深度学习声纹识别_一种基于机器学习及深度学习的声纹降噪方法及系统与流程…

The invention belongs to the field of speech character recognition and voiceprint noise reduction, and relates to a voiceprint noise reduction method and system based on machine learning and deep learning.

The intelligence in the process of geoscience data collection is a basic link in the construction of geoscience big data, which occupies an important position in the actual process of geological production. Among them, the collection of survey data that geologists observe and describe geological objects in the field is an important part. In order to support geologists to quickly collect the observed geoscience data in the field, the previous geoscience data acquisition systems attach great importance to the efficiency and convenience of data input, but the traditional text keyboard input mode is inefficient and inoperable in the field. In order to improve the efficiency of data input, voice input and character recognition system are adopted to improve the efficiency of data acquisition. In use, it is found that when using the voice recording system in the field, there may be wind sound, rain sound, animal sound, etc., or there may be a large number of operating machine noise in the drilling site, mining site or oil and gas production environment. when these noises are mixed with human voices, the text recognition accuracy of the current speech recognition system will be greatly reduced. As a result, the accuracy of speech input and character recognition in field geoscience data acquisition is very low. Because of this problem, the current geoscience data acquisition system has low operation efficiency and poor availability in the special geological working environment.

In view of the above shortcomings, there is an urgent need for a speech input noise reduction technology to remove as much environmental noise as possible in the field geological work environment, so as to lay a foundation for improving the accuracy of later speech text recognition. The invention provides a voiceprint noise reduction method based on machine learning and deep learning, which is used to solve the technical problem that the background noise is large and the effective speech is difficult to accurately identify when collecting geoscience data in the field voice mode, which includes the following steps:

According to another aspect of the invention, in order to solve its technical problems, the invention provides a voiceprint noise reduction system based on machine learning and deep learning, which comprises the following modules:

Initialization module, which is used to obtain the audio of a specific person’s description of geological phenomena and geological knowledge in the field working environment.

The machine learning and deep learning processing recognition module is used to learn the speech audio acquired in the initialization module through machine learning and deep learning model respectively to identify and distinguish the ambient sounds mixed in the speech audio.

The speech filtering module is used to filter the recognized speech obtained by the machine learning and deep learning processing recognition module, to remove the ambient sound in the speech that does not belong to the specific person’s speech audio, and to obtain the speech that has been preliminarily screened.

The speech filtering judgment module is used to determine whether the filtered speech signal-to-noise ratio reaches the preset threshold; if not, jump to the voice filtering module; if so, continue with the voiceprint extraction module

The voiceprint contrast extraction module is used to compare and extract the speech obtained in the filter judgment module with the person’s voiceprint recognition model, and to retain the speech frequency and speech spectrum images consistent with the person’s voiceprint recognition model. the speech that does not accord with the voiceprint recognition model is removed, and the speech processed by voiceprint noise reduction is obtained.

Speech purity judgment module: used to judge whether the extracted human voiceprint purity reaches the preset threshold, if not, jump to the voiceprint extraction module, if yes, carry on the speech enhancement module

The speech enhancement module further enhances the voice after noise reduction obtained in the voiceprint extraction module.

The speech articulation judgment module is used to determine whether the purity of the extracted human voiceprint reaches the preset threshold, if not, jump to the speech enhancement module, if yes, continue the voice input module

The speech output module outputs the results obtained in the speech enhancement module to the speech character recognition system to complete the subsequent processing.

In the voiceprint noise reduction system based on machine learning and deep learning, the invention also includes before initializing the module:

A machine learning and deep learning model is established, specifically, the speech of a large number of people speaking in a noisy wild environment is obtained, and all the speech is converted into a spectrogram and imported into the computer. a large number of repeated training of machine learning and deep learning models are carried out, and machine learning and deep learning are used to distinguish the ambient sounds in human speech audio, so as to obtain a trained machine learning and deep learning model.

In the voiceprint noise reduction system based on machine learning and deep learning, the voiceprint recognition model of a specific speaker is established before the voiceprint extraction module, specifically, the voiceprint spectrogram is established with the existing voiceprint of the specific speaker, and after the feature extraction on the voiceprint spectrogram, the voiceprint recognition model that only belongs to that person can be established, and after repeated training of a large number of voiceprint data of that person. A voice print recognition model with high voiceprint recognition rate is obtained.

