CNN을 위한 사투리 음성 데이터의 Spectrogram 이미지 변환 적용의 POC 검증

초록

In essence, audio data is time-series data. Therefore, for audio classification, time-series algorithms such as ARIMA (Autoregressive integrated moving average), ES (Exponential smoothing), or RNN (Recurrent neural network) in machine learning terms are commonly employed. Another method is to use a spectrogram as an image representing audio data rather than a time series number array as the input to the CNN(Convolutional neural network) learning process instead of RNN. In this paper, we use spectrogram analysis and mel-spectrogram analysis images of audio data as input data, using the CNN technique instead of RNN for time series data analysis, and we propose a CNN model to extract patterns of dialect audio. In addition, the feasibility was verified by performing a POC(Proof of concept) using a Python-based prototype of the proposed model.

키워드