A low-resource method for detecting infant crying

Abstract

The present invention relates to the field of Internet technology, and discloses a low-resource baby cry detection method, comprising the following steps: using a monophone model of 1 mixture, with a state number of 209, to perform preliminary matching and screening; sorting the viterbi matching scores of the monophone model, and selecting the phoneme models corresponding to the N results with the highest scores for further fine matching; using a triphone model of 3 mixtures, with a state number of 358, to perform fine matching and screening; sorting the viterbi matching scores of multiple triphone models, and selecting the sequence number corresponding to the model with the highest score as the final recognition result. The technical solution of the present invention can greatly reduce the model structure complexity and calculation complexity under the premise of ensuring the accuracy of the recognition algorithm by adopting a model dimensionality reduction method based on low mixture number + multi-state sharing and a non-real-time calculation method based on VAD triggering, thereby significantly reducing the chip hardware computing resource consumption required for the baby cry detection algorithm.

Description

Technical Field

[0001] The present invention relates to the field of Internet technology, and in particular to a low-resource baby crying detection method. Background Art

[0002] Baby cry detection technology has a wide range of applications in education, smart homes, and infant products. However, due to the high computational complexity of traditional baby cry detection algorithms, robust recognition algorithms typically require computing power exceeding 200 MIPS. Consequently, many chips with only a few dozen MIPS of computing power are unable to implement high-performance baby cry detection algorithms, hindering the widespread adoption of this technology. Summary of the Invention

[0003] The main purpose of this invention is to propose a low-resource baby cry detection method, which aims to significantly reduce the hardware system resource requirements of the baby cry recognition algorithm while ensuring recognition performance.

[0004] To achieve the above objectives, the present invention proposes a low-resource infant crying detection method, comprising the following steps:

[0005] Use 1mixture's monophone sub-model (number of states is 209) for preliminary matching and screening;

[0006] Sort the Viterbi matching scores of the monophone models and select the phoneme models corresponding to the N results with the highest scores for further fine matching;

[0007] Use 3mixture's triphone model (358 states) for fine matching and screening;

[0008] Sort the Viterbi matching scores of the N triphone models, and select the serial number corresponding to the model with the highest score as the final recognition result.

[0009] Furthermore, the preliminary matching and screening using the monophone sub-model of 1mixture (number of states is 209) includes the following steps:

[0010] The initial number of states of the monophonic sub-model is 1024;

[0011] The Baum-Welch parameter reestimation method is used to cluster the initial monophonic submodel states using the labeled crying training database, reducing the number of valid states M from 1024 to 209 (the overall likelihood score of the output probability density function is the highest). The specific process is as follows: Assume that S is the set of HMM model states, F is the training set, L(S) is the log-likelihood value, and assume that the bound states share the mean and variance of the Gaussian distribution, and that state binding does not affect the alignment of frames and states.

[0012] L(S)= in is the posterior probability of the state S generating the feature. Assuming that the output probability density function is Gaussian distribution, then L(S)=

[0013] In the formula, n is the data dimension (n is 27 in this example);

[0014] Select L(S) As the first number of L(S) is the likelihood value of the valid state 0, compare L(S) To L(S) Zhong and L(S) States with a difference of less than 5% are merged into the same state;

[0015] Repeat the above step S103 to obtain L(S) ~L(S) .

[0016] Furthermore, the fine matching and screening using the 3mixture triphone model (number of states is 358) includes the following steps:

[0017] The initial number of states of the triphone model is 1500;

[0018] Using the Baum-Welch parameter reestimation method, the initial three-phone model states were clustered using the labeled crying training database, reducing the number of valid states T from 1500 to 358 (the overall likelihood score of the output probability density function was the highest);

[0019] The specific process is as follows: Assume that S is the set of HMM model states, F is the training set, L(S) is the log-likelihood value, and assume that the bound states share the mean and variance of the Gaussian distribution, and that state binding does not affect the alignment of frames and states, then

[0020] L(S)=

[0021] in is the posterior probability of the state S generating the feature. Assuming that the output probability density function is Gaussian distribution, then L(S)=

[0022] In the formula, n is the data dimension (n is 27 in this example);

[0023] S304: Select L(S) The first number of L(S) is the likelihood value of the valid state 0, and L(S) is compared. To L(S) Zhong and L(S) States with a difference of less than 5% are merged into the same state;

[0024] S305: Repeat the above step S304 to obtain L(S) ~L(S) .

