Estimation of speaker characteristics from speech signal

Abstract

Estimating speaker attributes like age and height is a difficult task with sev eral applications in speech forensic analysis and potential applications in speaker verification and speaker adaptation techniques. We present a bi-encoder (Mixture of Experts (MoE) inspired) transformer mixture model for estimating speaker age and height in this project. For the extraction of specific male and female voice characteristic features, we suggest the use of two different transformer encoders, while making use of wav2vec 2.0 as a common feature extraction method. The bi encoder architecture is chosen due to the significant variances in male and female voice characteristics. This architecture increases the model’s generalizability by re ducing interference effects during the model training. We conduct our tests using the TIMIT corpus and find that our results on age estimation surpass the present state-of-the-art. For male and female age estimation, we obtain 5.54 years and 6.49 years as root mean squared error (RMSE), respectively. Further research into the relative impact of various phonetic sound kinds for speaker profiling reveals that vowel phonemes are the most distinctive for age estimate.