Speech recognition method for multi-dimensional AI platform

Abstract

The present invention relates to the field of speech recognition, and in particular to a speech recognition method for a multi-dimensional AI platform, and the technical scheme adopted is as follows: obtaining speech data sent from at least two audio receiving devices; correcting the corresponding original filter group according to the speech response degree of the speech data at different frequencies, and obtaining the corrected filter group; using the corrected filter group to generate a Mel spectrogram of the speech data; inputting the Mel spectrograms into a preset speech recognition network, and obtaining the speech recognition results corresponding to each of the speech data. The present invention can solve the problem that the speech recognition effect of the vehicle-mounted speech recognition system is poor in a multi-user environment.

Description

Technical Field

[0001] The present invention relates to the field of speech recognition technology, and in particular to a speech recognition method for a multi-dimensional AI platform. Background Art

[0002] With the rapid development of artificial intelligence (AI) and speech recognition technology, voice interaction has become an important function in intelligent driving and in-vehicle systems. Traditional in-vehicle speech recognition systems usually rely on a single microphone or a simple microphone array to capture the driver's voice commands. These in-vehicle speech recognition systems perform well in quiet environments, but the speech recognition effect is often not accurate and reliable in complex in-vehicle environments (such as background noise, multiple passenger conversations, etc.). When processing multiple smart cockpit audio data, the existing technology often needs to convert the obtained audio data into mel spectrograms in order to better perform speech recognition. However, the traditional mel spectrogram conversion method uses the same frequency conversion strategy for all input voices and lacks the ability to adaptively process the voice features of different users. This will result in the performance of the mel spectrogram in the multi-user environment obtained by the in-vehicle multi-dimensional AI platform, that is, the situation of voice aliasing of multiple different users in the vehicle. It is not ideal, and it is difficult to effectively distinguish and process multiple voice data, which reduces the accuracy of speech recognition. Summary of the invention

[0003] In order to solve the problem that the in-vehicle speech recognition system has poor speech recognition effect in a multi-user environment, the purpose of the present invention is to provide a speech recognition method for a multi-dimensional AI platform. The technical solution adopted is as follows:

[0004] Acquire voice data sent from at least two audio receiving devices;

[0005] Modifying the corresponding original filter group according to the speech response degree of the speech data at different frequencies to obtain a modified filter group;

[0006] Using the modified filter bank to generate a Mel-spectrogram of the speech data;

[0007] The Mel-spectrograms are input into a preset speech recognition network to obtain speech recognition results corresponding to each of the speech data.

[0008] Wherein, the corresponding original filter group is modified according to the speech response degree of the speech data at different frequencies, and the modified filter group is obtained, which includes:

[0009] Performing independent component analysis on each of the speech data to obtain independent components, and converting each of the speech data and each of the independent components into a spectrogram;

[0010] Acquiring the speech response degree of each of the speech data at different frequencies based on the spectrogram;

[0011] The original filter bank is modified accordingly based on the speech response level.

[0012] Wherein, obtaining the speech response degree of each of the speech data at different frequencies based on the spectrogram includes:

[0013] Obtaining the voice confusion degree of each of the voice data at each time;

[0014] Dividing the speech data into time periods according to the corresponding speech confusion degree;

[0015] The speech response degree of the speech data at different frequencies is calculated.

[0016] The obtaining of the voice confusion degree of each voice data at each time comprises:

[0017] Acquire energy change data of independent components in each of the speech data;

[0018] Selecting corresponding energy change data to be analyzed according to the energy change data;

[0019] Respectively collecting the frequencies to be analyzed corresponding to the energy change data to be analyzed and sorting the frequencies to be analyzed;

[0020] The speech confusion degree is calculated based on the sorted frequencies to be analyzed.

[0021] Wherein, dividing the voice data into time periods according to the corresponding voice confusion degree includes:

[0022] Obtaining a mean of the first-order differences of the speech confusion degree with respect to a first time, where the first time is within a time range of a preset time neighborhood from a time point to be checked;

[0023] When the first-order difference of the speech confusion degree about the to-be-tested time point is greater than the mean value, the to-be-tested time point is a time segmentation point, and the time segmentation point is used to divide the speech data into the time periods.

