A baby crying detection method integrating GMM model and HMM model

Abstract

The present invention relates to the field of Internet technology, and discloses a method for detecting infant crying by integrating a GMM model and an HMM model, comprising the following steps: establishing a short crying data model C1; establishing a medium crying data model C2; establishing a long crying data model C3: performing recognition training tests on each of the crying data models obtained above for multiple infant crying sounds, and obtaining the corresponding HMM model through training; establishing a short noise data model N1; establishing a medium noise data model N2: establishing a long noise data model N3: performing recognition training tests on each of the noise data models obtained above for multiple infant crying sounds, and obtaining the corresponding GMM model through training, and the GMM model can be fully utilized. The technical scheme of the present invention can significantly improve the accuracy of crying detection, and reduce the false alarm rate and the missed alarm rate.

Description

Technical Field

[0001] The present invention relates to the field of Internet technology, and in particular to a method for detecting baby crying by integrating a GMM model and an HMM model. Background Art

[0002] Baby cry detection technology has broad application prospects in the fields of education and accompanying care. At present, the industry mainly uses statistical model methods to achieve this. By establishing a baby cry database, training a GMM model or HMM model of baby cry, and then scoring it with the corresponding decoder to get the matching score, and judging whether it is a real baby cry by setting the confidence threshold. The main disadvantage of this type of method is that it only uses a single model structure (only using the GMM model structure, or only using the HMM model structure), and cannot give full play to the advantages of the two mainstream model structures according to the needs of specific application scenarios, while avoiding their shortcomings. Summary of the invention

[0003] The main purpose of the present invention is to propose a method for detecting infant crying sounds by integrating the GMM model and the HMM model, aiming to make full use of the characteristics of the GMM model that is independent of time series and the characteristics of the HMM model that is related to time series, establish an infant crying sound model and a noise model adapted thereto, improve the clustering and recognition performance of infant crying sound detection, improve the accuracy of crying sound detection, and reduce the false alarm rate and missed alarm rate.

[0004] To achieve the above object, the present invention proposes a method for detecting baby crying by integrating a GMM model and an HMM model, comprising the following steps:

[0005] Establish short cry data model C1;

[0006] Establish the crying sound data model C2;

[0007] Establish the long cry data model C3:

[0008] Each crying data model obtained above is used to perform recognition training tests on multiple infant crying sounds, and the corresponding HMM model is obtained through training;

[0009] Establish short noise data model N1;

[0010] Build the medium noise data model N2:

[0011] Build long noise data model N3:

[0012] Each noise data model obtained above is used for recognition training and testing on multiple baby crying sounds, and its corresponding GMM model is obtained through training.

[0013] Furthermore, the establishment of the short cry data model C1 comprises the following steps:

[0014] (1) For the pure crying sounds after removing silence, 5000 short-duration infant crying sounds with effective speech duration between 0.5 seconds and 1 second were selected, and the corresponding GMM model C1_G was obtained through training;

[0015] (2) From the 5000 crying sounds in the previous step, randomly select 2000 short-duration infant crying sounds with an overall speech duration of 0.5 seconds to 1 second, and obtain their corresponding HMM model C1_H through training.

[0016] Furthermore, the establishment of the crying data model C2 comprises the following steps:

[0017] (1) For pure crying sounds after eliminating silence, 5000 medium-length infant crying sounds with effective speech duration between 1 second and 2 seconds were selected, and the corresponding GMM model C2_G ​​was obtained through training;

[0018] (2) Randomly select 2000 short-duration infant crying sounds with an overall speech duration of 1 second to 2 seconds from the 5000 crying sounds in the previous step, and obtain their corresponding HMM model C2_H through training.

[0019] Furthermore, the establishment of the long cry data model C3 comprises the following steps:

[0020] (1) For pure crying sounds after eliminating silence, 5000 short-duration infant crying sounds with effective speech duration between 2 seconds and 3 seconds were selected, and the corresponding GMM model C3_G was obtained through training;

[0021] (2) Randomly select 2000 short-duration infant crying sounds with an overall speech duration of 2 seconds to 3 seconds from the 5000 crying sounds in the previous step, and obtain their corresponding HMM model C3_H through training.

