A method and device for on-board speech recognition of agricultural vehicles

Abstract

The invention relates to the technical field of speech recognition, and specifically discloses a method and device for on-board speech recognition of an agricultural vehicle, the method comprising: acquiring ambient sound data of the agricultural vehicle, finely filtering background noise in the ambient sound data, and obtaining finely filtered noise-free ambient sound data; extracting pure speech data from the finely filtered noise-free ambient sound data; determining the language type of the pure speech data; recognizing the dialect semantics in the pure speech data based on a preset dialect semantic recognition model of the language type in the pure speech data; performing standard language conversion on the dialect semantics in the pure speech data, and obtaining the standard language semantics in the pure speech data; sending the standard language semantics in the pure speech data to an agricultural vehicle control system; in an agricultural vehicle use environment with relatively large noise, realizing the pure extraction of speech data, and realizing the semantic conversion of the extracted speech data of different language types.

Description

Technical Field

[0001] The present invention relates to the technical field of speech recognition, and in particular to a speech recognition method and device mounted on an agricultural vehicle. Background Art

[0002] At present, in order to simplify the process of sending some simple commands to agricultural vehicles (such as starting / stopping the air conditioner, starting / stopping the wipers, and starting / closing the windows) to improve the driving experience of agricultural vehicle drivers, some agricultural vehicle manufacturers will install on-board voice recognition equipment or functional modules on agricultural vehicles to facilitate the control of agricultural vehicles by drivers who cannot adapt to the cumbersome operation methods of agricultural vehicles.

[0003] However, agricultural vehicles usually work outdoors. In agricultural work scenarios, the in-vehicle voice recognition process will be affected by various noises, such as road noise, engine noise, agricultural work noise, etc. These noises will interfere with voice recognition and reduce recognition accuracy. In addition, because different agricultural drivers may use different dialects, the difficulty of in-vehicle voice recognition for agricultural vehicles is greatly increased, and its recognition accuracy is also reduced.

[0004] Therefore, the present invention proposes a method and device for on-board speech recognition of an agricultural vehicle. Summary of the invention

[0005] The present invention provides an agricultural vehicle on-board speech recognition method and device, which are used to realize the pure extraction of speech data in an agricultural vehicle use environment with relatively large noise, and determine the language type of the pure speech data and the semantic conversion of the extracted speech data of different language types, thereby greatly improving the accuracy of agricultural vehicle on-board speech recognition.

[0006] The present invention provides an agricultural vehicle-mounted speech recognition method, comprising:

[0007] S1: Acquire the ambient sound data of agricultural vehicles, finely filter the background noise in the ambient sound data, and obtain finely filtered noise-free ambient sound data;

[0008] S2: Extract pure speech data from the finely filtered noise-free ambient sound data;

[0009] S3: Determine the language type of the clean speech data;

[0010] S4: Based on the preset dialect semantics recognition model of the language type in the clean speech data, the dialect semantics in the clean speech data are recognized;

[0011] S5: performing standard language conversion on the dialect semantics in the clean speech data to obtain the standard language semantics in the clean speech data;

[0012] S6: Send the standard language semantics in the clean speech data to the agricultural vehicle control system.

[0013] Preferably, S1: obtaining ambient sound data of agricultural vehicles, finely filtering background noise in the ambient sound data, and obtaining finely filtered noise-free ambient sound data, including:

[0014] S101: Based on recording devices arranged at different positions of the agricultural vehicle, recording data of the agricultural vehicle within a preset voice escape period is obtained as ambient sound data of the agricultural vehicle;

[0015] S102: removing regular background noise and irregular background noise in the ambient sound data to obtain finely filtered noise-free ambient sound data.

[0016] Preferably, S102: removing regular background noise and irregular background noise in the ambient sound data to obtain finely filtered noise-free ambient sound data, including:

[0017] Convert the audio signal corresponding to the ambient sound data from the time domain to the frequency domain to obtain the spectrum data of the ambient sound data;

[0018] Based on the bandpass filter with preset cutoff frequency, the regular background noise in the spectrum data is removed to obtain the coarse filtered noise-free ambient sound data;

[0019] The irregular background noise in the coarsely filtered noise-free ambient sound data is removed to obtain the finely filtered noise-free ambient sound data.

[0020] Preferably, removing irregular background noise from the coarsely filtered noise-free ambient sound data to obtain finely filtered noise-free ambient sound data includes:

[0021] Inputting the coarsely filtered noise-free ambient sound data into each irregular noise recognition model respectively, to obtain each irregular noise data contained in the coarsely filtered noise-free ambient sound data;

[0022] Each type of irregular noise data contained in the coarse-filtered noise-free ambient sound data is eliminated to obtain the fine-filtered noise-free ambient sound data.

