深層学習を利用した実環境下悲鳴検出システムの開発

使用深度学习在真实环境中开发尖叫检测系统

• 批准号: 19K04935
• 负责人: 早坂 昇
• 金额: $ 1.58万
• 依托单位: Osaka Electro-Communication University
• 依托单位国家: 日本
• 项目类别: Grant-in-Aid for Scientific Research (C)
• 财政年份: 2019
• 资助国家: 日本
• 起止时间: 2019-04-01 至 2024-03-31
• 项目状态: 已结题
• 来源: https://kaken.nii.ac.jp/en/grant/KAKENHI-PROJECT-19K04935/
• 关键词: 悲鳴検出; 悲鳴強調; 深層学習

2021年度まで，【課題1】の「劣悪な環境下における高い検出性能の達成」および【課題3】の「小型PCやモバイル端末への実装」に向けた演算量の削減に取り組み，一定の成果を達成することができた．しかしながら，【課題2】である「悲鳴とその類似音との識別精度向上」については十分な成果を上げることができていない状況であった．まず，【課題3】の実装状況については，Raspberry Pi 4 model Bへの実装が完了し，実時間でWave-U-Netを用いた悲鳴強調が可能であることを確認した．次に，【課題2】の悲鳴と類似音の識別精度向上については，2-stage Wave-U-Netについて検討した．これまでは雑音と悲鳴を分離するための深層ネットワーク（Wave-U-Net）を学習させ，当該ネットワークを利用し雑音重畳悲鳴から悲鳴のみを分離していたが，類似音は悲鳴と見なされ十分に分離することができていなかった．そこで，分離された悲鳴と類似音を分離するための深層ネットワーク（2-stage Wave-U-Net）を新たに学習させた．当該ネットワークに対し様々な類似音によるシミュレーションを実施したところ，平均-14dB程度となり高い類似音除去効果が得られた．一方，実運用した場合を考えると，類似音が常に混入するとは限らないため，同ネットワークに類似音が重畳されていない悲鳴を入力し，劣化の程度を測定した．その結果，約0.5dB程度の劣化に抑えられていることが確認され，2-stage Wave-U-Netの有効性が明らかとなった．2-stage Wave-U-Netの小型PCやモバイル端末への実装が未完了の状態であり，次年度の早期に実装し，実運用する予定である．

In the year of 2021, [title 1], "the performance of poor quality equipment in poor environmental conditions has reached the standard of" small PCs end-of-the-box equipment ". The calculation of the direction of the equipment is based on the calculation of the amount of equipment, and the results will be reduced to the performance of the poor environment. [topic 2] thanks for the warning that the accuracy of the error is up. On the results of the test, the Raspberry Pi 4 model B has finished loading, and it is possible to make sure that you are aware of the error when the Wave-U-Net is installed. [topic 2] the sound is similar to the accuracy of the sound, and the accuracy of the 2-stage Wave-U-Net sound is similar to that of the sound. The sound of the sound is separated from the sound, and the sound is separated from the sound of the sound. [topic 2] when the sound is similar to the accuracy of the sound, the sound of the sound is similar to the accuracy of the sound, and the sound of the sound is similar to that of the sound. The sound is very different from each other, the sound is very different, the sound is very different, the sound is as bad as the sound, the sound is like the sound of the sound, the sound is like the sound of the sound, the sound, the sound. On the one hand, the average-14dB level is similar to that of the other. On the one hand, the average level of sound is similar to that of others. On the other hand, on the one hand, the average level of sound is similar to that of others. On the one hand, the average level of sound is similar to that of noise. On the one hand, the average level of noise is similar to that of noise. On the one hand, the average level of sound is similar to that of noise. On the other hand, the average level of sound is similar to that of noise. On the other hand, on the one hand, the average level of sound is similar to that of the average level of sound. On the one hand, the average level of sound is similar to that of others. On the one hand, the average level of sound is similar to the level of noise. On the other hand, the average level of sound is similar to the level of noise. On the other hand, the average level of sound is similar to that of the average level of sound. On the one hand, the average level of sound is similar to that of 2-stage Wave-U-Net has the property to know that it is not finished at the end of the installation. In the next year, it will be installed in the early stage of the year. It will be predetermined by the software.

项目成果

Scream Enhancement using Wave-U-Net

使用 Wave-U-Net 增强尖叫

• 发表时间: 2021
• 作者: Riku Kasai;Noboru Hayasaka;Takuya Futagami;Yoshikazu Miyanaga
• 通讯作者: Yoshikazu Miyanaga