[0025] The technical solution of the present invention has the following beneficial effects: the technical solution of the present invention, by adopting a model dimensionality reduction method based on low mixture number + multi-state sharing and a non-real-time calculation method based on VAD triggering, can greatly reduce the model structure complexity and calculation complexity while ensuring the accuracy of the recognition algorithm, thereby significantly reducing the chip hardware computing resource consumption required for the baby crying detection algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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 the structures shown in these drawings without paying any creative work.

[0027] Figure 1 This is a schematic diagram of the overall framework process of a low-resource infant crying detection method proposed by the present invention.

[0028] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0029] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0030] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0031] In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the fact that ordinary technicians in this field can implement them. When the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0032] The present invention provides a low-resource baby crying detection method.

[0033] like Figure 1 As shown, in one embodiment of the present invention, the low-resource infant crying detection method includes the following steps:

[0034] S100: Use 1mixture's monophone sub-model (number of states is 209) for preliminary matching and screening:

[0035] S200: sorting the viterbi matching scores of the monophone models, and selecting the phoneme models corresponding to the N results with the highest scores for further fine matching;

[0036] S300: Use 3mixture's triphone model (358 states) for fine matching and screening:

[0037] S400: Sort the Viterbi matching scores of the N triphone models, and select the serial number corresponding to the model with the highest score as the final recognition result.

[0038] Specifically, the aforementioned S100: using the monophone sub-model of 1mixture (number of states is 209) for preliminary matching and screening includes the following steps:

[0039] S101: The initial number of states of the monophonic sub-model is 1024;

[0040] S102: Using the Baum-Welch parameter reestimation method, the initial monophonic sub-model states are clustered using the labeled crying training database, reducing the number of valid states M from 1024 to 209 (the overall likelihood score of the output probability density function is the highest). The specific process is as follows: Assume that S is the set of HMM model states, F is the training set, L(S) is the log-likelihood value, and that the bound states share the mean and variance of the Gaussian distribution, and that state binding does not affect the alignment of frames and states.

[0041] L(S)=

[0042] in is the posterior probability of the state S generating the feature. Assuming that the output probability density function is Gaussian distribution, then L(S)=

[0043] In the formula, n is the data dimension (n is 27 in this example);

[0044] S103: Select L(S) As the first number of L(S) is the likelihood value of the valid state 0, compare L(S) To L(S) Zhong and L(S) States with a difference of less than 5% are merged into the same state;

[0045] S104: Repeat the above step S103 to obtain L(S) ~L(S) .

[0046] Specifically, the S300: using the 3mixture triphone model (number of states is 358) for fine matching and screening includes the following steps:

[0047] S301: The number of initial triphone model states is 1500;

[0048] S302: Using the Baum-Welch parameter reestimation method, cluster the initial three-phone model states using the labeled crying training database, reducing the number of valid states T from 1500 to 358 (the overall likelihood score of the output probability density function is the highest);

[0049] S303: The specific process is as follows: Assume that S is the set of HMM model states, F is the training set, L(S) is the log-likelihood value, and assume that the bound states share the mean and variance of the Gaussian distribution, and that state binding does not affect the alignment of frames and states, then

[0050] L(S)=

[0051] in is the posterior probability of the state S generating the feature. Assuming that the output probability density function is Gaussian distribution, then L(S)=

[0052] In the formula, n is the data dimension (n is 27 in this example);

[0053] S304: Select L(S) The first number of L(S) is the likelihood value of the valid state 0, and L(S) is compared. To L(S) Zhong and L(S) States with a difference of less than 5% are merged into the same state;

[0054] S305: Repeat the above step S304 to obtain L(S) ~L(S) .

[0055] Specifically, compared with traditional methods, the computational complexity of the low-resource infant cry detection method proposed in the present invention is reduced by nearly 20 times, and the storage complexity is reduced by more than 10 times, so that a high-performance infant cry detection application solution can be implemented on an ordinary single-chip microcomputer with a main frequency below 100MHz.

[0056] On the other hand, the present invention also provides a non-real-time calculation method based on VAD triggering, which adopts a five-state switching technology to replace the real-time calculation with a large amount of computation, thereby reducing the computational complexity and operating power consumption. The method includes the following steps:

[0057] 1. The initial state of VAD is silence. At this time, the system runs at a low frequency (no more than 2MHz) and only determines the current state of the voice frame.