[0023] Preferably, S2: extracting pure voice data from the finely filtered noise-free ambient sound data comprises:

[0024] Extracting sub-fine-filtered noise-free ambient sound data obtained by each recording device from the fine-filtered noise-free ambient sound data;

[0025] Determine the characteristic parameter limit value of each short-term analysis feature of each sub-fine filtered noise-free ambient sound data;

[0026] Among all the sub-fine-filtered noise-free ambient sound data, the sub-fine-filtered noise-free ambient sound data that does not have a characteristic parameter limit value exceeding the characteristic parameter threshold value of the corresponding short-time analysis feature is used as the positioning reference ambient sound data;

[0027] Extract all reference voice data of each human voice data segment from all positioning reference ambient sound data;

[0028] Based on all the reference speech data of all the human voice data segments, pure speech data is obtained.

[0029] Preferably, obtaining pure voice data based on all reference voice data of all human voice data segments includes:

[0030] Acquire speech waveforms of all reference speech data of each vocal speech data segment as all reference speech waveforms of each vocal speech data segment;

[0031] Matching the bands in different reference speech waveforms of each vocal speech data segment to obtain a plurality of bands belonging to different reference speech waveforms and matching each other;

[0032] Based on the time difference between each two bands of different reference speech waveforms belonging to a single human voice data segment and matching each other, determining the time difference between the corresponding two reference speech waveforms in the corresponding two bands matching each other;

[0033] The average of the time difference between each two reference speech waveforms of a single human voice data segment on all the mutually matching bands included is regarded as the time difference between the corresponding two reference speech waveforms of the single human voice data segment;

[0034] Determine the sound source range of the corresponding human voice data segment based on the time difference between each two reference voice waveforms of a single human voice data segment;

[0035] Based on the sound source range of each vocal speech data segment, all reference speech data of each vocal speech data segment are processed for loss compensation and then merged to obtain sub-pure speech data of each vocal speech data segment;

[0036] The sub-clean voice data of all the vocal speech data segments are spliced ​​in time sequence to obtain the pure voice data.

[0037] Preferably, determining the sound source range of the corresponding human voice data segment based on the time difference between every two reference voice waveforms of a single human voice data segment includes:

[0038] The difference in distance between the two corresponding recording devices and the sound source position is calculated based on the preset sound propagation speed and the time difference between each two reference speech waveforms of a single human voice data segment;

[0039] The sound source range of the human voice data segment is determined based on the position coordinates of all recording devices in a preset coordinate system and the difference between the distances between every two recording devices of a single human voice data segment and the sound source position.

[0040] Preferably, based on the sound source range of each vocal speech data segment, all reference speech data of each vocal speech data segment are combined after being compensated, so as to obtain sub-clean speech data of each vocal speech data segment, including:

[0041] Based on the coordinate representation of the sound source range of each vocal speech data segment in a preset coordinate system and the position coordinates of all recording devices in the preset coordinate system, determine all possible sound conduction paths between the sound source position of each vocal speech data segment and each recording device;

[0042] Based on all possible sound conduction paths from the sound source position of each vocal speech data segment to each recording device, all reference speech data of each vocal speech data segment are processed for loss compensation and then merged to obtain sub-pure speech data of each vocal speech data segment.

[0043] Preferably, based on all possible sound conduction paths from the sound source position of each vocal speech data segment to each recording device, all reference speech data of each vocal speech data segment are combined after being compensated, so as to obtain sub-clean speech data of each vocal speech data segment, including:

[0044] Determine the average value of the acoustic loss coefficients of all acoustic conduction paths between the sound source position of each vocal speech data segment and each recording device, and use it as the total acoustic loss coefficient of the corresponding reference speech data;

[0045] Based on the total loss coefficient of all reference voice data of each vocal speech data segment, the corresponding reference voice data is subjected to loss compensation processing to obtain all lossless voice data of each vocal speech data segment;

[0046] All lossless voice data of each vocal speech data segment are merged to obtain sub-clean voice data of each vocal speech data segment.

[0047] The present invention provides an agricultural vehicle-mounted speech recognition device, which is used to execute the agricultural vehicle-mounted speech recognition method described in any one of Embodiments 1 to 9, comprising:

[0048] The ambient sound acquisition module is used to acquire the ambient sound data of agricultural vehicles, finely filter the background noise in the ambient sound data, and obtain finely filtered noise-free ambient sound data;

[0049] A voice data extraction module is used to extract pure voice data from the finely filtered noise-free ambient sound data;

[0050] A language type determination module, used to determine the language type of the clean voice data;

[0051] A dialect semantics recognition module, used to recognize the dialect semantics in the clean voice data based on a preset dialect semantics recognition model of the language type in the clean voice data;

[0052] A standard language conversion module is used to convert the dialect semantics in the clean voice data into standard language semantics to obtain the standard language semantics in the clean voice data;

[0053] The recognition result sending module is used to send the standard language semantics in the clean voice data to the agricultural vehicle control system.

