Automatic Recognition of Artificial Reverberation Settings in Speech Recordings
2019
artykuł naukowy
angielski
- artificial reverberation
- machine learning
- digital audio signal processing
EN The aim of this study is to create the method for automatic recognition of artificial reverberation settings extracted from a reference speech recordings. The proposed method employs machine-learning techniques to support the sound engineer in finding the ideal settings for artificial reverberation plugin available at a given Digital Audio Workstation (DAW), i.e. Gaussian Mixture Model (GMM) approach and deep Convolutional Neural Network (CNN) VGG13, which is a novel approach. Training set and data set are 1885 speech signals selected from a EMIME Bilingual Database which were processed with 66 artificial reverberation presets selected from Semantic Audio Labs’s SAFE Reverb plugin database. Performance of the proposed automatic recognition method was evaluated using similarity measures between features of reference and analysed speech recordings. Evaluation procedure showed that a classical GMM approach gives 43.8% of recognition accuracy while proposed method with VGG13 deep CNN gives 99.94% of accuracy.
2019125-1 - 2019125-8
CC BY (uznanie autorstwa)
publiczny
40
70