An ordering method based on AI voice interaction

Abstract

The present invention discloses an ordering method based on AI voice interaction, which relates to the field of voice interaction technology, including: constructing a commodity keyword database, and pre-storing a standard voice signal sequence corresponding to each record in the commodity keyword database; training a finite-length impulse response adaptive voice filter; processing the user's voice input by the finite-length impulse response adaptive voice filter to obtain a user voice signal sequence without background sound; cutting the user voice signal sequence according to the pause in the user voice signal sequence to obtain a recombined sequence of each user voice signal; performing voice feature matching on each recombined sequence of user voice signals and each standard voice signal sequence; and completing the ordering according to the voice feature matching result. The present invention significantly improves the ordering efficiency, reduces the complexity of user operations, enhances the user interaction experience, adapts to a variety of usage scenarios, and has good intelligent learning and optimization capabilities.

Description

Technical Field

[0001] The present invention relates to the field of voice interaction technology, and in particular to an order placing method based on AI voice interaction. Background Art

[0002] With the rapid development of Internet technology, online shopping has become a part of people's daily life.

[0003] The traditional ordering method mainly relies on users to manually input product information. This method is cumbersome and inefficient in operation, not convenient for use in mobile scenarios, and has certain usage barriers for specific groups such as the visually impaired and the elderly.

[0004] Therefore, it is particularly important to develop a more convenient and intelligent way to place orders. Summary of the invention

[0005] In response to the above-mentioned deficiencies in the prior art, the present invention provides an ordering method based on AI voice interaction, which solves the problems of existing online shopping ordering methods such as cumbersome user operations, low input efficiency, inconvenience for use in mobile scenarios, and lack of friendliness to specific user groups.

[0006] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0007] An order placing method based on AI voice interaction includes the following steps:

[0008] S1. Build a product keyword database and pre-store standard voice signal sequences corresponding to each record in the product keyword database;

[0009] S2, according to each standard speech signal sequence, training a finite length impulse response adaptive speech filter for background sound suppression and speech signal enhancement;

[0010] S3, receiving the user's voice input, and processing it by a finite-length impulse response adaptive voice filter to obtain a user voice signal sequence without background sound;

[0011] S4, cutting the user voice signal sequence according to the pauses in the user voice signal sequence to obtain a recombined sequence of each user voice signal;

[0012] S5, performing voice feature matching on each user voice signal recombined sequence and each standard voice signal sequence;

[0013] S6. Complete the order based on the voice feature matching results.

[0014] The beneficial effects of the present invention are as follows: the present invention pre-stores product information and its standard voice through database technology, suppresses the background noise of user voice through adaptive filtering, obtains effective user voice data, and learns the user's order intention based on database traversal matching to complete the order. The present invention significantly improves order efficiency, reduces user operation complexity, enhances user interaction experience, adapts to a variety of usage scenarios, and has good intelligent learning and optimization capabilities.

[0015] Furthermore, the commodity keyword database is a relational database, whose fields include commodity name, commodity specification and commodity quantity, and uses the field commodity name as the primary key.

[0016] The beneficial effect of the above further scheme is that the product name is an item in the product keyword database that cannot be empty and must have a unique value. A relational database is built on this basis to effectively organize the product data, facilitate subsequent voice recognition and matching, and complete user orders.

[0017] Furthermore, the operation expression of the finite length impulse response adaptive speech filter is:

[0018] ,

[0019] Among them, voice out (k) is the value of the speech signal sequence at time k after being processed by the finite-length impulse response adaptive speech filter, voice in (k) is the value of the speech signal sequence at time kn before being processed by the finite-length impulse response adaptive speech filter, w n is the nth order coefficient of the finite length impulse response adaptive speech filter, w N-1 is the N-1th order coefficient of the finite length impulse response adaptive speech filter, n is a positive integer greater than or equal to 0 and less than or equal to N-2, k is an integer, and N is the order of the finite length impulse response adaptive speech filter.

