Consideration of Varying Training Lengths for Short-Duration Speaker Verification

We present an efficient training scheme for speaker verification (SV) networks in short-duration speech input scenarios. We analyze the effects of varying training lengths on SV performance, with a particular focus on short utterances. Despite the high demand for short-duration SV in real-world applications, state-of-the-art SV systems have primarily been evaluated on long utterances, and little research has been conducted on short-duration SV. By considering the innate characteristics of SV architectures and the performance discrepancies associated with varying training data lengths, we propose a training scheme that accounts for varying length conditions. We categorize speaker characteristics as coarse-grained and fine-grained features and demonstrate that training models to learn both features can result in length-robust speaker embeddings. Our proposed training scheme improves model performance by 28.7% and 37.9% in terms of equal error rate on short-duration speech scenarios compared to baseline models.