[0054] The beneficial effects of the present invention compared with the prior art are: in the noisy agricultural vehicle use environment, it can achieve pure extraction of voice data, determine the language type of the pure voice data and the semantic conversion of the extracted voice data of different language types, thereby greatly improving the accuracy of on-board voice recognition of agricultural vehicles.

[0055] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in this application document.

[0056] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0058] Figure 1 This is a flow chart of a method for on-board speech recognition for agricultural vehicles in an embodiment of the present invention;

[0059] Figure 2 is a specific method flow chart of step S1 in an embodiment of the present invention;

[0060] Figure 3 Schematic diagram of the internal functional modules of the on-board speech recognition device for agricultural vehicles in an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0062] Embodiment 1:

[0063] The present invention provides a method for agricultural vehicle-mounted speech recognition, referring to Figure 1 ,include:

[0064] S1: Acquire the ambient sound data of agricultural vehicles, finely filter the background noise in the ambient sound data, and obtain finely filtered noise-free ambient sound data;

[0065] S2: Extract pure speech data from the finely filtered noise-free ambient sound data;

[0066] S3: Determine the language type of the clean speech data;

[0067] S4: Based on the preset dialect semantics recognition model of the language type in the clean speech data, the dialect semantics in the clean speech data are recognized;

[0068] S5: performing standard language conversion on the dialect semantics in the clean speech data to obtain the standard language semantics in the clean speech data;

[0069] S6: Send the standard language semantics in the clean speech data to the agricultural vehicle control system.

[0070] In this embodiment, the ambient sound data is sound data of the ambient sound of the agricultural vehicle directly recorded by a recording device installed on the agricultural vehicle.

[0071] In this embodiment, the finely filtered noise-free ambient sound data is sound data obtained by finely filtering the background included in the ambient sound data.

[0072] In this embodiment, the pure voice data is the sound data with better sound quality and better ability to express the voice meaning obtained by performing loss compensation processing on the sound data of the human voice part contained in the finely filtered noise-free ambient sound data.

[0073] In this embodiment, the background noise includes, for example, regular noise generated by a working tractor or occasional (irregular) noise data emitted by animals.

[0074] In this embodiment, the language type is a dialect type, and the division detail and method of the dialect type specifically correspond to the type and division detail of the existing preset dialect semantic recognition model used in this embodiment.

[0075] In this embodiment, the preset dialect semantic recognition model is an existing model that can convert preset types of speech that uses dialect to express semantics into dialect text, or a model trained using a large amount of self-collected dialect speech and corresponding dialect semantics as training samples.

[0076] In this embodiment, the dialect semantics is text content that expresses the speech that uses the dialect to express the semantics in text.

[0077] In this embodiment, standard language conversion is the process of converting dialect semantics into standard language (Mandarin Chinese or English or a language that is convenient for the agricultural vehicle control system to parse, such as program statements, etc.).

[0078] In this embodiment, the standard language semantics is the conversion of text sentences expressed in dialect into text sentences expressed in standard language.

[0079] In this embodiment, the agricultural vehicle control system is a central control system for controlling the agricultural vehicle.

[0080] In this embodiment, the language type of the clean speech data can be determined by using a language type recognition model in the prior art, or by using a model trained with self-collected training samples containing different speech data and corresponding dialect types.

[0081] In this embodiment, based on the preset dialect semantics recognition model of the language type in the clean voice data, the dialect semantics in the clean voice data are recognized as follows:

[0082] The clean speech data is input into the preset dialect semantic recognition model of the corresponding language type to obtain the corresponding dialect semantics.

[0083] In this embodiment, the dialect semantics in the clean speech data are converted into standard language. The standard language semantics in the clean speech data can be obtained by using the standard language conversion model of the corresponding language type in the prior art, or by using a model obtained by training with self-collected training samples containing dialect semantics corresponding to different language types and corresponding standard language semantics.

[0084] The beneficial effects of the above technology are: in the noisy agricultural vehicle use environment, it can achieve pure extraction of voice data, determine the language type of the pure voice data and the semantic conversion of the extracted voice data of different languages, thereby greatly improving the accuracy of on-board voice recognition of agricultural vehicles.

[0085] Embodiment 2:

[0086] Based on Example 1, S1: Acquire the ambient sound data of the agricultural vehicle, finely filter the background noise in the ambient sound data, and obtain finely filtered noise-free ambient sound data. Figure 2 ,include:

[0087] S101: Based on recording devices arranged at different positions of the agricultural vehicle, recording data of the agricultural vehicle within a preset voice escape period is obtained as ambient sound data of the agricultural vehicle;

[0088] S102: removing regular background noise and irregular background noise in the ambient sound data to obtain finely filtered noise-free ambient sound data.

