Sound event detection method and device and readable storage medium

Abstract

The embodiment of the invention provides a sound event detection method and device and a readable storage medium. The method comprises the following steps: performing voice detection on original audio; when the voice is detected, sampling the voice; inputting the sampling points into a feature extraction module of a deep learning neural network in a streaming manner for feature extraction; inputting the extracted features into a global average pooling module of a deep learning neural network for global average pooling processing to obtain global average pooling features; and respectively inputting each global average pooling feature into a full connection layer of a deep learning neural network to carry out sound event detection so as to obtain the category of the sound event. According to the embodiment of the invention, the calculation burden of the NPU is reduced, the calculation resources are saved, the occupation of the cache is reduced, and the real-time performance of sound event detection is improved.

Description

technical field [0001] The present invention relates to the technical field of audio processing, and in particular, to a sound event detection method, device, readable storage medium and computer program product. Background technique [0002] The task of Sound Event Detection (SED) involves classifying sound events from real-life environments, such as crying babies, people walking, and dogs barking. [0003] Usually, a deep learning neural network is used for sound event detection. The specific process is: firstly detect the voice in the input original audio, and if the voice is detected, then perform the sound event detection after sampling the voice, specifically: every first time length to the depth The learning neural network inputs the sampled speech of the second duration, wherein the first duration is less than the second duration, and the deep learning neural network outputs the sound event detection result of the speech, such as baby crying, human walking or dog bar...

AI Technical Summary

Problems solved by technology

[0005] 1. A large part of the speech processed by the deep learning neural network is repetitive, and the calculation of repeated speech will bring unnecessary computational burden to the NPU (Neural Processing Unit, neural network processor) running the deep learning neural network , consumes resources, and the NPU computing power and resources of edge-end AI (artificial intelligence) devices are limited

For example: when the first duration is 100ms (milliseconds) and the second duration is 2s (seconds), that is, every 100ms, input 2s of speech to the deep learning neural network, specifically: first input the speech of 0-2s into the deep learning neural network Network, after an interval of 100ms, input the voice of 0.1-2.1s into the deep learning neural network, and so on. In this way, the voice of 1.9s will be repeated twice adjacent to the voice input into the deep learning neural network, which increases the NPU The computational burden is large

[0006] 2. The system needs to cache the voice for at least the second duration. For some edge AI devices such as DDR less, the cache will increase the cost of the equipment

[0007] 3. The deep learning neural network must wait for the second-long voice to be processed before it can be processed. In this way, there will be a certain delay in the detection of sound events, and the real-time performance will decline.

Images