The invention adopts a voiceprint noise reduction method and system based on machine learning and deep learning, which can be targeted to solve the problem of low noise reduction rate after noise reduction by using common noise reduction technology or simple speech noise reduction algorithm in a specific variety of complex field geological environments with different high and low signal-to-noise ratios, so as to provide a good prerequisite for the speech character recognition process of geoscience data acquisition in the later stage. It provides technical guarantee for reducing the difficulty of speech character recognition in the process of geoscience data acquisition and greatly improving the accuracy of speech character recognition in the process of geoscience data acquisition.

The invention will be further explained in combination with the drawings and embodiments below, in the drawings:

In order to make the purpose, technical scheme and advantages of the invention more clear, the invention is further described in detail in combination with drawings and examples.

In the field, the speech recording system is used to convert the speech audio clips that specific people describe geological phenomena and geological knowledge in the field working environment into spectrograms through machine learning and deep learning model. aiming at the speech introduced into the model, the environmental murmur in the speech is recognized, and the environmental murmur is eliminated, that is, the ambient sound that does not belong to human voice in the speech segment is removed. Repeated cyclic processing to determine whether the noise filtered speech is qualified, specifically setting the signal-to-noise ratio in advance, when the preset signal-to-noise ratio is reached, then proceed to the next step, and if not, continue to carry out noise filtering.

The filtered voice is compared with the established voiceprint recognition model, and the flow chart is shown in figure 4. The input voiceprint is compared with the voiceprint recognition model, and the speech frequency and speech spectrum image consistent with the model are retained. The speech that does not conform to the model is eliminated and processed many times to determine whether the voiceprint extracted speech is pure or not, specifically as the preset voiceprint purity threshold. When the speech purity reaches the preset threshold, analyze whether there is any noise other than the person’s voice on the speech spectrogram, if so, continue the voiceprint noise reduction process, if not, the voice is processed by the voiceprint noise reduction.

According to the embodiment of the invention, the following modules are also included:

The initialization module is used to obtain the speech audio of the description of geological phenomena and geological knowledge of a specific person in the field working environment.

The machine learning and deep learning processing recognition module is used to learn the speech audio acquired in the initialization module through machine learning and deep learning model respectively to identify and distinguish the ambient sounds mixed in the speech audio.

The speech filtering module is used to filter the recognized speech obtained by the machine learning and deep learning processing recognition module, remove the ambient sound in the speech that does not belong to the human speech audio, and get the speech that has been screened preliminarily.

The speech filtering judgment module is used to determine whether the filtered speech signal-to-noise ratio reaches the preset threshold; if not, jump to the voice filtering module; if so, continue with the voiceprint extraction module

The voiceprint contrast extraction module is used to compare and extract the speech obtained in the filter judgment module with the person’s voiceprint recognition model, and to retain the speech frequency and speech spectrum images consistent with the person’s voiceprint recognition model. the speech that does not accord with the voiceprint recognition model is removed, and the speech processed by voiceprint noise reduction is obtained.

Speech purity judgment module: used to judge whether the extracted human voiceprint purity reaches the preset threshold, if not, jump to the voiceprint extraction module, if yes, carry on the speech enhancement module

The speech enhancement module further enhances the voice after noise reduction obtained in the voiceprint extraction module.

The speech articulation judgment module is used to determine whether the purity of the extracted human voiceprint reaches the preset threshold, if not, jump to the speech enhancement module, if yes, continue the voice input module

The speech output module outputs the results obtained in the speech enhancement module to the speech character recognition system to complete the subsequent processing.

In the voiceprint noise reduction system based on machine learning and deep learning, the invention also includes before initializing the module:

To establish a machine learning and deep learning model, specifically, to obtain the voice of a large number of human speech audio, convert all the speech into spectrograms and import them into the computer, and train the machine learning and deep learning models repeatedly. Machine learning and deep learning recognition to distinguish the ambient sound in human speech audio, so as to obtain a trained machine learning and deep learning model.

In the voiceprint noise reduction system based on machine learning and deep learning, the voiceprint recognition model of a specific speaker is established before the voiceprint extraction module, specifically, the voiceprint spectrogram is established with the existing voiceprint of the specific speaker, and after the feature extraction on the voiceprint spectrogram, the voiceprint recognition model that only belongs to that person can be established, and the voiceprint recognition model that only belongs to that person can be established after repeated training of a large number of voiceprint data. A voice print recognition model with high voiceprint recognition rate is obtained.

Above, the embodiments of the invention are described in combination with the drawings, but the invention is not limited to the above-mentioned specific embodiments, which are only schematic rather than restrictive. under the inspiration of the invention, ordinary technicians in the field can also make a number of improvements and deformations without departing from the scope protected by the purpose and claims of the invention. These are within the protection of the present invention.