Language Accent Detection with CNN Using Sparse Data from a Crowd-Sourced Speech Archive

The problem of accent recognition has received a lot attention with the development Automatic Speech Recognition (ASR) systems. crux is that conventional acoustic language models adapted to fit standard corpora are unable satisfy requirements for accented speech. In this research, we contribute task group up nine European accents in English and try provide some evidence favor specific hyperparameter choices neural network together search best input speech signal parameters ameliorate baseline accuracy. Specifically, used CNN-based model trained on audio features extracted from Accent Archive dataset, which crowd-sourced collection recordings. We show harnessing time–frequency energy (such as spectrogram, chromogram, spectral centroid, rolloff, fundamental frequency) Mel-frequency cepstral coefficients (MFCC) may increase accuracy classification compared feature sets MFCC and/or raw spectrograms. Our experiments demonstrate most impact brought about by amplitude mel-spectrograms linear scale fed into model. Amplitude scale, correlates energy, allow produce state-of-the-art results brings Germanic, Romance Slavic ranged 0.964 0.987; thus, outperforming existing classifying use Archive. also investigated how rhythm affects Based our preliminary experiments, recordings their original form (i.e., all pauses preserved) other experiments.