ECAPA2: A Hybrid Neural Network Architecture and Training Strategy...

Jenthe/ECAPA2 · Hugging Face

1. Introduction

Model fusion

• TDNN 기반 모델의 kernel이 frequency range를 모두 span (모두 커버) 하기 때문에
• 2D-conv는 TDNN의 freqency-independent의 한정적인 capability를 완화해야함.

Proposal

• model interpretability analysis
  • Feature ablation analysis (impact of diffent kernel types)
  • ⇒ 이를 기반으로 ECAPA2 arch. 제안
• Model training stategy
  • overlapped speech와 short utt. 를 위함

2. Architectural Analysis

2.1. Input feature robustness

2.2 Effeictive Receptive Fields (ERFs)

2.3. Neuron conductance

• Ref: How Important Is a Neuron?

Neuron conductance is based on the Integrated Gradients [12] attribution method and assigns importance scores to input features by integrating the gradients of the output of the model with respect to the inputs along a path from a baseline to the desired input.

• Figure
• TDNN은 ERF에서 보여주듯 freq. dim에 대해서 uniform하고,