[0022] Furthermore, the establishment of the short noise data model N1 comprises the following steps:

[0023] (1) Mix the existing M-type noise library sound data in proportion;

[0024] (2) After removing the silent sounds, we selected 5,000 short-duration noise speech with effective speech duration between 0.5 seconds and 1 second, and obtained the corresponding GMM model N1_G through training;

[0025] (3) Randomly select 2000 short-duration noisy speech pieces with an overall speech duration of 0.5 seconds to 1 second from the 5000 noisy speech pieces in the previous step, and obtain their corresponding HMM model N1_H through training.

[0026] Furthermore, the establishment of the noise data model N2 comprises the following steps:

[0027] (1) Mix the existing M-type noise library sound data in proportion;

[0028] (2) After removing the silent sounds, we selected 5,000 medium-length noise speech with effective speech duration between 1 second and 2 seconds, and obtained the corresponding GMM model N2_G through training;

[0029] (3) Randomly select 2000 medium-length noisy speech with an overall speech duration of 1 second to 2 seconds from the 5000 noisy speech in the previous step, and obtain their corresponding HMM model N2_H through training.

[0030] Furthermore, the establishment of the long noise data model N3 comprises the following steps:

[0031] (1) Mix the existing M-type noise library sound data in proportion;

[0032] (2) After removing the silent sounds, we selected 5,000 long noise speech with effective speech duration between 2 seconds and 3 seconds, and obtained the corresponding GMM model N3_G through training;

[0033] (3) Randomly select 2000 long noisy speech pieces with an overall speech duration of 2 seconds to 3 seconds from the 5000 noisy speech pieces in the previous step, and obtain their corresponding HMM model N3_H through training.

[0034] Furthermore, each crying data model obtained above is subjected to recognition training test for 5000 infant crying sounds in its respective training set: if the recognition rate is ≥97%, the crying model is retained; if the recognition rate is lower than 97%, the model is discarded, and 2000 crying voices are reassembled from the corresponding original 5000 crying voice database and repeated training is performed until the recognition rate meets the above requirements.

[0035] Furthermore, each noise data model obtained above is subjected to recognition training and testing for multiple baby crying sounds, and its corresponding GMM model is obtained through training: if the recognition rate is ≥97%, the noise model is retained; if the recognition rate is lower than 97%, the model is discarded, and 2000 noise voices are reassembled from the corresponding original 5000 noise database and repeated training is performed until the recognition rate meets the above requirements.

[0036] The technical solution of the present invention has the following beneficial effects: the technical solution of the present invention can make full use of the time-series-independent characteristics of the GMM model and the time-series-dependent characteristics of the HMM model, establish a baby crying model and a noise model adapted thereto, achieve the improvement of baby crying detection clustering and recognition performance, and can significantly improve the accuracy of crying detection and reduce the false alarm rate and missed alarm rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be 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 the structures shown in these drawings without paying creative work.

[0038] Figure 1 A schematic diagram of the overall framework process structure of a method for detecting baby crying sounds that integrates a GMM model and an HMM model proposed in the present invention;

[0039] Figure 2 A GMM model and HMM model fusion flow chart for a baby crying database of a baby crying detection method that integrates a GMM model and an HMM model proposed by the present invention;

[0040] Figure 3 The present invention provides a flow chart of the fusion of the GMM model and the HMM model for a baby crying detection method that integrates the GMM model and the HMM model for a noise database.

[0041] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] 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 components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0044] In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the fact that ordinary technicians in the field can implement it. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0045] The present invention proposes a method for detecting infant crying by integrating a GMM model and an HMM model.

