Supervector LDA: a new approach to reduced-complexity i-vector language recognition

Alan McCree, Bengt BorgströmInterspeech 2012

TL;DR

This paper introduces Supervector LDA (SV-LDA), a novel method for reducing the complexity of i-vector based language recognition systems. By applying Linear Discriminant Analysis (LDA) directly in the supervector space, SV-LDA achieves performance comparable to higher-dimensional i-vectors but with significantly reduced computational cost during extraction.

What Problem Does This Paper Address?

Language recognition systems often rely on sophisticated statistical models like Gaussian Mixture Models (GMMs) and their subspace extensions, such as Joint Factor Analysis (JFA) and i-vectors. While i-vectors have proven effective, extracting them can be computationally intensive, especially for high-dimensional models. This paper addresses the challenge of reducing this computational complexity without sacrificing recognition accuracy. Specifically, it explores how to efficiently perform scoring and extract more compact i-vectors by leveraging the underlying statistical framework of GMM supervectors and subspace modeling.

Methods

The authors build upon their previous work on GMM subspace compensation, extending it to the supervector space. They demonstrate that full Gaussian supervector scoring can be efficiently performed in a low-dimensional subspace, akin to i-vector scoring. A key contribution is the introduction of Supervector LDA (SV-LDA). Unlike traditional LDA applied after i-vector extraction, SV-LDA integrates the discriminative criterion directly into the design of the total variability subspace. This is achieved by reformulating the LDA objective in terms of subspace covariance matrices, allowing for the construction of a lower-dimensional subspace that is optimized for discrimination. The paper also discusses hyperparameter estimation and the application of PCA for dimension reduction in the supervector space.

Datasets

The experiments were conducted using the NIST LRE09 Evaluation Set. This dataset includes conversational telephone speech (CTS) and broadcast news (BN) data, covering 23 language classes. Evaluation was performed on test segments of 30s, 10s, and 3s durations.

Key Insights and Ideas

  • Supervector scoring equivalence: The paper shows that under certain subspace assumptions, full Gaussian supervector scoring is equivalent to i-vector scoring, especially for long-duration speech segments. This highlights a deeper statistical connection between these methods.
  • Efficient subspace scoring: The authors demonstrate that complex GMM supervector scoring can be performed efficiently in a low-dimensional subspace, reducing computational load.
  • SV-LDA for discriminative dimension reduction: The core innovation, SV-LDA, allows for discriminative dimension reduction during i-vector extraction. This is a significant improvement over applying LDA post-extraction, as it reduces the complexity of the entire system, not just the scoring phase.
  • Hyperparameter estimation from i-vectors: The relationships derived allow for hyperparameter estimation using i-vector covariance matrices, simplifying the training process.

Key Results

  • Performance parity with reduced complexity: The proposed Supervector LDA (SV-LDA) method achieved performance comparable to higher-dimensional i-vector systems. For instance, a 100-dimensional SV-LDA system matched the performance of a 600-dimensional i-vector system.
  • Significant speedup: The 100-dimensional SV-LDA system reduced the CPU time required for i-vector extraction from approximately 3 seconds to just 0.1 seconds per utterance, a substantial improvement in efficiency.
  • Comparison of dimension reduction techniques: Table 2 shows that while SV-PCA offers some complexity reduction, it leads to performance degradation. SV-LDA, however, demonstrates much better performance retention at lower dimensions, with the 100-dimensional system being particularly effective.
  • Subspace full Gaussian scoring vs. i-vector scoring: Table 1 indicates that subspace full Gaussian scoring provides equivalent performance to i-vector scoring, with only a slight degradation for shorter test durations.

Limitations and Future Work

  • Vocal tract normalization: The authors note that their performance is not quite as good as some prior work, attributing this to the lack of vocal tract normalization. This suggests a potential avenue for future improvement.
  • Generalizability of SV-LDA: While SV-LDA shows promise, further investigation into its generalizability across different datasets and acoustic conditions would be beneficial.
  • Exploration of other discriminative criteria: The paper focuses on LDA. Future work could explore other discriminative criteria for designing the total variability subspace.

Summary updated

BibTeX
@inproceedings{mccree2012supervector,
  author = {Alan McCree and Bengt Borgström},
  title = {{Supervector LDA: a new approach to reduced-complexity i-vector language recognition}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure