio signal.

[0010] Optionally, the preset function is: Among them, V is the feature estimate, is the weight of short-time energy, E is the short-time energy value, It is the maximum value of the 16-bit integer data type. is the minimum value of the 16-bit integer data type, L is the number of sampling points in each frame, is the weight of the short-time zero-crossing rate, and Z is the short-time zero-crossing rate value.

[0011] Optionally, the step of extracting the target acoustic features of the target audio signal includes: using fast Fourier transform to split the target audio signal into odd-sequence time domain signals and even-sequence time domain signals, and then performing recursive operations and butterfly operations to convert the target audio signal into a frequency domain signal; using a Mel filter group to filter the frequency domain signal; performing discrete cosine transform on the filtered frequency domain signal to obtain Mel-frequency cepstral coefficients, and using the Mel-frequency cepstral coefficients as target acoustic features.

[0012] Optionally, the step of extracting the target acoustic features of the target audio signal further includes: projecting the target acoustic features into a principal component space for dimensionality reduction using a principal component analysis method.

[0013] On the other hand, the present invention also provides a speech recognition device, which is applied to an embedded system, and the speech recognition device includes: an acquisition module, which is used to acquire audio signals; a preprocessing module, which is used to preprocess the audio signals to obtain target audio signals; a feature extraction module, which is used to extract target acoustic features of the target audio signals; a recognition module, which is used to use a dynamic time warping algorithm to perform similarity matching between the target acoustic features and pre-stored historical acoustic features; and output the target acoustic features with the highest similarity as a result; wherein the preprocessing module includes: a pre-emphasis module, which is used to pre-emphasize the audio signal; a non-overlapping framing module, which is used to perform non-overlapping framing on the pre-emphasized audio signal to obtain a number of non-overlapping signal frames, and add a connection frame at the front end or the back end of each of the signal frames; and a windowing module, which is used to window each of the signal frames after the connection frame is added.

[0014] On the other hand, the present invention further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned speech recognition method when executing the computer program.

[0015] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the above-mentioned speech recognition method when executed by a processor.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] In the traditional framing method, in order to retain the continuous features of the previous and next frames, a 20%-50% overlapping area is usually set. This method will result in a large amount of calculation and high computational complexity. In the present invention, the pre-emphasized audio signal is divided into non-overlapping frames to obtain a number of non-overlapping signal frames, and a connecting frame is added to the front end or the back end of each signal frame, so that the frames are non-overlapping and coherent, which ensures the effectiveness of the acoustic features while greatly reducing the computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The figure is a schematic flow chart of the steps of the method of the present invention.

[0019] Figure 2 The figure is a schematic flow chart of the steps of the pretreatment method of the present invention.

[0020] Figure 3 It is a schematic diagram for comparing the short-time cooling waveforms of two audio segments of the present invention.

[0021] Figure 4 Schematic diagram of a time domain waveform of an audio signal sampled in an embodiment of the present invention.

[0022] Figure 5 FIG. 4 is a schematic diagram of a time domain waveform of a reference audio signal pre-stored in an embodiment of the present invention.

[0023] Figure 6 Schematic diagram of a dynamic time warping path in an embodiment of the present invention.

[0024] Figure 7 Schematic diagram of the structure of an embedded speech recognition device in an embodiment of the present invention.

[0025] Figure 8 Schematic diagram of the structure of an embedded speech recognition device in an embodiment of the present invention.

[0026] In the figure: 1-acquisition module, 2-preprocessing module, 3-feature extraction module, 4-recognition module, 5-pre-emphasis module, 6-non-overlapping framing module, 7-windowing module, 8-embedded speech recognition device, 9-processor, 10-memory, 11-storage medium, 12-operating system, 13-data, 14-application, 15-network interface, 16-input and output interface, 17-power supply. DETAILED DESCRIPTION

[0027] The scheme 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, rather than all the embodiments.

[0028] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0029] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present application refers to the presence of features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be an intermediate element. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The term "and / or" used herein includes all or any unit and all combinations of one or more associated listed items.

[0030] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art to which this application belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with those in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as here.

[0031] It should be understood that the sequence numbers and sizes of the steps in this embodiment do not mean the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

[0032] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0033] Please refer to Figure 1 and Figure 2 A speech recognition method of the present invention is applied to an embedded system, and the method steps include:

[0034] S100: Collect audio signals.

[0035] Specifically, the speech recognition method of the present invention is applied in an embedded system to capture a real-time audio signal through an audio acquisition device.

[0036] S200: Preprocess the audio signal to obtain a target audio signal.