[0046] like Figures 1 to 3 As shown, in one embodiment of the present invention, the method for detecting a baby crying sound by integrating the GMM model and the HMM model comprises the following steps:

[0047] S100: Establishing a short crying sound data model C1;

[0048] S200: establishing a crying sound data model C2;

[0049] S300: Establishing long crying data model C3:

[0050] S400: Performing recognition training tests on each crying data model obtained above for multiple infant crying sounds, and obtaining the corresponding HMM model through training;

[0051] S500: Establishing a short noise data model N1;

[0052] S600: Establishing a medium noise data model N2:

[0053] S700: Establishing long noise data model N3:

[0054] S800: Performing recognition training tests on each noise data model obtained above for a plurality of baby crying sounds, and obtaining the corresponding GMM model through training.

[0055] Specifically, the establishment of the short cry data model C1 includes the following steps:

[0056] (1) For the pure crying sounds after removing silence, 5000 short-duration infant crying sounds with effective speech duration between 0.5 seconds and 1 second were selected, and the corresponding GMM model C1_G was obtained through training;

[0057] (2) From the 5000 crying sounds in the previous step, randomly select 2000 short-duration infant crying sounds with an overall speech duration of 0.5 seconds to 1 second, and obtain their corresponding HMM model C1_H through training.

[0058] Specifically, the establishment of the crying data model C2 includes the following steps:

[0059] (1) For pure crying sounds after eliminating silence, 5000 medium-length infant crying sounds with effective speech duration between 1 second and 2 seconds were selected, and the corresponding GMM model C2_G ​​was obtained through training;

[0060] (2) Randomly select 2000 short-duration infant crying sounds with an overall speech duration of 1 second to 2 seconds from the 5000 crying sounds in the previous step, and obtain their corresponding HMM model C2_H through training.

[0061] Specifically, the establishment of the long cry data model C3 includes the following steps:

[0062] (1) For pure crying sounds after eliminating silence, 5000 short-duration infant crying sounds with effective speech duration between 2 seconds and 3 seconds were selected, and the corresponding GMM model C3_G was obtained through training;

[0063] (2) Randomly select 2000 short-duration infant crying sounds with an overall speech duration of 2 seconds to 3 seconds from the 5000 crying sounds in the previous step, and obtain their corresponding HMM model C3_H through training.

[0064] Specifically, the establishment of the short noise data model N1 includes the following steps:

[0065] (1) Mix the existing M-type noise library sound data in proportion;

[0066] (2) After removing the silent sounds, we selected 5,000 short-duration noise speech with effective speech duration between 0.5 seconds and 1 second, and obtained the corresponding GMM model N1_G through training;

[0067] (3) Randomly select 2000 short-duration noisy speech pieces with an overall speech duration of 0.5 seconds to 1 second from the 5000 noisy speech pieces in the previous step, and obtain their corresponding HMM model N1_H through training.

[0068] Specifically, the establishment of the noise data model N2 includes the following steps:

[0069] (1) Mix the existing M-type noise library sound data in proportion;

[0070] (2) After removing the silent sounds, we selected 5,000 medium-length noise speech with effective speech duration between 1 second and 2 seconds, and obtained the corresponding GMM model N2_G through training;

[0071] (3) Randomly select 2000 medium-length noisy speech with an overall speech duration of 1 second to 2 seconds from the 5000 noisy speech in the previous step, and obtain their corresponding HMM model N2_H through training.

[0072] Specifically, the establishment of the long noise data model N3 includes the following steps:

[0073] (1) Mix the existing M-type noise library sound data in proportion;

[0074] (2) After removing the silent sounds, we selected 5,000 long noise speech with effective speech duration between 2 seconds and 3 seconds, and obtained the corresponding GMM model N3_G through training;

[0075] (3) Randomly select 2000 long noisy speech pieces with an overall speech duration of 2 seconds to 3 seconds from the 5000 noisy speech pieces in the previous step, and obtain their corresponding HMM model N3_H through training.

[0076] Specifically, each crying data model obtained above is subjected to recognition training test for 5000 infant crying sounds in its respective training set: if the recognition rate is ≥97%, the crying model is retained; if the recognition rate is lower than 97%, the model is discarded, and 2000 crying voices are reassembled from the corresponding original 5000 crying voice database and repeated training is performed until the recognition rate meets the above requirements.

[0077] Specifically, each noise data model obtained above is subjected to recognition training and testing for multiple baby crying sounds, and its corresponding GMM model is obtained through training: if the recognition rate is ≥97%, the noise model is retained; if the recognition rate is lower than 97%, the model is discarded, and 2000 noise voices are reassembled from the corresponding original 5000 noise databases and repeated training is performed until the recognition rate meets the above requirements.

[0078] During specific identification, if the highest score of the viterbi decoding of the sample and the three baby cry models C1_G / C2_G / C3_G and the three noise models N1_G / N2_G / N3_G (must correspond to any one of C1 / C2 / C3) is greater than the GMM threshold TG, and the highest score of the viterbi decoding of the three baby cry models C1_H / C2_H / C3_H and the three noise models N1_H / N2_H / N3_H (must correspond to any one of C1 / C2 / C3) is greater than the HMM threshold TH, then the sample can be determined as a baby cry, otherwise it is determined as noise;

[0079] Specifically, the present invention proposes a new baby crying detection method that integrates the GMM model and the HMM model, which fully utilizes the time-series-independent characteristics of the GMM model and the time-series-dependent characteristics of the HMM model, establishes a baby crying model and a noise model adapted to it, and improves the clustering and recognition performance of baby crying detection, which can significantly improve the accuracy of crying detection and reduce the false alarm rate and missed alarm rate.

[0080] 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 changes made by using the contents of the present invention specification and drawings under the inventive concept of the present invention, or directly / indirectly applied in other related technical fields are included in the patent protection scope of the present invention.

Claims

1. A method for detecting baby crying by integrating a GMM model and an HMM model, characterized in that: The steps include: Establish a short crying data model C1, extract the features of short crying and train the corresponding GMM model C1_G and HMM model C1_H; Establish the medium-duration crying data model C2, extract the characteristics of medium-duration crying and train the corresponding GMM model C2_G ​​and HMM model C2_H; Establish a long crying data model C3, extract the characteristics of long crying and train the corresponding GMM model C3_G and HMM model C3_H; Establish a short noise data model N1, extract the characteristics of short-term noise and train the corresponding GMM model N1_G and HMM model N1_H; Establish a medium noise data model N2, extract the features of medium-duration noise and train the corresponding GMM model N2_G and HMM model N2_H; Establish a long noise data model N3, extract the characteristics of long-term noise and train the corresponding GMM model N3_G and HMM model N3_H; Each crying data model and noise data model obtained above is used for recognition training and testing on multiple samples of baby crying sounds. The sample has the highest Viterbi decoding score with the three baby crying models C1_G / C2_G / C3_G and the three noise models N1_G / N2_G / N3_G, which must correspond to any one of C1_G / C2_G / C3_G and be greater than the GMM threshold TG. At the same time, the sample has the highest Viterbi decoding score with the three baby crying models C1_H / C2_H / C3_H and the three noise models N1_H / N2_H / N3_H, which must correspond to any one of C1_H / C2_H / C3_H and be greater than the HMM threshold TH. In this case, the sample can be determined as a baby crying sound, otherwise it is determined as noise.

2. The method for detecting baby crying by integrating the GMM model and the HMM model according to claim 1, characterized in that: The establishment of the short cry data model C1 comprises the following steps: (1) For the pure crying sounds after removing silence, 5000 short-duration infant crying sounds with effective speech duration between 0.5 seconds and 1 second were selected, and the corresponding GMM model C1_G was obtained through training; (2) From the 5000 crying sounds in the previous step, randomly select 2000 short-duration infant crying sounds with an overall speech duration of 0.5 seconds to 1 second, and obtain their corresponding HMM model C1_H through training.