[0024] Wherein, calculating the voice response degree of the voice data at different frequencies includes:

[0025] Normalizing the inverse of the mean value of the speech confusion degree in the first time segment to obtain a normalized result, wherein the first time segment is any time segment after the time period is divided;

[0026] Acquire a first segment initial voice response degree of the voice data at the frequency to be analyzed within the first time segment;

[0027] Multiplying the first segmented initial voice response degree by the normalization result to obtain a first segmented voice response degree of the voice data at the frequency to be analyzed;

[0028] The above steps are repeated to sequentially traverse each of the time segments and change the frequency value to be analyzed, so as to obtain the voice response degree of the voice data at different frequencies in each time segment.

[0029] Wherein, obtaining the first segment initial voice response degree of the voice data at the frequency to be analyzed in the first time segment includes:

[0030] In the first time segment, the energy of the independent components in the speech data at different frequencies is obtained, and the maximum value of the energy ratio at the frequency to be analyzed is calculated, where the maximum value is the probability that the frequency to be analyzed is the frequency of the speech data;

[0031] The first segment initial speech response degree of the speech data at the frequency to be analyzed is obtained based on the energy of the speech data in the first time segment according to the probability.

[0032] The method further includes: the process of obtaining the energy ratio includes:

[0033] Acquire a minimum value of energy of the independent component of the speech data at a first frequency, where the first frequency is a frequency corresponding to a minimum energy of the independent component of the speech data;

[0034] Add a preset positive number to the minimum value to obtain a preset denominator;

[0035] The energy ratio is obtained by dividing the energy value of the independent component of the speech data at the frequency to be analyzed by the preset denominator.

[0036] The method further includes: obtaining, according to the probability, a first segment initial voice response degree of the voice data at the frequency to be analyzed in the first time segment, including:

[0037] Obtaining the sum of the products of the energy value corresponding to the frequency to be analyzed and the probability corresponding to the frequency to be analyzed at different times within the first time segment;

[0038] The sum of the products is divided by the duration of the first time segment to obtain the first segment initial voice response degree.

[0039] Wherein, modifying the original filter bank accordingly based on the speech response degree comprises:

[0040] The center frequency of the original filter to be corrected is the initial center frequency, and the previous initial center frequency of the previous filter before correction and the previous corrected center frequency after correction of the previous filter of the original filter to be corrected are obtained;

[0041] Obtaining a mean value of the speech response degree in a frequency interval from the previous initial center frequency to the initial center frequency, taking the inverse of the mean value of the speech response degree and normalizing it to obtain a normalized result;

[0042] Subtracting the previous initial center frequency from the initial center frequency to obtain a difference result;

[0043] After the normalization result is multiplied by the difference result, the result is added to the previous corrected center frequency to obtain the corrected center frequency of the original filter to be corrected.

[0044] The present invention has the following beneficial effects:

[0045] The present invention uses a multi-dimensional AI platform to identify the voice data in the car. First, the audio receiving device built in the smart cockpit is used to collect the voice data emitted by the people in the cockpit. The number of audio receiving devices is at least two. These voice data often include the voice data of each person in the vehicle, and also include various noises such as vehicle operation noise. In order to make the voice data of different people in the smart cabin more accurately separated and enhanced, the corresponding original filter group is corrected according to the voice response degree of the voice data at different frequencies, and the center frequency of the original filter group is corrected. The mel spectrogram generated by the corrected mel filter group can better characterize the voice characteristics of the corresponding people in the cockpit. In the process of mel spectrum conversion, the frequency component of the voice part of the main character is enhanced, so that the mel spectrum can adaptively adjust the voice part component of its concern. In this way, the voice characteristics of the main character can be captured more effectively, the sensitivity and recognition effect of the voice recognition system to the voice of the main character can be improved, and the robustness of the vehicle-mounted voice recognition system in a complex in-vehicle environment has also been greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 A flowchart of a speech recognition method for a multi-dimensional AI platform provided by one embodiment of the present invention;

[0048] Figure 2 A schematic diagram of an audio receiving device for a smart cockpit provided by an embodiment of the present invention;

[0049] Figure 3 A flowchart of step S102 provided in one embodiment of the present invention;

[0050] Figure 4 A flowchart of step S302 provided in one embodiment of the present invention;

[0051] Figure 5 A flowchart of step S401 provided in one embodiment of the present invention;

[0052] Figure 6 A flowchart of step S403 provided in one embodiment of the present invention;

[0053] Figure 7 The effect diagram of the original filter bank provided by one embodiment of the present invention;

[0054] Figure 8 This is a diagram showing the effect of a modified filter bank provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the speech recognition method for a multi-dimensional AI platform proposed by the present invention, its specific implementation method, structure, features and effects are described in detail below in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0056] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0057] The specific scheme of the speech recognition method for a multi-dimensional AI platform provided by the present invention is described in detail below with reference to the accompanying drawings.