[0089] In this embodiment, the preset speech escape period is the interval period during which the vehicle-mounted speech recognition method or device recognizes the speech near the agricultural vehicle.

[0090] In this embodiment, the regular background noise is noise with regular sound characteristics (for example, the interval periods of the same decibel or the same short-time energy are approximately equal).

[0091] In this embodiment, the irregular background noise is noise whose sound features are not regular.

[0092] The beneficial effect of the above technology is: it is clarified that the process of fine filtering of ambient sound data is the process of removing regular background noise and irregular background noise therein.

[0093] Embodiment 3:

[0094] On the basis of Example 1, S102: removing regular background noise and irregular background noise in the ambient sound data to obtain finely filtered noise-free ambient sound data, including:

[0095] Convert the audio signal corresponding to the ambient sound data from the time domain to the frequency domain to obtain the spectrum data of the ambient sound data;

[0096] Based on the bandpass filter with preset cutoff frequency, the regular background noise in the spectrum data is removed to obtain the coarse filtered noise-free ambient sound data;

[0097] The irregular background noise in the coarsely filtered noise-free ambient sound data is removed to obtain the finely filtered noise-free ambient sound data.

[0098] In this embodiment, the preset cutoff frequency is set in such a way that only the frequency components of the human voice signal are allowed to pass through, thereby achieving the purpose of removing regular noise in the ambient sound data.

[0099] The beneficial effect of the above technology is: it is clear that a bandpass filter with a preset cutoff frequency is used to remove regular noise data in the ambient sound data.

[0100] Embodiment 4:

[0101] On the basis of Example 3, the irregular background noise in the coarsely filtered noise-free ambient sound data is removed to obtain the finely filtered noise-free ambient sound data, including:

[0102] Inputting the coarsely filtered noise-free ambient sound data into each irregular noise recognition model respectively, to obtain each irregular noise data contained in the coarsely filtered noise-free ambient sound data;

[0103] Each type of irregular noise data contained in the coarse-filtered noise-free ambient sound data is eliminated to obtain the fine-filtered noise-free ambient sound data.

[0104] In this embodiment, the irregular noise recognition model is a model that is pre-trained using sound data containing a single type of labeled irregular noise (such as dog barking, agricultural vehicle engine starting sound) as training samples. The irregular noise recognition model can recognize sound data of the corresponding type of irregular noise contained in the sound data.

[0105] In this embodiment, the irregular noise data is sound data of irregular noise contained in the environmental sound data.

[0106] The beneficial effects of the above technology are: it clearly defines the use of each irregular noise recognition model to realize the recognition and elimination of irregular noise data in the ambient sound data.

[0107] Embodiment 5:

[0108] On the basis of Example 1, S2: extracting pure voice data from the finely filtered noise-free ambient sound data, including:

[0109] Extracting sub-fine-filtered noise-free ambient sound data obtained by each recording device from the fine-filtered noise-free ambient sound data;

[0110] Determine the characteristic parameter limit value of each short-term analysis feature of each sub-fine filtered noise-free ambient sound data;

[0111] Among all the sub-fine-filtered noise-free ambient sound data, the sub-fine-filtered noise-free ambient sound data that does not have a characteristic parameter limit value exceeding the characteristic parameter threshold value of the corresponding short-time analysis feature is used as the positioning reference ambient sound data;

[0112] Extract all reference voice data of each human voice data segment from all positioning reference ambient sound data;

[0113] Based on all the reference speech data of all the human voice data segments, pure speech data is obtained.

[0114] In this embodiment, the sub-fine filtered noise-free ambient sound data is sound data obtained after fine filtering of the ambient sound data acquired by each recording device.

[0115] In this embodiment, the short-time analysis features include: short-time energy and short-time average amplitude.

[0116] In this embodiment, the characteristic parameter limit values ​​are the maximum short-time energy and the maximum short-time average amplitude.

[0117] In this embodiment, the characteristic parameter thresholds include an allowable maximum short-time energy value (ie, a short-time energy threshold) and an allowable maximum short-time average amplitude value (ie, a short-time average amplitude threshold).

[0118] In this embodiment, the human voice data segment is a general term for all voice data segments (ie, reference voice data) with similar short-term analysis features contained in all positioning reference ambient sound data.

[0119] The beneficial effects of the above technology are: based on the characteristic parameter limit value of each short-time analysis feature, the ambient sound data obtained by each recording device after fine filtering is screened, all available finely filtered recording data of the same human voice are screened out, and based on all available finely filtered recording data of all human voices, high-purity voice data is obtained.

