Source-linked AI summary

Exploring the Encoding Layer and Loss Function in End-to-End Speaker and Language Recognition System

Weicheng Cai, Jinkun Chen, Ming Li

arXiv:1804.05160v1eess.AScs.LGcs.SD

TL;DR

The paper addresses how to build utterance-level speaker and language recognition systems that handle variable-length speech and produce discriminative representations for open-set verification. It develops a unified end-to-end framework that compares pooling and encoding layers and introduces discriminative loss functions. Experiments on VoxCeleb and NIST LRE 07 report significant improvements from suitable encoding layers and loss functions.

  • Problem

    Speaker and language recognition need robust utterance-level representations from variable-length speech, while open-set verification requires discriminative large-margin features.

  • Method

    The paper develops a unified end-to-end system using CNN feature extraction, temporal pooling or learnable dictionary encoding, and center or A-Softmax loss for open-set verification.

  • Results

    Suitable encoding layers and loss functions significantly improve end-to-end performance across VoxCeleb speaker recognition and NIST LRE 07 language recognition tasks.

  • Takeaways & Limitations

    Self-attentive pooling and learnable dictionary encoding outperform temporal average pooling, while discriminative losses generally outperform softmax loss.

  • Takeaways & Limitations

    The comparison focuses on systems that do not require additional transcribed speech data or an extra DNN acoustic model.

Abstract

from arXiv · show

In this paper, we explore the encoding/pooling layer and loss function in the end-to-end speaker and language recognition system. First, a unified and interpretable end-to-end system for both speaker and language recognition is developed. It accepts variable-length input and produces an utterance level result. In the end-to-end system, the encoding layer plays a role in aggregating the variable-length input sequence into an utterance level representation. Besides the basic temporal average pooling, we introduce a self-attentive pooling layer and a learnable dictionary encoding layer to get the utterance level representation. In terms of loss function for open-set speaker verification, to get more discriminative speaker embedding, center loss and angular softmax loss is introduced in the end-to-end system. Experimental results on Voxceleb and NIST LRE 07 datasets show that the performance of end-to-end learning system could be significantly improved by the proposed encoding layer and loss function.

1. Introduction

Language and speaker recognition require robust utterance-level representations from variable-length speech, with open-set verification additionally requiring discriminative feature spaces. The paper therefore develops a unified end-to-end system and investigates its encoding layers and loss functions.

  • Motivation: Language and text-independent speaker recognition are utterance-level sequence-to-one tasks that seek time-invariant representations from variable-length input.Unlike speech recognition, they retrieve information about an entire utterance rather than specific word content.
  • Conventional systems: The conventional pipeline uses local feature descriptors, a dictionary, and vector encoding to aggregate variable-length sequences into utterance-level vectors.Examples include Fbank, MFCC, PLP, SDC, bottleneck, PPP, and tandem features, with dictionaries learned by K-means, GMM, or DNNs.
  • End-to-end motivation: End-to-end learning offers an alternative to hand-crafted i-vector pipelines by placing a general encoding layer above a convolutional neural network.The learnable dictionary encoding layer connects conventional GMM Supervector processing with end-to-end neural networks.
  • Recognition settings: Closed-set identification is treated as classification, whereas open-set speaker verification is treated as metric learning requiring discriminative large-margin features.In open-set verification, testing identities are not restricted to identities seen during training.
  • Paper objective: The paper builds a unified end-to-end system for language and speaker recognition and focuses on improving it through encoding layers and loss functions.The pipeline includes input data, frame-level feature extraction, encoding, loss, and similarity-metric modules.

2. End-to-End System Overview

The end-to-end framework processes speech of arbitrary duration by extracting frame-level features and aggregating them into a fixed-dimensional utterance representation. A similarity metric is added specifically for open-set verification.

  • Framework: The framework accepts variable-length speech segments and produces an utterance-level result.This accommodates testing speech whose duration is not known in advance.
  • Output and verification: The utterance representation is passed through a fully connected layer and output layer, with an additional similarity metric for open-set verification.The similarity metric is specifically designated for the open-set verification task.
  • Frame-level feature extraction: A convolutional neural network extracts high-level local patterns from input features such as log mel-filterbank energies.The resulting frame-level representations retain temporal order before aggregation.
  • Encoding: The encoding layer aggregates temporally ordered frame representations into a fixed-dimensional utterance-level representation.This aggregation resolves the mismatch between variable-length input sequences and utterance-level outputs.

3. Encoding layer

The encoding layer converts variable-length frame sequences into fixed-dimensional utterance-level representations. The paper compares temporal average pooling, self-attentive pooling, and learnable dictionary encoding as alternative aggregation mechanisms.