[0020] Furthermore, the S2 comprises the following sub-steps:

[0021] S21, recording the sounds of each standard voice signal sequence through a device for receiving user voice input, to obtain each voice sample collection sequence;

[0022] S22, splicing each voice sample collection sequence and each standard voice signal sequence into a voice sample collection long sequence and a standard voice signal long sequence respectively;

[0023] S23, setting the initial values ​​of each order coefficient of the finite length impulse response adaptive speech filter, wherein the initial values ​​of the 0th to N-2th order coefficients are set to 0, and the initial value of the N-1th order coefficient is set according to the following formula:

[0024] ,

[0025] Among them, w N-1 (0) is the initial value of the N-1th order coefficient of the finite length impulse response adaptive speech filter, is the mean value of each moment in the long sequence of standard speech signals, Collect the mean value of each moment in the long sequence for the speech sample;

[0026] S24, cyclically inputting a long sequence of speech samples into the finite length impulse response adaptive speech filter, and taking the value of the long sequence of standard speech signals at the corresponding time as the output target of the finite length impulse response adaptive speech filter, and training the finite length impulse response adaptive speech filter until all order coefficients of the finite length impulse response adaptive speech filter converge.

[0027] Further, the S24 trains a finite-length impulse response adaptive speech filter by the following formula:

[0028] ,

[0029] ,

[0030] Among them, w n (k+1) is the training value of the nth order coefficient of the finite length impulse response adaptive speech filter at time k+1, w n (k) The training value of the nth order coefficient of the finite-length impulse response adaptive speech filter at time k, ς is the learning rate, ς is greater than 0 and less than 2, P voice is the average power of each standard voice signal sequence, voice st (k) is the value of the standard speech signal long sequence at time k, Collect the values ​​of the long sequence kn times for the speech sample, is the value at the kth moment obtained after a long sequence of speech samples is input into the finite length impulse response adaptive speech filter, w N-1 (k+1) is the training value of the N-1th order coefficient of the finite length impulse response adaptive speech filter at time k+1, w N-1 (k) is the training value of the N-1th order coefficient of the finite length impulse response adaptive speech filter at time k.

[0031] The beneficial effects of the above further scheme are: a finite-length impulse response adaptive speech filter based on time domain taps, multiplication and addition operations is constructed, which has good intelligent learning and optimization capabilities. Through target learning, background sounds, noise and other noises are suppressed, leaving only the effective user voice signal, which greatly improves the signal-to-noise ratio, which is equivalent to enhancing the voice signal, and is also convenient for subsequent cutting and reorganization operations.

[0032] Furthermore, the S5 comprises the following sub-steps:

[0033] S51, respectively calculating the speech feature vectors of the reconstructed sequences of the user speech signals and the speech feature vectors of the standard speech signal sequences;

[0034] S52. For each speech feature vector of the reconstructed sequence of the user speech signal, the similarity between it and the speech feature vectors of all standard speech signal sequences is traversed. If the similarity is greater than a threshold, it is regarded as a speech feature match, and the user is regarded as selecting the corresponding record in the product keyword database.

[0035] Furthermore, the format of the speech feature vector in S51 and S52 is [λ(1), λ(2), λ(3), λ(4), λ(5), λ(6), λ(7), λ(8), λ(9), λ(10), λ(11), λ(12), λ(13), λ(14), λ(15)];

[0036] Where, λ(1) is the voiced frequency;

[0037] λ(2) is the fundamental frequency;

[0038] λ(3) is the peak frequency of the power spectrum;

[0039] λ(4) to λ(15) are the 2nd to 13th Mel-frequency cepstrum coefficients respectively.

[0040] The beneficial effects of the above further scheme are as follows: voiced sounds are sounds produced by the vibration of the vocal cords when people pronounce words. The vowels in most languages ​​are voiced sounds, and their frequencies are the "fingerprints" of language vocabulary; fundamental frequency and power spectrum peak frequency also constitute important features of language vocabulary; Mel-Frequency Cepstral Coefficients (MFCCs) are coefficients that make up the Mel-Frequency Cepstral, which are nonlinear scales of human frequency perception and are used to measure phonemes. The present invention extracts the 2nd to 13th Mel-Frequency Cepstral coefficients, which helps to identify language content and discards all other information such as emotions. The present invention combines these voiceprint features to form a 15-dimensional speech feature vector, and performs speech recognition by traversal matching.

[0041] Furthermore, the S52 calculates the similarity of the speech feature vectors by the following formula:

[0042] ,

[0043] Among them, ρ a,bis the similarity between speech feature vector a and speech feature vector b, tanh is the hyperbolic tangent function, λ a (i) is the i-th element of the speech feature vector a, λ b (i) is the i-th element of the speech feature vector b, and L is the length of the speech feature vector.

[0044] The beneficial effect of the above further scheme is: the above formula can compare the value differences of each element in the two speech feature vectors, and provide gain for small differences through the hyperbolic tangent function, attenuate large differences, improve the recognition of small differences, and by taking the inverse, the more similar they are, the larger the similarity value is, which is convenient for subsequent threshold judgment.