[0120] Embodiment 6:

[0121] On the basis of Example 5, pure voice data is obtained based on all reference voice data of all human voice data segments, including:

[0122] Acquire speech waveforms of all reference speech data of each vocal speech data segment as all reference speech waveforms of each vocal speech data segment;

[0123] Matching the bands in different reference speech waveforms of each vocal speech data segment to obtain a plurality of bands belonging to different reference speech waveforms and matching each other;

[0124] Based on the time difference between each two bands of different reference speech waveforms belonging to a single human voice data segment and matching each other, determining the time difference between the corresponding two reference speech waveforms in the corresponding two bands matching each other;

[0125] The average of the time difference between each two reference speech waveforms of a single human voice data segment on all the mutually matching bands included is regarded as the time difference between the corresponding two reference speech waveforms of the single human voice data segment;

[0126] Determine the sound source range of the corresponding human voice data segment based on the time difference between each two reference voice waveforms of a single human voice data segment;

[0127] Based on the sound source range of each vocal speech data segment, all reference speech data of each vocal speech data segment are processed for loss compensation and then merged to obtain sub-pure speech data of each vocal speech data segment;

[0128] The sub-clean voice data of all the vocal speech data segments are spliced ​​in time sequence to obtain the pure voice data.

[0129] In this embodiment, the speech waveform diagram is a waveform representation diagram of the sound signal of the human voice data segment.

[0130] In this embodiment, the bands in different reference speech waveforms of each vocal speech data segment are matched, for example:

[0131] Determine the peaks and troughs in two reference speech waveforms of a single human voice data segment, divide the two reference speech waveforms into a plurality of bands based on the peaks and troughs, regard the average of the ratios of all waveform amplitudes at the same time in the two bands belonging to the two reference speech waveforms as the amplitude ratio of the two bands, and regard the band whose amplitude ratio is closest to 1 with respect to the single band as the matching band;

[0132] Based on the above principle, the matching of the bands in different reference speech waveforms of each vocal speech data segment is achieved.

[0133] In this embodiment, the difference in timing between each two bands of different reference speech waveforms belonging to a single human voice data segment and matching each other is:

[0134] The average of the difference between the start time and the end time of each two bands that match each other and belong to different reference speech waveforms of a single human voice data segment is calculated as the difference in timing between the corresponding two bands.

[0135] In this embodiment, based on the difference in timing between each two bands of different reference speech waveforms belonging to a single human voice data segment and matching each other, the time difference between the corresponding two reference speech waveforms in the corresponding two bands matching each other is determined as:

[0136] The time difference between each two bands of different reference speech waveforms belonging to a single human voice data segment and matching each other is regarded as the time difference between the corresponding two reference speech waveforms in the corresponding two bands matching each other.

[0137] In this embodiment, the sound source range is the possible spatial range of the sound source position of the corresponding human voice, which is represented by three-dimensional spatial coordinates.

[0138] In this embodiment, the loss compensation process is a process of amplifying the amplitude of the speech waveform corresponding to the reference speech data by a coefficient.

[0139] In this embodiment, the amplitudes of the speech waveform graphs of all reference speech data after loss compensation of a single human voice data segment at the same time are averaged and taken as the amplitude of the merged speech waveform graph at the corresponding time, thereby obtaining the merged speech waveform graph, and the sound data corresponding to the merged speech waveform graph is taken as the merged speech data.

[0140] In this embodiment, the sub-clean voice data is voice data obtained by compensating and merging all reference voice data of a single human voice data segment.

[0141] The beneficial effect of the above technology is: by matching the bands in different reference speech waveforms of each vocal speech data segment and obtaining the time difference, the time difference of different recording devices obtaining the same vocal speech is determined, thereby further inferring the possible spatial range of the sound source position, and achieving a rough estimation of the sound source position of each vocal speech data segment.

[0142] Embodiment 7:

[0143] On the basis of Example 6, based on the time difference between every two reference speech waveforms of a single human voice data segment, determining the sound source range of the corresponding human voice data segment includes:

[0144] The difference in distance between the two corresponding recording devices and the sound source position is calculated based on the preset sound propagation speed and the time difference between each two reference speech waveforms of a single human voice data segment;

[0145] The sound source range of the human voice data segment is determined based on the position coordinates of all recording devices in a preset coordinate system and the difference between the distances between every two recording devices of a single human voice data segment and the sound source position.

[0146] In this embodiment, the preset sound propagation speed is the preset propagation speed of human voice in the agricultural vehicle use environment, for example, the propagation speed in the air is 340 meters per second.

[0147] In this embodiment, the difference in distance between the corresponding two recording devices and the sound source position is calculated based on the preset sound propagation speed and the time difference between each two reference speech waveforms of a single human voice data segment, which is:

[0148] The product of the preset sound propagation speed and the time difference between each two reference speech waveforms of a single human voice data segment is regarded as the difference in distance between the corresponding two recording devices and the sound source position.