[0037] Specifically, preprocessing includes denoising, gain control and / or endpoint detection, wherein endpoint detection includes pre-emphasis, framing and windowing; endpoint detection is used to determine the starting and ending positions of the audio signal, so that valid audio segments can be extracted from the audio stream. The effectiveness of endpoint detection will directly affect the accuracy of speech recognition and the occupancy rate of system resources.

[0038] More specifically, the pre-emphasis formula is ,in is the emphasized audio signal, is the audio signal before emphasis, a is the pre-emphasis coefficient and is , N is a natural number.

[0039] More specifically, windowing includes a Hamming window, and the specific formula is: ,in, is the nth point of the window function, S is the window length, and N is a natural number.

[0040] More specifically, the signal of each frame after windowing is expressed as ,in is the frame signal after windowing, for The nth sampling point of the frame, is the nth point of the window function, S is the window length, and N is a natural number.

[0041] S300: extracting target acoustic features of a target audio signal.

[0042] Specifically, the target acoustic features of the target audio signal are extracted by using a Mel-frequency cepstral coefficient method, a linear prediction cepstral coefficient method and / or a perceptual linear prediction method, wherein the target acoustic features include Mel-frequency cepstral coefficients and / or linear prediction coefficients.

[0043] S400, using a dynamic time warping algorithm to perform similarity matching between the target acoustic feature and the pre-stored reference acoustic feature, and outputting the target acoustic feature with the highest similarity as the result.

[0044] Please refer to Figure 4 , Figure 5 and Figure 6 Specifically, the dynamic time warping algorithm uses the nonlinear transformation of the time axis to find the best matching path between two sequences; since an audio segment is affected by factors such as intonation, tone, and speech speed, the lengths of two audios to be recognized are basically impossible to be the same, for example Figure 3The two audio segments shown in the figure, although they both represent the semantics of "hello", have great differences in the timing sense. In this scenario, we can use the dynamic time warping algorithm to find the minimum distance between the two audio segments. The smaller the distance, the higher the similarity between the two audio segments. By storing the reference isolated word audio data in the device in advance, the acquired audio sequence and the reference audio data sequence are operated by the dynamic time warping algorithm to identify the closest audio data and output it as the recognition result.

[0045] The step of preprocessing the audio signal includes:

[0046] S201: Pre-emphasize the audio signal.

[0047] Specifically, pre-emphasis is used to enhance the high-frequency signal in the audio signal.

[0048] S202: Perform non-overlapping frame division on the pre-emphasized audio signal to obtain a plurality of non-overlapping signal frames.

[0049] Specifically, the relationship between the number of frames without overlapping framing of the pre-emphasized audio signal, the number of sampling points contained in each frame, and the number of overlapping sampling points between adjacent frames is expressed as follows: , where X is the total number of frames of an audio signal, n is the number of sampling points contained in an audio signal, L is the number of sampling points in each frame, and M is the number of overlapping sampling points between two adjacent frames; it can be seen that the larger the number of overlapping sampling points M, the larger the total number of frames of the audio signal, and the greater the amount of calculation and difficulty of calculation in subsequent feature extraction. The present application sets the overlapping sampling points to 0, that is, a non-overlapping setting, which greatly reduces the amount of calculation and difficulty of calculation, and can be suitable for the application environment of embedded systems.

[0050] S203: Add a connection frame at the front end or the back end of each signal frame.

[0051] Specifically, the backend connection position set of each frame is: , where P is a non-overlapping signal frame, is the position in the non-overlapping signal frame sampling point sequence, X is the total number of frames of an audio signal, and N is a natural number; the front connection position set of each frame is: , where P is a sampling point sequence, X is the total number of frames of an audio signal, and N is a natural number; by adding a connection frame after the back-end connection position of each frame or before the front-end connection position of each frame, the continuity of non-overlapping frames can be guaranteed.

[0052] S204, windowing is performed on each signal frame after a connection frame is added.

[0053] Specifically, windowing the audio signal of each frame can effectively avoid spectrum leakage and ensure the accuracy of speech recognition.

[0054] It is understandable that in the traditional framing method, in order to retain the continuous features of the previous and next frames, a 20%-50% overlapping area is usually set. This method will result in a large amount of calculation and high computational complexity. In the present invention, the pre-emphasized audio signal is divided into frames without overlapping to obtain a number of non-overlapping signal frames, and a connecting frame is added to the front or back end of each signal frame, so that the frames are non-overlapping and coherent, which ensures the effectiveness of the acoustic features while greatly reducing the computational complexity.

[0055] In some embodiments, the step of dividing the pre-emphasized audio signal into frames without overlap includes: dividing a segment of the pre-emphasized audio signal into a plurality of frames in a non-overlapping segmentation manner, and the sampling points between two adjacent frames do not overlap with each other.