[0058] See also Figure 1 , which shows a flow chart of a speech recognition method for a multi-dimensional AI platform provided by an embodiment of the present invention, the method comprising:

[0059] S101. Obtaining voice data sent from at least two audio receiving devices;

[0060] The vehicle-mounted smart cockpit system is equipped with multiple multi-channel audio receiving devices to obtain the voice data corresponding to each cockpit at the same time. The schematic diagram of the audio receiving device of the smart cockpit is shown in the figure. Figure 2 As shown, the number of the multi-channel audio receiving devices is the same as the number of the cockpits, and one multi-channel audio receiving device corresponds to obtaining the voice data of one cockpit. The multi-channel audio receiving device can be installed in a location suitable for receiving voice data, such as the side below the seat or behind the backrest of the previous cockpit, etc. The smart cockpit can be equipped with an audio acquisition module, a signal processing module, a voice recognition module, a user interaction module, etc. The audio acquisition module or the audio receiving device can collect voice data.

[0061] After receiving the voice data sent by the audio receiving devices at different locations, record the The voice data of an audio receiving device is , the frequency of voice data is 50kHz, that is, there are 50k data points in 1s. The number of audio receiving devices installed in the same vehicle is .

[0062] S102. Modifying the corresponding original filter group according to the speech response degree of the speech data at different frequencies to obtain a modified filter group;

[0063] The original filter group, such as the mel filter group, does not consider the difference between the main frequencies of the voices of different people separately, which will lead to inaccurate voice feature extraction for certain specific individuals. In order to solve the above problem, the present invention creatively uses the voice response degree of the voice data at different frequencies to correct the center frequency of the original filter group, so that the corrected filter group can extract voice features for specific individuals more accurately.

[0064] S103. Using the modified filter bank to generate a Mel spectrogram of the speech data;

[0065] The modified filter group converts the spectrogram of the voice data of the audio receiving device into a mel spectrogram, i.e., a mel spectrogram. The energy information in the mel spectrogram can better characterize the voice characteristics of the person speaking in the cockpit.

[0066] S104. Input the Mel-spectrograms into a preset speech recognition network to obtain speech recognition results corresponding to each of the speech data.

[0067] The mel spectrogram is used as the input layer of the speech recognition network, with a size of time step × number of frequency channels. The model output layer is the probability distribution of characters or words at each time step. A hybrid neural network architecture is constructed by connecting a convolutional layer (CNN) to a long short-term memory network (LSTM). The feature extraction capability of the convolutional network and the temporal modeling capability of the recursive network are combined to perform speech recognition, and the CTC loss function is used to train the network.

[0068] For subsequent voice input, voice recognition is performed based on the obtained hybrid network model. Based on the voice recognition results, it is fed back to the multi-dimensional AI platform, which can provide more accurate voice recognition and natural language understanding, and provide a more smooth and natural human-computer interaction experience.

[0069] The speech recognition network is a prior art and will not be described in detail here.

[0070] Figure 3 The flowchart of step S102 provided in one embodiment of the present invention, wherein the corresponding original filter group is modified according to the speech response degree of the speech data at different frequencies, and the modified filter group is obtained, comprising:

[0071] S301. Performing independent component analysis on each of the speech data to obtain independent components, and converting each of the speech data and each of the independent components into a spectrogram;

[0072] The voice data received by multiple audio receiving devices are subjected to ICA independent component analysis. ICA is a blind source separation technology, which usually assumes that the number of input mixed signals is the same as the number of output independent signals. The first one of the voice data corresponding to the smart cockpit The independent components are , where the number of independent components is , that is, Among them, each component corresponds to a different sound source signal. In order to analyze the frequency and other conditions of the sound source of each voice data, each voice data and its independent components are converted into a spectrogram, and subsequent analysis is performed based on the obtained spectrogram information.

[0073] S302. Acquire the speech response degree of each of the speech data at different frequencies based on the spectrogram;

[0074] The process of obtaining the speech response degree of each speech data at different frequencies based on the spectrogram will be described in detail in the subsequent embodiments and will not be repeated here.