[0149] In this embodiment, the sound source position is the sound source position of the human voice that emits the corresponding human voice speech data segment in the three-dimensional space.

[0150] In this embodiment, based on the position coordinates of all recording devices in the preset coordinate system and the difference between the distances between each two recording devices of a single human voice data segment and the sound source position, the sound source range of the human voice data segment is determined as follows:

[0151] Step 1: Taking the unknown distance between a single recording device and the sound source as the unknown variable, construct a spherical representation function with the recording device as the center of the sphere;

[0152] Step 2: construct a spherical coordinate function with other recording devices as the center of the sphere based on the difference between the distance between other recording devices and the recording device in step 1 and the sound source position and the position coordinates of the recording devices in a preset (three-dimensional) coordinate system, and the spherical coordinate function with other recording devices as the center of the sphere also only includes four variables: the three-dimensional coordinate quantity and the unknown distance between the single recording device in step 1 and the sound source position, wherein the position coordinates of the recording device in the preset coordinate system are the coordinate values ​​of the physical center of the recording device in three-dimensional space;

[0153] Step 3: Using all currently constructed spherical representation functions as constraints, determine a spatial range representation function that satisfies all currently constructed spherical representation functions at the same time. The spatial range corresponding to the spatial range representation function is the sound source range of the human voice data segment.

[0154] The beneficial effect of the above technology is: based on the preset sound propagation speed and the time difference between each two reference speech waveforms of a single human voice data segment, the difference in distance between the corresponding two recording devices and the sound source position is calculated, so as to realize the estimation of the possible spatial range of the sound source position of each human voice data segment.

[0155] Embodiment 8:

[0156] On the basis of Example 6, based on the sound source range of each vocal speech data segment, all reference speech data of each vocal speech data segment are combined after being compensated, so as to obtain sub-clean speech data of each vocal speech data segment, including:

[0157] Based on the coordinate representation of the sound source range of each vocal speech data segment in a preset coordinate system and the position coordinates of all recording devices in the preset coordinate system, determine all possible sound conduction paths between the sound source position of each vocal speech data segment and each recording device;

[0158] Based on all possible sound conduction paths from the sound source position of each vocal speech data segment to each recording device, all reference speech data of each vocal speech data segment are processed for loss compensation and then merged to obtain sub-pure speech data of each vocal speech data segment.

[0159] In this embodiment, the coordinate representation of the sound source range in the preset coordinate system is the spatial range representation function corresponding to the sound source range obtained in Embodiment 7.

[0160] In this embodiment, all possible sound conduction paths between the sound source position of each human voice data segment and each recording device are:

[0161] The path between the position coordinates of each recording device in the preset coordinate system and a single position point within the sound source range is a single possible sound conduction path, and the number of all possible sound conduction paths corresponds to the number of position points within the sound source range.

[0162] The beneficial effect of the above technology is: by determining all possible sound conduction paths between the sound source position of each vocal speech data segment and each recording device, it is possible to compensate and merge all reference speech data of each vocal speech data segment, thereby obtaining speech data with better sound quality and better ability to represent its speech semantics.

[0163] Embodiment 9:

[0164] On the basis of Example 8, based on all possible sound conduction paths from the sound source position of each vocal speech data segment to each recording device, all reference speech data of each vocal speech data segment are subjected to loss compensation processing and then merged to obtain sub-clean speech data of each vocal speech data segment, including:

[0165] Determine the average value of the acoustic loss coefficients of all acoustic conduction paths between the sound source position of each vocal speech data segment and each recording device, and use it as the total acoustic loss coefficient of the corresponding reference speech data;

[0166] Based on the total loss coefficient of all reference voice data of each vocal speech data segment, the corresponding reference voice data is subjected to loss compensation processing to obtain all lossless voice data of each vocal speech data segment;

[0167] All lossless voice data of each vocal speech data segment are merged to obtain sub-clean voice data of each vocal speech data segment.

[0168] In this embodiment, the acoustic loss coefficients of all acoustic conduction paths between the sound source position of each vocal speech data segment and each recording device are determined by retrieving a preset acoustic conduction path representation function-acoustic loss coefficient list, wherein the acoustic loss coefficient is the degree of amplitude loss of the sound signal.

[0169] In this embodiment, the total sound loss coefficient is to determine the amplitude loss degree of the sound signal corresponding to each human voice data segment by using all reference voice data of each human voice data segment.

[0170] In this embodiment, based on the total loss coefficient of all reference speech data of each vocal speech data segment, the corresponding reference speech data is subjected to loss compensation processing, that is,

[0171] The value obtained by dividing the amplitude of the sound signal of each reference speech data at each moment by (1-acoustic loss coefficient) is used as the new sound signal amplitude at the corresponding moment;

[0172] A new sound signal waveform is fitted using the new sound signal amplitude at all times, and the lossless speech data after the compensation processing is obtained using the new sound signal waveform.