[0056] Specifically, for general endpoint detection scenarios, system resources are usually abundant, and methods such as deep learning, machine learning, and spectral feature extraction can be used, but these are very difficult to implement for embedded systems. Based on this situation, this application adopts an endpoint detection method, highlighting the modifications to the frame division method and the composite weighted model, which can achieve better detection effects with lower computational complexity.

[0057] In some embodiments, the step of adding a connection frame at the front end or the back end of each signal frame includes: adding 8%-12% of the number of sampling points in the signal frame at the front end or the back end of each signal frame, and taking an integer number of sampling points as the connection frame.

[0058] Specifically, the number of sampling points in the connecting frame is 8%-12% of the number of sampling points in the signal frame, and is rounded to an integer; for example, at the connecting position after the current frame, 10% of the sampling points of the current frame are added as the connecting frame, that is, the number of sampling points in the connecting frame is , where L is the number of sampling points in each frame; at this time, the added connection frame is T , ,in, is the position in the sampling point sequence of the connecting frame, and Take the integer, is the position in the sequence of sampling points of non-overlapping signal frames.

[0059] It should be understood that the present application simulates features by self-building connection points, and the corresponding range of adding 8%-12% of the number of sampling points in the signal frame as connection frames is a relatively appropriate range. If the number of sampling points is less than 8% of the number of sampling points in the current frame, it is easy to cause the features to be invalid. If the number of sampling points is greater than 12% of the number of sampling points in the current frame, it will cause over-fitting of the features and affect the overall performance of the system.

[0060] In some embodiments, the step of preprocessing the audio signal to obtain the target audio signal also includes: presetting the endpoint threshold of the feature estimation as the speech activation threshold and the speech exit threshold; calculating the short-time energy and the short-time zero-crossing rate of each windowed signal frame; substituting the short-time energy and the short-time zero-crossing rate into a preset function, and adjusting the weight of the short-time energy and the weight of the short-time zero-crossing rate to obtain the feature estimation of the signal frame; and taking the signal frames corresponding to several consecutive feature estimations within the endpoint threshold range as a target audio signal.

[0061] Specifically, during the endpoint detection process, the voice activation threshold and voice exit threshold need to be pre-set. The conditions that each frame in the audio signal needs to meet are: ,in, is the position in the sampling point sequence of the non-overlapping signal frame, is the speech activation threshold, is the speech exit threshold, L is the number of sampling points in each frame, and N is a natural number.

[0062] More specifically, the short-time energy calculation formula is: , where E is the short-time energy value, is the position in the sampling point sequence of the non-overlapping signal frame; the short-time zero-crossing rate calculation formula is: , where Z is the short-time zero-crossing rate value, is the position in the sampling point sequence of the non-overlapping signal frame, and the sgn function is .

[0063] In some embodiments, the preset function is: Among them, V is the feature estimate, is the weight of short-time energy, E is the short-time energy value, It is the maximum value of the 16-bit integer data type. is the minimum value of the 16-bit integer data type, L is the number of sampling points in each frame, is the weight of the short-time zero-crossing rate, and Z is the short-time zero-crossing rate value.

[0064] Specifically, the present application substitutes the weights of short-time energy and short-time zero-crossing rate into a preset function according to each different embedded environment. The weights of short-time energy and short-time zero-crossing rate can be adjusted by calculation to obtain the optimal value, thereby avoiding algorithm performance problems caused by environmental differences and further improving the accuracy of speech recognition.

[0065] More specifically, according to the preset function, in order to obtain the most suitable weight in the current environment, multiple groups of weight values ​​are set to perform speech detection tasks respectively, taking setting five groups as an example: ,in is the weight of short-time energy, is the weight of the short-time zero-crossing rate. When setting the weight, first pre-mark the front end of the collected audio signal and backend The endpoint of the target audio signal corresponding to each set of weights is calculated. and backend The target endpoint is selected, and a set of weights corresponding to the minimum distance T between the target endpoint and the endpoint is selected and substituted into the preset function, which satisfies .

[0066] In some embodiments, the step of extracting the target acoustic features of the target audio signal includes: using fast Fourier transform to split the target audio signal into odd-sequence time domain signals and even-sequence time domain signals, and then performing recursive operations and butterfly operations to convert the target audio signal into a frequency domain signal; using a Mel filter group to filter the frequency domain signal; performing discrete cosine transform on the filtered frequency domain signal to obtain Mel-frequency cepstral coefficients, and using the Mel-frequency cepstral coefficients as the target acoustic features.

