End-to-End Speaker Age and Height Estimation using Attention Mechanism and Triplet Loss

Automatic age and height estimation of speakers using acoustic features is widely used for the purpose of human-computer interaction, forensics, etc. In this work, we study end-to-end framework for age and height estimation. We first propose a novel attention mechanism, named cross-attention. Different from conventionally used attention, which calculates context vector as the sum of attention only across timeframes, the proposed approach introduces a modified context vector which takes into account total attention across both time-frames and encoder units. We further propose using triplet loss to enhance the discriminative power of the encoder. We evaluate the Root Mean Square Error (RMSE) of pro-posed approaches on the TIMIT corpus. The proposed cross-attention outperforms the conventional counterpart for both age and height estimation while the triplet loss brings 8% relative improvement for age estimation. We obtain RMSE of 6.92/6.42 cm for male/female height estimation and 7.20/7.10 years for male/female age estimation which outperforms all the previous baselines achieving state-of-the-art performance for age estimation on TIMIT dataset. By tracking the attention weights allocated to different phones, we find that Vowel phones are most important while Stop and Fricative phones are least important for the estimation task.