[0075] S303. Modify the original filter group accordingly based on the speech response degree.

[0076] Optionally, modifying the original filter bank accordingly based on the speech response degree comprises:

[0077] The center frequency of the original filter to be corrected is the initial center frequency, and the previous initial center frequency of the previous filter before correction and the previous corrected center frequency after correction of the previous filter of the original filter to be corrected are obtained;

[0078] Obtaining a mean value of the speech response degree in a frequency interval from the previous initial center frequency to the initial center frequency, taking the inverse of the mean value of the speech response degree and normalizing it to obtain a normalized result;

[0079] Subtracting the previous initial center frequency from the initial center frequency to obtain a difference result;

[0080] After the normalization result is multiplied by the difference result, the result is added to the previous corrected center frequency to obtain the corrected center frequency of the original filter to be corrected.

[0081] For obtaining the corrected center frequency of the first original filter, the initial center frequency of the 0th original filter is 0, and the corrected center frequency of the 0th original filter is of course also 0. The average voice response degree of the first original filter is the audio data obtained by the first smart cockpit in the frequency range The mean value of the speech response degree within the range is obtained by taking the inverse of the mean value of the speech response degree and normalizing it to obtain a normalized result. According to the step of multiplying the normalized result by the difference result and then adding it to the previous corrected center frequency to obtain the corrected center frequency of the original filter to be corrected, the corrected center frequency of the first original filter can be obtained as the normalized result multiplied by .

[0082] After the corrected center frequency of the first original filter is obtained, the corrected center frequency of the second original filter, the corrected center frequency of the third original filter, and so on can be obtained in a similar manner. In one embodiment, the number of filters in the original filter group can be , among which, The initial center frequency of the filter is , while arrive There are multiple voice response levels at different frequencies. For the audio data obtained by a smart cockpit, record the audio data in the interval The average level of voice response in For the intervals with higher mean values ​​of speech response, the corresponding interval length should be shortened to better preserve the original spectrum information after filtering.

[0083] After correction The center frequency of the filter for:

[0084]

[0085] in, Indicates the revised The center frequency of the filter, represents the normalization function, Indicates The audio data obtained by the smart cockpit is in the frequency range The mean level of voice response within Indicates The initial center frequency of the filter, Indicates The initial center frequency of the filter.

[0086] The correction of the center frequency of the original filter is performed one by one. Each original filter to be corrected needs the center frequency of the previous corrected original filter as a basis, and the center frequency of the original filter to be corrected is determined according to the deviation between the original filter to be corrected and the previous original filter. The deviation is precisely the innovation of the present invention. The center frequency of the original filter to be corrected can be obtained by obtaining the center frequency of the original filter after the previous correction and adding the deviation between the original filter to be corrected and the original filter after the previous correction.

[0087] Figure 7 This is a rendering of an original filter bank (original Mel filter bank) provided by an embodiment of the present invention. Figure 8 For an effect diagram of a modified filter bank (improved Mel filter bank) provided by an embodiment of the present invention, please refer to Figure 7 and Figure 8 ,in, Figure 7 and Figure 8 In the figure, the horizontal axis represents frequency (Hz) and the vertical axis represents amplitude. Compared with the original filter group, the improved filter group has a lower degree of preservation at around 50 frequencies, while the filter group after 60 has a better effect of preserving the original spectrum, indicating that the speech frequency of the people in the cockpit is mainly concentrated in within the frequency range.

[0088] Figure 4 The flowchart of step S302 provided in one embodiment of the present invention, obtaining the speech response degree of each speech data at different frequencies based on the spectrogram includes:

[0089] S401. Obtaining the degree of speech confusion of each of the speech data at each time;

[0090] After obtaining the spectrograms of each voice data and its corresponding components, considering that the vehicle is moving, the voice data that can be obtained in each cabin not only includes the voice information of each person in the vehicle, but also includes the noise emitted by the vehicle itself during operation, the noise outside the vehicle, etc. The number of independent components of the voice data is fixed, which will lead to the result of traditional ICA independent component analysis, which cannot better separate the voice data of different people, that is, each independent component contains not only the sound data of a single person, but also a certain amount of uncertain noise.

[0091] The energy changes of the spectrogram are different for different independent components at different frequencies. For human voice, the voice frequency of the same person is relatively stable over a period of time, while the external noise of the vehicle, electrical noise, brake noise, air conditioning noise, etc. during vehicle operation are highly random, that is, at some moments, the noise will appear in multiple different independent components.