  • Temporal average pooling: Temporal average pooling equally aggregates CNN-extracted features across time.It is the basic pooling approach described for end-to-end architectures.
  • Self-attentive pooling: Self-attentive pooling weights frames according to their importance for classification before aggregating them into an utterance representation.A learned context vector and softmax-derived weights identify informative frames.
  • Self-attentive pooling: The self-attention mechanism is self-contained and does not require extra guiding source information.This distinguishes the described architecture from an earlier attention-based recurrent approach that used language-category pretraining.
  • Learnable dictionary encoding: Learnable dictionary encoding combines dictionary learning and vector encoding into one end-to-end trainable layer.It imitates the GMM Supervector mechanism while learning directly from the loss function.
  • Learnable dictionary encoding: The LDE layer transforms a D × L temporal feature sequence into a length-independent D × C utterance-level representation.It assigns features to learnable dictionary components, aggregates residual vectors, and concatenates the resulting encoder outputs.

4. Loss function

The paper replaces explicitly separated front-end and back-end processing with unified end-to-end losses, addressing the need for discriminative embeddings in open-set speaker verification. It examines softmax, center, and angular softmax objectives for closed-set identification and open-set verification.

  • 4.1. Loss function for closed-set identification: Closed-set identification jointly learns the front-end feature extractor and back-end classifier under a unified softmax loss.The classifier uses the penultimate representation, target label, and final-layer weights and bias.
  • 4.2. Loss function for open-set verification: Open-set speaker verification requires embeddings that are both separable and discriminative because testing identities cannot all be pre-collected for training.Consequently, basic label-prediction softmax is not always applicable to this setting.
  • 4.2. Loss function for open-set verification: Pairwise contrastive and triplet losses formulate open-set verification as metric learning but require non-trivial pair or triplet mining.The mining procedure is described as time-consuming and performance-sensitive.
  • 4.2.1. Center loss: Center loss learns one center per class and penalizes distances between features and their corresponding centers while preserving inter-class separability through joint softmax supervision.The scalar λ balances center loss against softmax loss, and λ = 0 recovers conventional softmax.
  • 4.2.2. Angular Softmax loss: Angular softmax learns an angular margin with integer m controlling its size; m = 1 yields modified softmax, while m ≥ 2 imposes stronger angular classification requirements.Its geometric formulation produces an angular distance metric on a hypersphere and separates features by angular margins.

5. Experiments

Experiments evaluate unified end-to-end systems for speaker and language recognition on VoxCeleb and NIST LRE 07, comparing encoding layers, losses, and similarity metrics. The results favor LDE encoding and discriminative losses, with task-dependent gains over TAP and i-vector baselines.

  • Datasets and tasks: Experiments use VoxCeleb for speaker verification and identification, and NIST LRE 07 for closed-set language detection across 3-, 10-, and 30-second durations.VoxCeleb verification uses Cdet and EER; identification reports top-1 and top-5 accuracy, while LRE 07 uses Cavg and EER.
  • Overall comparison: End-to-end learning systems significantly outperform the conventional i-vector approach on both speaker-recognition and language-recognition tasks.The comparison covers the verification and identification results in Tables 2–4.
  • Encoding layers: Both SAP and LDE outperform TAP, while LDE also surpasses SAP across the reported experiments.These comparisons concern the encoding layer used to form utterance-level representations.
  • Losses and similarity metrics: Discriminative center loss and A-Softmax generally outperform softmax, whereas PLDA provides little or negative gain for end-to-end systems trained with discriminative losses.PLDA reduces error in the conventional i-vector system but can worsen end-to-end results.
  • Reported gains: 25%, 45%, and 63% relative error reductions are achieved by CNN-LDE versus CNN-TAP on NIST LRE 07 3s, 10s, and 30s tasks.CNN-LDE also achieves 12% relative error reduction for VoxCeleb speaker identification; LDE-A-Softmax verification reaches Cdet 0.441 and EER 4.56%.

6. Conclusions

The paper develops a unified, interpretable end-to-end system for speaker and language recognition that accepts variable-length input and produces utterance-level results. It improves performance by using suitable encoding layers and loss functions, including SAP, LDE, center loss, and A-Softmax.

  • Conclusions: A unified end-to-end system is developed for speaker and language recognition, accepting variable-length input and producing utterance-level results.The system investigates encoding layers and loss functions within one pipeline.
  • Conclusions: SAP and LDE replace basic temporal average pooling to produce utterance-level representations from variable-length sequences.The paper introduces both self-attentive pooling and learnable dictionary encoding.
  • Conclusions: Center loss and A-Softmax loss are introduced to obtain more discriminative speaker embeddings for open-set speaker verification.These losses address the speaker-embedding component of the end-to-end system.
  • Conclusions: Experimental results show that suitable encoding layers and loss functions significantly improve end-to-end system performance.The reported experiments use VoxCeleb and NIST LRE 07 datasets.
Loading 1804.05160v1…