Transformer-based language-independent gender recognition in noisy audio environments

Area of Science:

• Speech processing
• Machine learning
• Computational linguistics

Background:

• Speaker gender identification is crucial for voice recognition systems.
• Existing systems often exhibit gender bias due to imbalanced training data.
• Noise and language variations pose significant challenges in accurate gender detection.

Purpose of the Study:

• To develop an independent method for identifying speaker gender from audio clips in noisy environments.
• To compare the effectiveness of Mel-spectrograms versus Wav2Vec2 acoustic model emissions for gender identification.
• To address and mitigate gender bias in voice recognition by ensuring balanced datasets.

Main Methods:

• Audio clips were processed using Mel-spectrograms and Wav2Vec2 acoustic model emissions.
• Experiments were conducted across five languages: English, Arabic, Spanish, French, and Russian.
• A balanced dataset with equivalent male and female audio clips was used to mitigate bias.

Main Results:

• The Mel-spectrogram method demonstrated superior performance over the Wav2Vec2 transformer method.
• Spectrogram analysis achieved 99% accuracy for Russian, while Wav2Vec2 achieved 89%.
• Models trained on diverse languages and both noisy/silent conditions showed improved accuracy.

Conclusions:

• Mel-spectrograms offer a robust approach for acoustic gender detection, outperforming transformer-based models in this study.
• Balanced datasets are essential for reducing gender bias in speaker recognition systems.
• Future systems should incorporate multi-lingual and varied environmental data for enhanced reliability and fairness.