3. The method for detecting baby crying by integrating the GMM model and the HMM model according to claim 1, characterized in that: The establishment of the crying data model C2 comprises the following steps: (1) For pure crying sounds after eliminating silence, 5000 medium-length infant crying sounds with effective speech duration between 1 second and 2 seconds were selected, and the corresponding GMM model C2_G ​​was obtained through training; (2) Randomly select 2000 short-duration infant crying sounds with an overall speech duration of 1 second to 2 seconds from the 5000 crying sounds in the previous step, and obtain their corresponding HMM model C2_H through training.

4. The method for detecting baby crying by integrating the GMM model and the HMM model according to claim 1, characterized in that: The establishment of the long cry data model C3 comprises the following steps: (1) For pure crying sounds after eliminating silence, 5000 short-duration infant crying sounds with effective speech duration between 2 seconds and 3 seconds were selected, and the corresponding GMM model C3_G was obtained through training; (2) Randomly select 2000 short-duration infant crying sounds with an overall speech duration of 2 seconds to 3 seconds from the 5000 crying sounds in the previous step, and obtain their corresponding HMM model C3_H through training.

5. The method for detecting baby crying by integrating the GMM model and the HMM model according to claim 1, characterized in that: The establishment of the short noise data model N1 comprises the following steps: (1) Mix the existing M-type noise library sound data in proportion; (2) After removing the silent sounds, we selected 5,000 short-duration noise speech with effective speech duration between 0.5 seconds and 1 second, and obtained the corresponding GMM model N1_G through training; (3) Randomly select 2000 short-duration noisy speech pieces with an overall speech duration of 0.5 seconds to 1 second from the 5000 noisy speech pieces in the previous step, and obtain their corresponding HMM model N1_H through training.

6. The method for detecting baby crying by integrating the GMM model and the HMM model according to claim 1, characterized in that: The establishment of the noise data model N2 comprises the following steps: (1) Mix the existing M-type noise library sound data in proportion; (2) After removing the silent sounds, we selected 5,000 medium-length noise speech with effective speech duration between 1 second and 2 seconds, and obtained the corresponding GMM model N2_G through training; (3) Randomly select 2000 medium-length noisy speech with an overall speech duration of 1 second to 2 seconds from the 5000 noisy speech in the previous step, and obtain their corresponding HMM model N2_H through training.

7. The method for detecting baby crying by integrating the GMM model and the HMM model according to claim 1, characterized in that: The establishment of the long noise data model N3 comprises the following steps: (1) Mix the existing M-type noise library sound data in proportion; (2) After removing the silent sounds, we selected 5,000 long noise speech with effective speech duration between 2 seconds and 3 seconds, and obtained the corresponding GMM model N3_G through training; (3) Randomly select 2000 long noisy speech pieces with an overall speech duration of 2 seconds to 3 seconds from the 5000 noisy speech pieces in the previous step, and obtain their corresponding HMM model N3_H through training.

8. The method for detecting baby crying by integrating the GMM model and the HMM model according to claim 1, characterized in that: Each crying data model obtained above is subjected to recognition training test for 5000 infant crying sounds in its respective training set: if the recognition rate is ≥97%, the crying data model is retained; if the recognition rate is lower than 97%, the model is discarded, and 2000 crying voices are reassembled from the corresponding original 5000 crying voice database and repeated training is performed until the recognition rate meets the above requirements.

9. The method for detecting baby crying by integrating the GMM model and the HMM model according to claim 1, characterized in that: Each noise data model obtained above is subjected to recognition training and testing for multiple baby crying sounds, and its corresponding GMM model is obtained through training: if the recognition rate is ≥97%, the noise data model is retained; if the recognition rate is lower than 97%, the model is discarded, and 2000 noise voices are reassembled from the corresponding original 5000 noise databases and repeated training is performed until the recognition rate meets the above requirements.

Images