[0173] In this embodiment, the lossless voice data is voice data obtained after performing loss compensation processing on the reference voice data.

[0174] In this embodiment, all lossless speech data of each vocal speech data segment are merged into:

[0175] The average of all signal amplitudes at each moment in the sound signal waveform diagram of all lossless voice data of each vocal speech data segment is taken as the new signal amplitude at the corresponding moment, and the new signal amplitude at all moments is used to fit the sub-pure voice data of the corresponding vocal speech data segment.

[0176] The beneficial effects of the above technology are: by averaging the sound loss coefficients of all sound conduction paths between the sound source position of each vocal speech data segment to each recording device, a total sound loss coefficient is obtained, and the total sound loss coefficient is used to compensate and merge the sound loss coefficients of all sound conduction paths between the sound source position of each vocal speech data segment to each recording device, thereby obtaining speech data with better sound quality and better ability to represent its speech semantics.

[0177] Embodiment 10:

[0178] The present invention provides an agricultural vehicle vehicle-mounted speech recognition device for executing the agricultural vehicle vehicle-mounted speech recognition method described in any one of embodiments 1 to 9, referring to Figure 3 ,include:

[0179] The ambient sound acquisition module is used to acquire the ambient sound data of agricultural vehicles, finely filter the background noise in the ambient sound data, and obtain finely filtered noise-free ambient sound data;

[0180] A voice data extraction module is used to extract pure voice data from the finely filtered noise-free ambient sound data;

[0181] A language type determination module, used to determine the language type of the clean voice data;

[0182] A dialect semantics recognition module, used to recognize the dialect semantics in the clean voice data based on a preset dialect semantics recognition model of the language type in the clean voice data;

[0183] A standard language conversion module is used to convert the dialect semantics in the clean voice data into standard language semantics to obtain the standard language semantics in the clean voice data;

[0184] The recognition result sending module is used to send the standard language semantics in the clean voice data to the agricultural vehicle control system.

[0185] The beneficial effects of the above technology are: in the noisy agricultural vehicle use environment, it can achieve pure extraction of voice data, determine the language type of the pure voice data and the semantic conversion of the extracted voice data of different languages, thereby greatly improving the accuracy of on-board voice recognition of agricultural vehicles.

[0186] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for voice recognition on an agricultural vehicle, characterized in that: include: S1: Acquire the ambient sound data of agricultural vehicles, finely filter the background noise in the ambient sound data, and obtain finely filtered noise-free ambient sound data; S2: Extract pure speech data from the finely filtered noise-free ambient sound data; S3: Determine the language type of the clean speech data; S4: Based on the preset dialect semantics recognition model of the language type in the clean speech data, the dialect semantics in the clean speech data are recognized; S5: performing standard language conversion on the dialect semantics in the clean speech data to obtain the standard language semantics in the clean speech data; S6: sending the standard language semantics in the clean voice data to the agricultural vehicle control system; Among them, S2: extracting pure voice data from the finely filtered noise-free ambient sound data, including: Extracting sub-fine-filtered noise-free ambient sound data obtained by each recording device from the fine-filtered noise-free ambient sound data; Determine the characteristic parameter limit value of each short-time analysis feature of each sub-fine filtered noise-free ambient sound data, wherein the characteristic parameter limit value is the maximum short-time energy and the maximum short-time average amplitude; Among all the sub-fine-filtered noise-free ambient sound data, the sub-fine-filtered noise-free ambient sound data that does not have a characteristic parameter limit value exceeding the characteristic parameter threshold value of the corresponding short-time analysis feature is used as the positioning reference ambient sound data; Extract all reference voice data of each human voice data segment from all positioning reference ambient sound data; Based on all reference voice data of all human voice data segments, pure voice data is obtained; Wherein, based on all reference voice data of all human voice data segments, pure voice data is obtained, including: Acquire speech waveforms of all reference speech data of each vocal speech data segment as all reference speech waveforms of each vocal speech data segment; Matching the bands in different reference speech waveforms of each vocal speech data segment to obtain a plurality of bands belonging to different reference speech waveforms and matching each other; Based on the time difference between each two bands of different reference speech waveforms belonging to a single human voice data segment and matching each other, determining the time difference between the corresponding two reference speech waveforms in the corresponding two bands matching each other; The average of the time difference between each two reference speech waveforms of a single human voice data segment on all the mutually matching bands included is regarded as the time difference between the corresponding two reference speech waveforms of the single human voice data segment; Based on the time difference between each two reference speech waveforms of a single human voice data segment, determining the sound source range of the corresponding human voice data segment; Based on the sound source range of each vocal speech data segment, all reference speech data of each vocal speech data segment are processed for loss compensation and then merged to obtain sub-pure speech data of each vocal speech data segment; splicing the sub-clean voice data of all the vocal voice data segments in time sequence to obtain the pure voice data; Wherein, based on the sound source range of each vocal speech data segment, all reference speech data of each vocal speech data segment are combined after being compensated, so as to obtain sub-clean speech data of each vocal speech data segment, including: Based on the coordinate representation of the sound source range of each vocal speech data segment in a preset coordinate system and the position coordinates of all recording devices in the preset coordinate system, determine all possible sound conduction paths between the sound source position of each vocal speech data segment and each recording device; Based on all possible sound conduction paths from the sound source position of each vocal speech data segment to each recording device, all reference speech data of each vocal speech data segment are processed for loss compensation and then merged to obtain sub-pure speech data of each vocal speech data segment; Wherein, based on the time difference between every two reference speech waveforms of a single human voice data segment, determining the sound source range of the corresponding human voice data segment includes: The difference in distance between the two corresponding recording devices and the sound source position is calculated based on the preset sound propagation speed and the time difference between each two reference speech waveforms of a single human voice data segment; The sound source range of the human voice data segment is determined based on the position coordinates of all recording devices in a preset coordinate system and the difference between the distances between every two recording devices of a single human voice data segment and the sound source position.