[0045] Furthermore, S6 determines the product name, product specifications and product quantity input by the user's voice according to the voice feature matching result, confirms the order result to the user, and completes the order. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flowchart of an ordering method based on AI voice interaction provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0048] like Figure 1 As shown, in one embodiment of the present invention, a method for placing an order based on AI voice interaction includes the following steps:

[0049] S1. Build a commodity keyword database, and pre-store a standard voice signal sequence corresponding to each record in the commodity keyword database.

[0050] The commodity keyword database is a relational database, whose fields include commodity name, commodity specification and commodity quantity, and uses the commodity name field as the primary key.

[0051] The product name is an item in the product keyword database that cannot be empty and must have a unique value. A relational database is built on this basis to effectively organize product data, facilitate subsequent voice recognition and matching, and complete user orders.

[0052] S2. According to each standard speech signal sequence, a finite-length impulse response adaptive speech filter for background sound suppression and speech signal enhancement is trained.

[0053] The operational expression of the finite-length impulse response adaptive speech filter is:

[0054] ,

[0055] Among them, voice out (k) is the value of the speech signal sequence at time k after being processed by the finite-length impulse response adaptive speech filter, voice in (k) is the value of the speech signal sequence at time kn before being processed by the finite-length impulse response adaptive speech filter, w n is the nth order coefficient of the finite length impulse response adaptive speech filter, w N-1 is the N-1th order coefficient of the finite length impulse response adaptive speech filter, n is a positive integer greater than or equal to 0 and less than or equal to N-2, k is an integer, and N is the order of the finite length impulse response adaptive speech filter.

[0056] S2 includes the following sub-steps:

[0057] S21, recording the sounds of each standard voice signal sequence through a device for receiving user voice input, to obtain each voice sample collection sequence;

[0058] S22, splicing each voice sample collection sequence and each standard voice signal sequence into a voice sample collection long sequence and a standard voice signal long sequence respectively;

[0059] S23, setting the initial values ​​of each order coefficient of the finite length impulse response adaptive speech filter, wherein the initial values ​​of the 0th to N-2th order coefficients are set to 0, and the initial value of the N-1th order coefficient is set according to the following formula:

[0060] ,

[0061] Among them, w N-1 (0) is the initial value of the N-1th order coefficient of the finite length impulse response adaptive speech filter, is the mean value of each moment in the long sequence of standard speech signals, Collect the mean value of each moment in the long sequence for the speech sample;

[0062] S24, cyclically inputting a long sequence of speech samples into the finite length impulse response adaptive speech filter, and taking the value of the long sequence of standard speech signals at the corresponding time as the output target of the finite length impulse response adaptive speech filter, and training the finite length impulse response adaptive speech filter by the following formula until all order coefficients of the finite length impulse response adaptive speech filter converge:

[0063] ,

[0064] ,

[0065] Among them, w n (k+1) is the training value of the nth order coefficient of the finite length impulse response adaptive speech filter at time k+1, w n (k) The training value of the nth order coefficient of the finite-length impulse response adaptive speech filter at time k, ς is the learning rate, ς is greater than 0 and less than 2, P voice is the average power of each standard voice signal sequence, voice st (k) is the value of the standard speech signal long sequence at time k, Collect the values ​​of the long sequence kn times for the speech sample, is the value at the kth moment obtained after a long sequence of speech samples is input into the finite length impulse response adaptive speech filter, w N-1 (k+1) is the training value of the N-1th order coefficient of the finite length impulse response adaptive speech filter at time k+1, w N-1 (k) is the training value of the N-1th order coefficient of the finite length impulse response adaptive speech filter at time k.

[0066] The above steps of this embodiment construct a finite-length impulse response adaptive speech filter based on time domain taps, multiplication and addition operations, which has good intelligent learning and optimization capabilities. It suppresses background sounds, noises and other noises through target learning, leaving only effective user voice signals, which greatly improves the signal-to-noise ratio, is equivalent to enhancing the voice signal, and is also convenient for subsequent cutting and reorganization operations.

[0067] The learning rate ς determines the iteration speed of the finite-length impulse response adaptive speech filter. In the range of 0 to 2, the larger the value, the more obvious the iteration change, which can quickly learn, but may cause over-limit oscillation; the smaller the value, the slower the iteration change, but the more stable it is. It can be set according to preference.

[0068] S3. Receive the user's voice input, and process it with a finite-length impulse response adaptive voice filter to obtain a user voice signal sequence without background sound.

