Gender and Age Estimation Methods Based on Speech Using Deep Neural Networks.

Area of Science:

• Speech processing
• Machine learning
• Artificial intelligence

Background:

• Speech signals contain rich speaker information like gender and age.
• Automatic analysis of these features is crucial for various applications.
• Deep Neural Network (DNN)-based approaches have shown promise in speech analysis.

Purpose of the Study:

• To explore and develop an automatic system for speaker gender classification and age estimation using speech signals.
• To evaluate the effectiveness of different DNN embedder architectures (x-vector, d-vector).
• To investigate the impact of transfer learning on system performance.

Main Methods:

• Utilized Deep Neural Network (DNN) embedder architectures, specifically x-vector and d-vector.
• Employed a transfer learning strategy: pre-training on speaker recognition (Vox-Celeb1) and fine-tuning for joint age/gender tasks.
• Evaluated performance on the TIMIT dataset.

Main Results:

• Achieved new state-of-the-art results for age estimation.
• Reported Mean Absolute Error (MAE) of 5.12 years (male) and 5.29 years (female).
• Reported Root Mean Square Error (RMSE) of 7.24 years (male) and 8.12 years (female), with 99.60% gender recognition accuracy.

Conclusions:

• The proposed system effectively performs speaker gender classification and age estimation.
• Transfer learning significantly enhances performance for these tasks.
• DNN-based embedders show strong potential for analyzing speaker characteristics from speech.