2. The agricultural vehicle-mounted speech recognition method according to claim 1, characterized in that: S1: Obtain the ambient sound data of agricultural vehicles, finely filter the background noise in the ambient sound data, and obtain finely filtered noise-free ambient sound data, including: S101: Based on recording devices arranged at different positions of the agricultural vehicle, recording data of the agricultural vehicle within a preset voice escape period is obtained as ambient sound data of the agricultural vehicle; S102: removing regular background noise and irregular background noise in the ambient sound data to obtain finely filtered noise-free ambient sound data.

3. The agricultural vehicle-mounted speech recognition method according to claim 1, characterized in that: S102: removing regular background noise and irregular background noise in the ambient sound data to obtain finely filtered noise-free ambient sound data, including: Convert the audio signal corresponding to the ambient sound data from the time domain to the frequency domain to obtain the spectrum data of the ambient sound data; Based on the bandpass filter with preset cutoff frequency, the regular background noise in the spectrum data is removed to obtain the coarse filtered noise-free ambient sound data; The irregular background noise in the coarsely filtered noise-free ambient sound data is removed to obtain the finely filtered noise-free ambient sound data.

4. The agricultural vehicle-mounted speech recognition method according to claim 3, characterized in that: The irregular background noise in the coarse-filtered noise-free ambient sound data is removed to obtain the fine-filtered noise-free ambient sound data, including: Inputting the coarsely filtered noise-free ambient sound data into each irregular noise recognition model respectively, to obtain each irregular noise data contained in the coarsely filtered noise-free ambient sound data; Each type of irregular noise data contained in the coarse-filtered noise-free ambient sound data is eliminated to obtain the fine-filtered noise-free ambient sound data.

5. The agricultural vehicle-mounted speech recognition method according to claim 1, characterized in that: Based on all possible sound conduction paths from the sound source position of each vocal speech data segment to each recording device, all reference speech data of each vocal speech data segment are processed for loss compensation and then merged to obtain sub-clean speech data of each vocal speech data segment, including: Determine the average value of the acoustic loss coefficients of all acoustic conduction paths between the sound source position of each vocal speech data segment and each recording device, and use it as the total acoustic loss coefficient of the corresponding reference speech data; Based on the total loss coefficient of all reference voice data of each vocal speech data segment, the corresponding reference voice data is subjected to loss compensation processing to obtain all lossless voice data of each vocal speech data segment; All lossless voice data of each vocal speech data segment are merged to obtain sub-clean voice data of each vocal speech data segment.

6. A speech recognition device mounted on an agricultural vehicle, characterized in that: The method for performing the agricultural vehicle-mounted speech recognition method according to any one of claims 1 to 5 comprises: The ambient sound acquisition module is used to acquire the ambient sound data of agricultural vehicles, finely filter the background noise in the ambient sound data, and obtain finely filtered noise-free ambient sound data; A voice data extraction module is used to extract pure voice data from the finely filtered noise-free ambient sound data; A language type determination module, used to determine the language type of the clean voice data; A dialect semantics recognition module, used to recognize the dialect semantics in the clean voice data based on a preset dialect semantics recognition model of the language type in the clean voice data; A standard language conversion module is used to convert the dialect semantics in the clean voice data into standard language semantics to obtain the standard language semantics in the clean voice data; The recognition result sending module is used to send the standard language semantics in the clean voice data to the agricultural vehicle control system.

Images