Exploring Speaker Age Estimation on Different Self-Supervised Learning Models

Abstract

Self-supervised learning (SSL) has played an important role in various tasks in the field of speech and audio processing. However, there is limited research on adapting these SSL models to predict the speaker's age and gender using speech signals. In this paper, we investigate seven SSL models, namely PASE+, NPC, wav2vec 2.0, XLSR, HuBERT, WavLM, and data2vec in the joint age estimation and gender classification task on the TIMIT corpus. Additionally, we also study the effect of using different hidden encoder layers within these models on the age estimation result. Furthermore, we evaluate how the performance of different SSL models varies in predicting the speaker's age under simulated noisy conditions. The simulated noisy speech is created by mixing the clean utterance from the TIMIT test set with random noises from the Music and Noise category of the MUSAN corpus on multiple levels of signal-to-noise ratio (SNR). Our findings confirm that a recent SSL model, namely WavLM can obtain better and more robust speech representation than wav2vec 2.0 SSL model used in the current state-of-the-art (SOTA) approach by achieving a 3.6% and 11.32% mean average error (MAE) reduction on the clean and 5dB SNR TIMIT test set.