[0067] Specifically, splitting the target audio signal into odd-sequence time domain signals and even-sequence time domain signals, and then performing recursive operations and butterfly operations can effectively optimize the computing power required for the fast Fourier transform. Although the frame length is fixed, the complexity of the operation can be greatly reduced, which is more friendly to embedded devices.

[0068] More specifically, the formula for splitting the target audio signal into odd-sequence time domain signals and even-sequence time domain signals is: ,in is an odd sequence, is an even sequence, L is the number of sampling points in each frame, and N is a natural number.

[0069] More specifically, the recursive operation formula is ,

[0070] The formula for butterfly operation is This method can effectively optimize the computing power required for the fast Fourier transform. Although the frame length is fixed, it can greatly reduce the complexity of the calculation and is more friendly to embedded devices.

[0071] More specifically, after converting the target audio signal into a frequency domain signal, the frequency domain signal is squared by the absolute value to obtain the power spectrum density; the power spectrum density is filtered through a Mel filter bank, and then the logarithm is taken and a discrete cosine transform is performed to obtain the Mel-frequency cepstral coefficients, which can be used as the target acoustic features; or the dynamic features can be obtained by calculating the first-order and second-order differences of the Mel-frequency cepstral coefficients; the dynamic features and the Mel-frequency cepstral coefficients are normalized to obtain a set of comprehensive features, which are then used as the target acoustic features.

[0072] In some embodiments, the step of extracting the target acoustic features of the target audio signal further includes: projecting the target acoustic features into a principal component space for dimensionality reduction using a principal component analysis method.

[0073] Specifically, by using the principal component analysis method to reduce the dimension of the target acoustic features, the present application can reduce the 40 filters commonly used in traditional methods to 12, thereby retaining the main features of the audio data while reducing the density and amount of computation, making it possible to implement speech recognition on resource-constrained embedded systems, and can be designed and reused on any embedded system, which has high practical significance.

[0074] Please refer to Figure 7 On the other hand, the present invention also provides a speech recognition device, which is applied to an embedded system. The speech recognition device includes: an acquisition module 1, which is used to acquire audio signals; a preprocessing module 2, which is used to preprocess the audio signals to obtain target audio signals; a feature extraction module 3, which is used to extract target acoustic features of the target audio signals; a recognition module 4, which is used to use a dynamic time warping algorithm to match the target acoustic features with pre-stored historical acoustic features for similarity; and output the target acoustic features with the highest similarity as the result; wherein the preprocessing module 2 includes a pre-emphasis module 5, which is used to pre-emphasize the audio signal; a non-overlapping framing module 6, which is used to non-overlapping framing the pre-emphasized audio signal to obtain a number of non-overlapping signal frames, and to add a connection frame at the front end or the back end of each signal frame; and a windowing module 7, which is used to window each signal frame after the connection frame is added.

[0075] Please refer to Figure 8 On the other hand, the present invention also provides a computer device, including a memory 10 and a processor 9, the memory 10 stores a computer program, and the processor 9 implements the steps of the above-mentioned speech recognition method when executing the computer program.

[0076] Please refer to Figure 8 On the other hand, the present invention also provides a computer-readable storage medium 11 on which a computer program is stored, and when the computer program is executed by the processor 9, the steps of the above-mentioned speech recognition method are implemented.

[0077] Specifically, Figure 8 1 is a schematic diagram of the structure of an embedded speech recognition device provided by an embodiment of the present invention. The embedded speech recognition device 8 may have relatively large differences due to different configurations or performances, and may include one or more processors 9 (central processing units, CPU) (for example, one or more processors 9) and memories 10, and one or more storage media 11 (for example, one or more mass storage devices) storing applications 14 or data 13; wherein the memories 10 and the storage media 11 may be short-term storage or permanent storage; further, the processor 9 may be configured to communicate with the storage medium 11, and execute a series of instruction operations in the storage medium 11 on the embedded speech recognition device 8 to implement the steps of the speech recognition method provided in the above-mentioned method embodiments.

[0078] The embedded speech recognition device 8 may also include one or more power supplies 17, one or more wired or wireless network interfaces 15, one or more input and output interfaces 16, and / or one or more operating systems 12, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be appreciated by those skilled in the art that Figure 8 The structure of the embedded speech recognition device shown does not constitute a limitation on the embedded speech recognition device, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0079] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0080] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium 11. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods; wherein, any reference to the memory 10, storage, database or other medium used in the embodiments provided by the present invention can include non-volatile and / or volatile memory 10; the non-volatile memory 10 can include a read-only memory 10 (ROM), a programmable ROM (PROM), an electrically programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM) or a flash memory; the volatile memory 10 can include a random access memory 10 (RAM) or an external cache memory 10; as an illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synch li nk) DRAM (SLDRAM), memory 10 bus (Rambus) direct RAM (RDRAM), direct memory 10 bus dynamic RAM (DRDRAM), and memory 10 bus dynamic RAM (RDRAM), etc.