[0092] The process of obtaining the voice confusion degree of each voice data at each time will be described in detail in the subsequent embodiments and will not be repeated here.

[0093] S402. Divide the voice data into time periods according to the corresponding degree of voice confusion;

[0094] Wherein, dividing the voice data into time periods according to the corresponding voice confusion degree includes:

[0095] Obtaining a mean of the first-order differences of the speech confusion degree with respect to a first time, where the first time is within a time range of a preset time neighborhood from a time point to be checked;

[0096] When the first-order difference of the speech confusion degree about the to-be-tested time point is greater than the mean value, the to-be-tested time point is a time segmentation point, and the time segmentation point is used to divide the speech data into the time periods.

[0097] Specifically, for the voice data of the same cockpit, the degree of voice confusion at each moment is obtained. ,when Time, remember the time is a time segment point, where express About time The first-order difference of represents the mean function, Indicates the time neighborhood size. The recommended value here is 20, which can be customized.

[0098] According to the above steps, each of the voice data is divided into time periods.

[0099] S403: Calculate the speech response degree of the speech data at different frequencies.

[0100] According to the time division results corresponding to each cockpit, the noise mixing degree of the voice data obtained by a single cockpit is relatively stable within each time segment. In order to obtain the actual audio frequency situation of the personnel corresponding to the current single cockpit, it is considered to analyze the energy of the audio data at different frequencies within each time segment, and calculate the voice response degree of each cockpit at different frequencies.

[0101] The process of calculating the voice response degree of the voice data at different frequencies will be described in detail in the subsequent embodiments and will not be repeated here.

[0102] Figure 5 The flowchart of step S401 provided in one embodiment of the present invention, obtaining the voice confusion degree of each voice data at each time comprises:

[0103] S501. Obtain energy change data of independent components in each of the voice data;

[0104] Record Among the audio data obtained from the cockpit, The energy change data of each independent component is: .in, Indicates the selected frequency upper limit, which can be a preset value. The recommended value here is 1000; Indicates energy change data at time The frequency is The amount of energy below.

[0105] S502. Selecting corresponding energy change data to be analyzed according to the energy change data;

[0106] In order to avoid the influence of the energy change data with smaller energy on the judgment of the degree of speech confusion, it is possible to select The value corresponding to the first quarter dividing point after the energy values ​​are sorted from large to small is used as the threshold, and the first 25% of the energy change data are selected for subsequent analysis.

[0107] Of course, there is not only one way to select energy change data, and it can be adaptively adjusted according to actual conditions.

[0108] S503. Collect the frequencies to be analyzed corresponding to the energy change data to be analyzed respectively and sort the frequencies to be analyzed;

[0109] Optionally, when the first 25% of the energy change data are selected as the energy change data to be analyzed, the frequency corresponding to the selected energy change data, that is, the frequency to be analyzed, is sorted from large to small to form the sequence: .

[0110] S504. Calculate the speech confusion degree based on the sorted frequencies to be analyzed.

[0111] At the moment Next The degree of speech confusion of the speech data obtained from each cockpit for:

[0112]

[0113] in, represents the number of independent components obtained from the speech data, It represents the standard deviation function. Indicates the selected upper frequency limit. Indicates The first channel (smart cockpit) after data decomposition Independent components at time Select the next frequency value.

[0114] When the frequency values ​​are relatively concentrated, it means that the independent components have a better decomposition of the original speech data. At the same time, the speech data contains less noise and the mixed audio data has less speech confusion.

[0115] Figure 6 The flowchart of step S403 provided in one embodiment of the present invention, wherein calculating the speech response degree of the speech data at different frequencies includes:

[0116] S601. Normalizing the inverse of the mean of the speech confusion degree in the first time segment to obtain a normalized result, wherein the first time segment is any time segment after the time period is divided;

[0117] S602. Obtaining a first segment initial voice response level of the voice data at the frequency to be analyzed within the first time segment;

[0118] Wherein, obtaining the first segment initial voice response degree of the voice data at the frequency to be analyzed in the first time segment includes:

[0119] a. Obtaining the energy of the independent components in the speech data at different frequencies, and calculating the maximum value of the energy ratio at the frequency to be analyzed, wherein the maximum value is the probability that the frequency to be analyzed is the frequency of the speech data;

[0120] The process of obtaining the energy ratio includes: obtaining the minimum value of the energy of the independent component of the speech data at the first frequency; adding a preset number to the minimum value to obtain a preset denominator; and dividing the energy value of the independent component of the speech data at the frequency to be analyzed by the preset denominator to obtain the energy ratio.

