Meta-Learning for Short Utterance Speaker Recognition with Imbalance Length Pairs

1. Introduction

• The work targets short-utterance recognition because practical test utterances can be only 1 to 5 seconds, while existing approaches have not substantially improved realistic scenarios.The introduction situates this challenge alongside task-specific feature extraction and frame-aggregation approaches.Source: §1 · p. 1
• Episodes pair long support utterances with variable-length short queries, unlike vanilla training on same-length utterances.Varying query lengths represents different practical situations and supports a more length-robust speaker recognition model.Source: §1 · p. 1
• Global classification trains each episode's support and query samples against all training classes, complementing episodic classification.This is intended to make short-utterance embeddings discriminative against other speakers while matching their own long utterances.Source: §1 · p. 1
• The model achieves state-of-the-art results on short-utterance speaker verification and unseen speaker identification across VoxCeleb datasets.The implementation uses ResNet34, temporal average pooling, non-margin metric loss, and 40-dimensional log mel-filterbank inputs.Source: §1 · p. 1

2. Related Work

• Figure 1 visualizes vanilla training versus meta-learning with global classification using t-SNE embeddings for a 10-way 5-shot task.Each prototype averages support utterances of 5 seconds.Source: Figure 1 · p. 2
• DNN speaker-recognition systems extract frame-level features, aggregate them into utterance-level vectors, then classify them; angular-margin methods target more discriminative embeddings.The described systems use convolutional or recurrent feature extractors and pooling methods including ASP, SAP, LDE, and SPE.Source: §2 · pp. 1–2
• Few-shot speaker verification and unseen-speaker identification motivate metric-based meta-learning, including prior Prototypical Networks applications to speaker recognition.The paper differentiates its method by using imbalanced-length pairs and global classification over episode samples.Source: §2 · p. 2
• Short-utterance methods include attention-based NetVLAD/GhostVLAD pooling and TDV, which reports gains over GhostVLAD but performs relatively poorly on long utterances.Other cited approaches use knowledge distillation, generative adversarial networks, or angular-margin methods.Source: §2 · p. 2

3.1. Problem Definition

• Few enrollment examples can make conventional supervised learning prone to overfitting when recognizing short test utterances from unseen speakers.The stated test utterance may be as short as 5 seconds, while enrollment utterances may be longer.Source: §3.1 · p. 2

3.2. Meta-learning with imbalance length pairs

• The method targets a length mismatch that creates a trade-off: training only on short segments performs relatively poorly on long speech.The paper motivates variable-length queries as a way to address test utterances of varying duration.Source: §3.2 · p. 2
• Episodes pair longer support utterances with variable-length short queries, and query examples are trained toward prototypes formed by averaging support examples.The query utterances are described as 1 to 2 seconds, shorter than the support utterances.Source: §3.2 · pp. 2–3

3.3. Global classification

• Global classification applies to both support and query samples, and the final objective combines its loss with the episode loss.The combined objective is described as matching imbalanced-length pairs while classifying them over the whole training-class set.Source: §3.3 · p. 3

4.1. Dataset

• VoxCeleb1 and VoxCeleb2 contain 1,251 and 5,994 disjoint speakers; verification is measured with EER and minDCF, and identification reports average accuracy over 1,000 episodes.Source: §4.1 · p. 3

4.2. Experiment setting

• Training uses 40-dimensional log mel-filterbank features and 1-shot 100-way episodes; short-verification tests use full-duration enrollment with test crops of 1, 2, or 5 seconds.Source: §4.2 · p. 3, §4.4 · p. 4

4.3. Speaker verification for full utterance

• On VoxCeleb1, the proposed model outperforms previous state-of-the-art models, including a TDNN with attentive statistic pooling, and it consistently beats baselines on VoxCeleb2.Source: §4.3 · p. 3

4.4. Speaker verification for short utterance

• On 1–2-second verification, the model beats baseline models by a significantly large margin, while the combined scheme is best in the VoxCeleb1 ablation.Source: Table 2 · p. 4, §4.4 · p. 4
• Imbalance-length pairs outperform equal-length and fixed long-short pairs, while global classification of support and query samples targets more discriminative embeddings.Source: §4.4 · p. 4

4.5. Unseen speaker identification

• In unseen-speaker identification, the proposed method outperforms vanilla training in every setting, with the performance gap increasing as speaker count grows.Source: Table 4 · p. 4, §4.5 · p. 4

• The proposed meta-learning scheme combines different-length support and query utterances with global classification to obtain discriminative embeddings, achieving state-of-the-art performance on short-utterance speaker recognition.Source: §5 · p. 4