[0081] The above are only embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent transformations made using the contents of the specification and drawings of the present invention, or directly or indirectly applied in related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A speech recognition method, characterized in that: Applied to an embedded system, the method steps include: Collect audio signals; Preprocessing the audio signal to obtain a target audio signal; Extracting target acoustic features of the target audio signal; A dynamic time warping algorithm is used to perform similarity matching between the target acoustic feature and a pre-stored reference acoustic feature, and the target acoustic feature with the highest similarity is output as a result; The step of preprocessing the audio signal includes pre-emphasizing the audio signal, and performing non-overlapping framing on the pre-emphasized audio signal to obtain a plurality of non-overlapping signal frames, adding a connection frame to the front end or the rear end of each signal frame, and windowing each signal frame after adding the connection frame; The step of adding a connection frame at the front end or the back end of each signal frame includes: At the front end or the back end of each of the signal frames, 8%-12% of the number of sampling points in the signal frame are added accordingly, and an integer number of sampling points are taken as the connection frame.

2. The speech recognition method according to claim 1, characterized in that: The step of performing non-overlapping framing on the pre-emphasized audio signal comprises: A segment of the pre-emphasized audio signal is divided into a plurality of frames in a non-overlapping segmentation manner, and the sampling points between two adjacent frames do not overlap with each other.

3. The speech recognition method according to claim 1, characterized in that: The step of preprocessing the audio signal to obtain the target audio signal further comprises: Presetting endpoint thresholds of feature estimation as a speech activation threshold and a speech exit threshold; Calculate the short-time energy and short-time zero-crossing rate of each windowed signal frame; Substituting the short-time energy and the short-time zero-crossing rate into a preset function, and adjusting the weight of the short-time energy and the weight of the short-time zero-crossing rate to obtain a feature estimation of the signal frame; The signal frames corresponding to a plurality of continuous feature estimations within an endpoint threshold range are taken as a target audio signal.

4. The speech recognition method according to claim 3, characterized in that: The preset function is: Among them, V is the feature estimate, is the weight of short-time energy, E is the short-time energy value, It is the maximum value of the 16-bit integer data type. is the minimum value of the 16-bit integer data type, L is the number of sampling points in each frame, is the weight of the short-time zero-crossing rate, and Z is the short-time zero-crossing rate value.

5. The speech recognition method according to claim 1 or 4, characterized in that: The step of extracting the target acoustic feature of the target audio signal comprises: Using fast Fourier transform to split the target audio signal into an odd sequence time domain signal and an even sequence time domain signal, and then performing recursive operation and butterfly operation to convert the target audio signal into a frequency domain signal; Using a Mel filter bank to filter the frequency domain signal; The filtered frequency domain signal is subjected to discrete cosine transform to obtain Mel-frequency cepstral coefficients, and the Mel-frequency cepstral coefficients are used as target acoustic features.

6. The speech recognition method according to claim 1, characterized in that: The step of extracting the target acoustic feature of the target audio signal also includes: The target acoustic features are projected into the principal component space for dimensionality reduction using principal component analysis.

7. A speech recognition device, characterized in that: Applied to an embedded system, the speech recognition device comprises: An acquisition module, used for acquiring audio signals; A preprocessing module, used for preprocessing the audio signal to obtain a target audio signal; A feature extraction module, used to extract target acoustic features of the target audio signal; A recognition module, used for matching the target acoustic feature with the pre-stored historical acoustic features by similarity using a dynamic time warping algorithm, and outputting the target acoustic feature with the highest similarity as a result; Wherein, the preprocessing module includes: A pre-emphasis module, used for pre-emphasis of the audio signal; A non-overlapping framing module is used to perform non-overlapping framing on the pre-emphasized audio signal to obtain a plurality of non-overlapping signal frames, and to add a connection frame at the front end or the rear end of each signal frame; A windowing module, used for performing windowing on each of the signal frames after adding the connection frame; The step of adding a connection frame at the front end or the back end of each signal frame includes: At the front end or the back end of each of the signal frames, 8%-12% of the number of sampling points in the signal frame are added accordingly, and an integer number of sampling points are taken as the connection frame.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the speech recognition method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the speech recognition method according to any one of claims 1 to 6 are implemented.

Images