Reproduce “Analyzing And Improving Neural Speaker Embeddings for ASR” on OpenResearch

About this paper

Christoph Lüscher, Jingjing Xu, Mohammad Zeineldeen, Ralf Schlüter, Hermann Ney

Researchers at RWTH Aachen University and AppTek GmbH developed an improved pipeline for neural speaker embedding extraction, featuring an auxiliary reconstruction loss, enabling x-vectors and c-vectors to achieve performance on par with i-vectors for speaker adaptation in a state-of-the-art Conformer-based hybrid ASR system. This approach also leveraged a One Cycle Learning Rate schedule, reducing training time while maintaining high performance.

Key insights

• Neural speaker embeddings trained purely for speaker discrimination tend to discard acoustic and channel information beneficial for ASR adaptation.
• Incorporating an auxiliary reconstruction loss during neural embedding training effectively forces the network to retain ASR-relevant acoustic information, significantly improving adaptation performance.
• Temporal pooling strategies in neural embedding extraction, particularly attentive statistics pooling, are more effective for ASR when they capture statistical properties beyond pure speaker discrimination.

Summary by alphaXiv.