[0069] S4. Cut the user voice signal sequence according to the pauses in the user voice signal sequence to obtain a reconstructed sequence of each user voice signal.

[0070] The suppression effect of the finite-length impulse response adaptive speech filter on the background sound makes it easy to implement the cutting and reorganization in step S4. For example, the situation where a certain length of time value is less than a certain value in the user's speech signal sequence can be regarded as a pause, and then the cutting is performed. The time length and the value used for judgment can be adjusted and optimized, which will not be repeated in this article.

[0071] S5, performing voice feature matching on each user voice signal recombinant sequence and each standard voice signal sequence, including the following sub-steps:

[0072] S51, respectively calculating the speech feature vectors of each user speech signal reconstructed sequence and the speech feature vectors of each standard speech signal sequence.

[0073] The format of the speech feature vector is [λ(1), λ(2), λ(3), λ(4), λ(5), λ(6), λ(7), λ(8),λ(9), λ(10), λ(11), λ(12), λ(13), λ(14), λ(15)], where: λ(1) is the voiced frequency; λ(2) is the fundamental frequency; λ(3) is the power spectrum peak frequency; λ(4) to λ(15) are the 2nd to 13th Mel-frequency cepstrum coefficients respectively.

[0074] Voiced sounds are sounds produced by the vibration of the vocal cords when people pronounce words. Vowels in most languages ​​are voiced sounds, and their frequencies are the "fingerprints" of language vocabulary; fundamental frequency and power spectrum peak frequency also constitute important features of language vocabulary; Mel-Frequency Cepstral Coefficients (MFCCs) are coefficients that make up the Mel-Frequency Cepstral, which is a nonlinear scale of human frequency perception and is used to measure phonemes. The present invention extracts the 2nd to 13th Mel-Frequency Cepstral coefficients, which helps to identify language content and discards all other information such as emotions. The present invention combines these voiceprint features to form a 15-dimensional speech feature vector, and performs speech recognition by traversing and matching.

[0075] The calculation methods of voiced frequency and Mel frequency cepstrum coefficients can be found in the paper; the fundamental frequency and power spectrum peak frequency are obtained through FFT transformation. This article will not go into details.

[0076] S52. For each speech feature vector of the reconstructed sequence of the user speech signal, the similarity between it and the speech feature vectors of all standard speech signal sequences is traversed. If the similarity is greater than a threshold, it is regarded as a speech feature match, and the user is regarded as selecting the corresponding record in the product keyword database.

[0077] The similarity threshold can be debugged in actual working conditions.

[0078] The similarity is calculated by the following formula:

[0079] ,

[0080] Among them, ρ a,b is the similarity between speech feature vector a and speech feature vector b, tanh is the hyperbolic tangent function, λa (i) is the i-th element of the speech feature vector a, λ b (i) is the i-th element of the speech feature vector b, and L is the length of the speech feature vector.

[0081] The above formula can compare the value differences of each element in the two speech feature vectors, and provide gain for small differences through the hyperbolic tangent function, attenuate large differences, improve the recognition of small differences, and take the inverse so that the more similar they are, the larger the similarity value is, which is convenient for subsequent threshold judgment.

[0082] S6 determines the product name, product specifications and product quantity input by the user's voice according to the voice feature matching result, confirms the order result to the user, and completes the order.

[0083] In summary, the present invention pre-stores product information and its standard voice through database technology, suppresses the background noise of user voice through adaptive filtering, obtains effective user voice data, and learns the user's order intention based on database traversal matching to complete the order. The present invention significantly improves order efficiency, reduces user operation complexity, enhances user interaction experience, adapts to a variety of usage scenarios, and has good intelligent learning and optimization capabilities.

[0084] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0085] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. An order placing method based on AI voice interaction, characterized in that: The following steps are involved: S1. Build a product keyword database and pre-store standard voice signal sequences corresponding to each record in the product keyword database; S2, according to each standard speech signal sequence, training a finite length impulse response adaptive speech filter for background sound suppression and speech signal enhancement; S3, receiving the user's voice input, and processing it by a finite-length impulse response adaptive voice filter to obtain a user voice signal sequence without background sound; S4, cutting the user voice signal sequence according to the pauses in the user voice signal sequence to obtain a recombined sequence of each user voice signal; S5, performing voice feature matching on each user voice signal recombined sequence and each standard voice signal sequence; S6. Complete the order based on the voice feature matching result; The S2 comprises the following sub-steps: S21, recording the sounds of each standard voice signal sequence through a device for receiving user voice input, to obtain each voice sample collection sequence; S22, splicing each voice sample collection sequence and each standard voice signal sequence into a voice sample collection long sequence and a standard voice signal long sequence respectively; S23, setting the initial values ​​of each order coefficient of the finite length impulse response adaptive speech filter, wherein the initial values ​​of the 0th to N-2th order coefficients are set to 0, and the initial value of the N-1th order coefficient is set according to the following formula: ， Among them, w N-1 (0) is the initial value of the N-1th order coefficient of the finite length impulse response adaptive speech filter, is the mean value of each moment in the long sequence of standard speech signals, Collect the mean value of each moment in the long sequence for the speech sample; S24, cyclically inputting a long sequence of speech sample collection into the finite length impulse response adaptive speech filter, and taking the value of the long sequence of standard speech signals at the corresponding time as the output target of the finite length impulse response adaptive speech filter, and training the finite length impulse response adaptive speech filter until all order coefficients of the finite length impulse response adaptive speech filter converge; The S24 is trained with a finite-length impulse response adaptive speech filter as follows: ， ， Among them, w n (k+1) is the training value of the nth order coefficient of the finite length impulse response adaptive speech filter at time k+1, w n (k) The training value of the nth order coefficient of the finite-length impulse response adaptive speech filter at time k, ς is the learning rate, ς is greater than 0 and less than 2, P voice is the average power of each standard voice signal sequence, voice st (k) is the value of the standard speech signal long sequence at time k, Collect the values ​​of the long sequence kn times for the speech sample, is the value at the kth moment obtained after a long sequence of speech samples is input into the finite length impulse response adaptive speech filter, w N-1 (k+1) is the training value of the N-1th order coefficient of the finite length impulse response adaptive speech filter at time k+1, w N-1 (k) is the training value of the N-1th order coefficient of the finite length impulse response adaptive speech filter at time k.

2. The order placing method based on AI voice interaction according to claim 1, characterized in that: The commodity keyword database is a relational database, whose fields include commodity name, commodity specification and commodity quantity, and uses the commodity name field as the primary key.

3. The order placing method based on AI voice interaction according to claim 1, characterized in that: The operational expression of the finite length impulse response adaptive speech filter is: ， Among them, voice out (k) is the value of the speech signal sequence at time k after being processed by the finite-length impulse response adaptive speech filter, voice in (k) is the value of the speech signal sequence at time kn before being processed by the finite-length impulse response adaptive speech filter, w n is the nth order coefficient of the finite length impulse response adaptive speech filter, w N-1 is the N-1th order coefficient of the finite length impulse response adaptive speech filter, n is a positive integer greater than or equal to 0 and less than or equal to N-2, k is an integer, and N is the order of the finite length impulse response adaptive speech filter.

4. The order placing method based on AI voice interaction according to claim 1, characterized in that: The S5 comprises the following sub-steps: S51, respectively calculating the speech feature vectors of the reconstructed sequences of the user speech signals and the speech feature vectors of the standard speech signal sequences; S52. For each speech feature vector of the reconstructed sequence of the user speech signal, the similarity between it and the speech feature vectors of all standard speech signal sequences is traversed. If the similarity is greater than a threshold, it is regarded as a speech feature match, and the user is regarded as selecting the corresponding record in the product keyword database.

5. The order placing method based on AI voice interaction according to claim 4 is characterized in that: The format of the speech feature vector in S51 and S52 is [λ(1), λ(2), λ(3), λ(4), λ(5), λ(6), λ(7), λ(8), λ(9), λ(10), λ(11), λ(12), λ(13), λ(14), λ(15)]; Where, λ(1) is the voiced frequency; λ(2) is the fundamental frequency; λ(3) is the peak frequency of the power spectrum; λ(4) to λ(15) are the 1st to 12th Mel-frequency cepstrum coefficients, respectively.

6. The order placing method based on AI voice interaction according to claim 4 is characterized in that: The S52 calculates the similarity of the speech feature vectors by the following formula: ， Among them, ρ a,b is the similarity between speech feature vector a and speech feature vector b, tanh is the hyperbolic tangent function, λ a (i) is the i-th element of the speech feature vector a, λ b (i) is the i-th element of the speech feature vector b, and L is the length of the speech feature vector.

7. The order placing method based on AI voice interaction according to claim 2, characterized in that: The S6 determines the product name, product specifications and product quantity input by the user's voice according to the voice feature matching result, confirms the order result to the user, and completes the order.

Images