[0058] 2. If the current frame energy is greater than the activation threshold TA, it is in the active state. At this time, the system still operates at a low frequency (no more than 2MHz) and only determines the current state of the voice frame.

[0059] 3. If the current frame energy is greater than the voice threshold TO, or the current frame energy is greater than the activation threshold TA for more than three times in a row, the state switches to on. At this time, the system still runs at a low frequency (no more than 2MHz) and only determines the current state of the voice frame.

[0060] 4. If the current frame energy is less than the silence threshold TI, the state switches to inactive. At this time, the system still runs at a low frequency (no more than 2MHz) and only determines the current state of the voice frame.

[0061] 5. If the current frame energy is less than the silence threshold (TI) for more than 16 consecutive times, the state switches to end. At this time, the system switches to high frequency (greater than 50MHz) and calculates the matching score with each model for the speech between active and end.

[0062] 6. After the model matching score calculation is completed, the current frame state switches to silence; at this time, the system switches to a low frequency (no more than 2MHz) and only determines the current state of the voice frame;

[0063] 7. Repeat steps 2 to 6 above.

[0064] Compared with traditional methods, the VAD-triggered non-real-time calculation method proposed in the present invention only performs intensive calculations on valid speech, which can significantly reduce the amount of calculation and reduce operating power consumption.

[0065] Specifically, the present invention adopts a model dimensionality reduction method based on low mixture number + multi-state sharing and a non-real-time calculation method based on VAD triggering, which greatly reduces the model structure complexity and calculation complexity while ensuring the accuracy of the recognition algorithm, thereby significantly reducing the chip hardware computing resource consumption required for the baby crying detection algorithm.

[0066] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. All equivalent structural transformations made by using the contents of the present invention description and drawings under the inventive concept of the present invention, or direct / indirect application in other related technical fields are included in the patent protection scope of the present invention.

Claims

1. A low-resource baby crying detection method, characterized in that: The steps include: Use 1mixture's monophone model with 209 states for preliminary matching and screening; Sort the viterbi matching scores of the monophone models, and select the phoneme models corresponding to the N results with the highest scores for further fine matching; Use 3mixture's triphone model with 358 states for fine matching and screening; Sort the Viterbi matching scores of multiple triphone models, and select the serial number corresponding to the model with the highest score as the final recognition result; The monophone single-phone sub-model using 1mixture has 209 states, and the preliminary matching and screening includes the following steps: S101: The number of initial monophonic submodel states is 1024; S102: Use the Baum-Welch parameter reestimation method to cluster the initial single-tone sub-model states using the labeled crying training database, and reduce the number of valid states M from 1024 to 209; the specific process is as follows: Assume that S is the set of HMM model states, F is the training set, L(S) is the log-likelihood value, and assume that the bound states share the mean and variance of the Gaussian distribution, and that state binding does not affect the alignment of frames and states, then L(S)= in is the posterior probability that state S produces the feature, o f is the observed feature vector, which is a set of features extracted from the audio signal and is used to represent certain characteristics of the audio signal; u ( S ) is the mean vector of state S, Σ( S ) is the covariance matrix of state S; Assuming the output probability density function is Gaussian distribution, then L(S)= In the formula, n is the data dimension, and n=27; S103: Select L(S) As the first number of L(S) as the likelihood value of the valid state 0, compare L(S) To L(S) Middle and L(S) States with a difference of less than 5% are merged into the same state; S104: Repeat the above step S103 to obtain L(S) ~L(S) .

2. The low-resource infant crying detection method according to claim 1, characterized in that: The triphone model using 3mixture has 358 states, and the fine matching and screening includes the following steps: S301: the number of initial triphone model states is 1500; S302: using the Baum-Welch parameter reestimation method, clustering the initial three-phone model states using the labeled crying training database, and reducing the number of valid states T from 1500 to 358; S303: The specific process is as follows: Assume that S is the set of HMM model states, F is the training set, L(S) is the log-likelihood value, and assume that the bound states share the mean and variance of the Gaussian distribution, and that state binding does not affect the alignment of frames and states, then L(S)= in is the posterior probability of state S generating features. Assuming the output probability density function is Gaussian, then L(S)= In the formula, n is the data dimension, and n=27; S304: Select L(S) The first number of L(S) is used as the likelihood value of the valid state 0, and L(S) is compared. To L(S) Middle and L(S) States with a difference of less than 5% are merged into the same state; Repeat the above step S304 to obtain L(S) ~L(S) .

Images