[0121] Specifically, considering that there is a certain correlation between the voice data obtained by multiple audio receiving devices, when the first When a person in a cockpit produces a voice, multiple similar voice receiving devices will receive audio with different volumes. After the voice data is subjected to ICA independent component analysis, the voice will be better reflected in an independent component of the remaining audio receiving devices, that is, at a certain moment and with the same frequency. When the maximum value of the energy ratio of a certain independent component is large, the frequency is more likely to be the corresponding frequency of speech, then the time is recorded. Next, frequency The probability that the frequency of the voice data is , The calculation method is:

[0122]

[0123] in, Indicates The audio receiving device The spectrogram corresponding to the component, It means at time The frequency is The energy value under the condition, w is set to prevent the denominator from being 0, w can be 0.01, etc., max is the maximum value function, min is the minimum value function, It refers to the energy value at different frequency values ​​at time t, the first frequency The corresponding energy value is the smallest.

[0124] b. According to the probability, the energy of the speech data in the first time segment is used to obtain the first segment initial speech response degree of the speech data at the frequency to be analyzed.

[0125] Wherein, according to the probability, obtaining a first segment initial voice response degree of the voice data at the frequency to be analyzed in the first time segment includes:

[0126] Obtaining the sum of the products of the energy value corresponding to the frequency to be analyzed and the probability corresponding to the frequency to be analyzed at different times within the first time segment;

[0127] The sum of the products is divided by the duration of the first time segment to obtain the first segment initial voice response degree.

[0128] Specifically, the first The time segment corresponding to each segment is , then in this time segment, record the Audio data at frequency The degree of voice response under for:

[0129]

[0130] in, Indicates Time interval segmentation points, Similarly, Indicates Spectrogram energy data of speech data; Indicates energy data at time , the frequency is The energy size when Indicates time Next, frequency That is, the probability that the frequency to be analyzed is the frequency of voice data.

[0131] S603. Multiplying the first segmented initial voice response degree by the normalization result to obtain a first segmented voice response degree of the voice data at the frequency to be analyzed;

[0132]

[0133] in, represents the normalization function, represents the mean function, Indicates Time segment points, Indicates the time interval Internal time The degree of speech confusion, Indicates Audio data in the interval Intrinsic Frequency The degree of voice response under is in the time interval Internal time The degree of voice response under.

[0134] S604. Repeat the above steps to sequentially traverse each of the time segments and change the frequency value to be analyzed to obtain the voice response degree of the voice data at different frequencies in each time segment.

[0135] In summary, the present invention has the following beneficial effects:

[0136] The present invention uses a multi-dimensional AI platform to identify the voice data in the car. First, the audio receiving device built in the smart cockpit is used to collect the voice data emitted by the people in the cockpit. The number of audio receiving devices is at least two. These voice data often include the voice data of each person in the vehicle, and also include various noises such as vehicle operation noise. In order to make the voice data of different people in the smart cabin more accurately separated and enhanced, the corresponding original filter group is corrected according to the voice response degree of the voice data at different frequencies, and the center frequency of the original filter group is corrected. The mel spectrogram generated by the corrected mel filter group can better characterize the voice characteristics of the corresponding people in the cockpit. In the process of mel spectrum conversion, the frequency component of the voice part of the main character is enhanced, so that the mel spectrum can adaptively adjust the voice part component of its concern. In this way, the voice characteristics of the main character can be captured more effectively, the sensitivity and recognition effect of the voice recognition system to the voice of the main character can be improved, and the robustness of the vehicle-mounted voice recognition system in a complex in-vehicle environment has also been greatly improved.

[0137] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0138] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

Claims

1. A speech recognition method for a multi-dimensional AI platform, characterized in that: The method comprises: Acquire voice data sent from at least two audio receiving devices; Modifying the corresponding original filter group according to the speech response degree of the speech data at different frequencies to obtain a modified filter group; Using the modified filter bank to generate a Mel-spectrogram of the speech data; Input the Mel spectrograms into a preset speech recognition network to obtain speech recognition results corresponding to the speech data; Correspondingly modifying the original filter bank based on the speech response degree includes: The center frequency of the original filter to be corrected is the initial center frequency, and the previous initial center frequency of the previous filter before correction and the previous corrected center frequency after correction of the previous filter of the original filter to be corrected are obtained; Obtaining a mean value of the speech response degree in a frequency interval from the previous initial center frequency to the initial center frequency, taking the inverse of the mean value of the speech response degree and normalizing it to obtain a normalized result; Subtracting the previous initial center frequency from the initial center frequency to obtain a difference result; After multiplying the normalization result and the difference result, the result is added to the previous corrected center frequency to obtain the corrected center frequency of the original filter to be corrected; Modifying the corresponding original filter group according to the speech response degree of the speech data at different frequencies, and obtaining the modified filter group includes: Performing independent component analysis on each of the speech data to obtain independent components, and converting each of the speech data and each of the independent components into a spectrogram; Acquiring the speech response degree of each of the speech data at different frequencies based on the spectrogram; Correspondingly modifying the original filter bank based on the speech response degree; Acquiring the speech response degree of each of the speech data at different frequencies based on the spectrogram includes: Obtaining the voice confusion degree of each of the voice data at each time; Dividing the speech data into time periods according to the corresponding speech confusion degree; Calculating the speech response degree of the speech data at different frequencies; The obtaining of the voice confusion degree of each voice data at each time comprises: Acquire energy change data of independent components in each of the speech data; Selecting corresponding energy change data to be analyzed according to the energy change data; Respectively collecting the frequencies to be analyzed corresponding to the energy change data to be analyzed and sorting the frequencies to be analyzed; The speech confusion degree is calculated based on the sorted frequencies to be analyzed.

2. The speech recognition method for a multi-dimensional AI platform according to claim 1, characterized in that: Dividing the voice data into time periods according to the corresponding voice confusion degree includes: Obtaining a mean of the first-order differences of the speech confusion degree with respect to a first time, where the first time is within a time range of a preset time neighborhood from a time point to be checked; When the first-order difference of the speech confusion degree about the to-be-tested time point is greater than the mean value, the to-be-tested time point is a time segmentation point, and the time segmentation point is used to divide the speech data into the time periods.

3. The speech recognition method for a multi-dimensional AI platform according to claim 1, characterized in that: Calculating the speech response degree of the speech data at different frequencies includes: Normalizing the inverse of the mean value of the speech confusion degree in the first time segment to obtain a normalized result, wherein the first time segment is any time segment after the time period is divided; Acquire a first segment initial voice response degree of the voice data at the frequency to be analyzed within the first time segment; Multiplying the first segmented initial voice response degree by the normalization result to obtain a first segmented voice response degree of the voice data at the frequency to be analyzed; The above steps are repeated to sequentially traverse each of the time segments and replace the first segment voice response degree under the frequency to be analyzed, so as to obtain the voice response degree of the voice data at different frequencies in each time segment.

4. The speech recognition method for a multi-dimensional AI platform according to claim 3, characterized in that: Acquiring a first segment initial voice response degree of the voice data at the frequency to be analyzed in the first time segment includes: In the first time segment, the energy of the independent components in the speech data at different frequencies is obtained, and the maximum value of the energy ratio at the frequency to be analyzed is calculated, where the maximum value is the probability that the frequency to be analyzed is the frequency of the speech data; The first segment initial speech response degree of the speech data at the frequency to be analyzed is obtained based on the energy of the speech data in the first time segment according to the probability.

5. The speech recognition method for a multi-dimensional AI platform according to claim 4, characterized in that: The method further includes: the process of obtaining the energy ratio includes: Acquire a minimum value of energy of the independent component of the speech data at a first frequency, where the first frequency is a frequency corresponding to a minimum energy of the independent component of the speech data; Add a preset positive number to the minimum value to obtain a preset denominator; The energy ratio is obtained by dividing the energy value of the independent component of the speech data at the frequency to be analyzed by the preset denominator.

6. The speech recognition method for a multi-dimensional AI platform according to claim 4, characterized in that: The method further includes: obtaining, according to the probability, a first segment initial voice response degree of the voice data at the frequency to be analyzed in the first time segment, comprising: Obtaining the sum of the products of the energy value corresponding to the frequency to be analyzed and the probability corresponding to the frequency to be analyzed at different times within the first time segment; The sum of the products is divided by the duration of the first time segment to obtain the first segment initial voice response degree.

Images