A new spectral transformation for speaker normalization

Pierre L. Dognin, Amro El-JaroudiInterspeech 2003 · OpenReview

TL;DR

This paper introduces a novel "Band-Pass Transform" (BPT) for speaker normalization, offering a more flexible way to warp the frequency axis of speech signals compared to existing methods. By leveraging the Bilinear Transformation with two degrees of freedom, BPT achieves better performance in Automatic Speech Recognition (ASR) tasks, demonstrating significant word error rate reductions.

What Problem Does This Paper Address?

Automatic Speech Recognition (ASR) systems struggle with variations in speech, particularly those arising from differences in speakers' vocal tract anatomy, such as vocal tract length (VTL). Traditional methods like linear frequency warping are too simplistic to capture the complex, non-linear way formants shift with VTL changes. While previous research has explored non-linear spectral transformations, this paper aims to introduce a more powerful and flexible technique that can better model these variations, ultimately improving ASR accuracy.

Methods

The core of this paper is the introduction of the "Band-Pass Transform" (BPT), a new spectral transformation for speaker normalization. BPT is derived from the Bilinear Transformation (BLT) framework but extends it by allowing the mapping of a prototype Band-Pass (BP) filter to a general BP filter, rather than being restricted to mapping a Low-Pass (LP) filter. This approach provides two degrees of freedom (controlled by parameters $\alpha$ and $\gamma$, which are related to $k$) for warping the frequency axis, enabling more complex transformations than previous one-parameter methods. The paper also proposes using the Nelder-Mead algorithm to estimate these two BPT parameters for each speaker, a method that avoids the need for derivative calculations.

Datasets

The experiments were conducted using the BYBLOS system on the Hub-5 English Dev01 and Eval01 benchmark datasets.

  • Hub-5 English Dev01: A 2-hour set with 48 speakers.
  • Hub-5 English Eval01: A larger 6-hour set with 120 speakers.

The paper also mentions using a 40-hour gender-balanced dataset for building gender-dependent acoustic models.

Key Insights and Ideas

  • Beyond Linear Warping: The paper highlights that vocal tract length variations don't translate to simple linear shifts in formants, necessitating more sophisticated non-linear transformations.
  • Two Degrees of Freedom for Flexibility: The BPT's ability to warp the frequency axis using two parameters ($\alpha$ and $\gamma$, related to $k$) offers significantly more flexibility in modeling speaker-specific spectral variations compared to single-parameter methods.
  • Physical Interpretation: The BPT is grounded in the physical transformation of a prototype Band-Pass filter into a general Band-Pass filter, providing an intuitive basis for its application in VTLN.
  • Nelder-Mead for Parameter Estimation: The use of the Nelder-Mead algorithm is a practical approach for estimating the two BPT parameters without requiring computationally expensive derivative calculations, making it suitable for real-world ASR systems.
  • BPT Outperforms Existing Methods: Experimental results consistently show that BPT achieves lower word error rates than linear warping, Eide VTLN, and 1st-BLT VTLN.

Key Results

  • Performance Improvement: BPT demonstrated superior performance across both Dev01 and Eval01 test sets.
  • On the Hub-5 Dev01 set, BPT achieved a gain of 1.32% absolute in word error rate reduction.
  • On the larger Hub-5 Eval01 set, BPT provided a gain of 1.13% absolute for the n-best rescoring method and 1.01% absolute for the 1-best rescoring method.
  • Parameter Estimation: The Nelder-Mead algorithm, when used with n-best or 1-best rescoring, provided parameter estimates that were close to those obtained from an "Oracle" grid search, indicating its effectiveness.
  • Comparison to Baselines: For speaker sw04537A, BPT achieved an absolute gain of 8.88% over the no-VTLN baseline. For speaker sw20316A, BPT provided a 3.45% absolute gain.

Limitations and Future Work

  • Jacobian Ignored: The paper acknowledges that the Jacobian of the non-linear transformation from unwarped to warped features was not taken into account in the current results. This could potentially affect the accuracy of the parameter estimation.
  • Comparative Studies: The authors suggest future work to conduct more extensive comparative studies of BPT against various VTLN methods on full test sets like Eval01.
  • Improving NM Procedure: Further research is planned to refine the Nelder-Mead procedure, specifically by incorporating the contribution of the transformation's Jacobian.

Summary updated

Abstract

This paper proposes a new spectral transformation for speaker normalization. We use the Bilinear Transformation (BLT) to introduce a new frequency warping resulting from a mapping of a prototype Band-Pass (BP) filter into a general BP filter. This new transformation called "Band-Pass Transform" (BPT) offers two degrees of freedom enabling complex warpings of the frequency axis and different from previous works with BLT. A procedure based on the Nelder-Mead algorithm is proposed to estimate the BPT parameters. Our experimental results include a detailed study of the performance of the BPT compared to other VTLN methods for a subset of speakers and results on large test sets. BPT performs better than other VTLN methods and offers a gain of 1.13% absolute on Hub-5 English Eval01 set.
BibTeX
@inproceedings{dognin2003new,
  author = {Pierre L. Dognin and Amro El-Jaroudi},
  title = {{A new spectral transformation for speaker normalization}},
  booktitle = {Interspeech},
  year = {2003},
  url = {https://openreview.net/forum?id=ONzNG6wanN},
}
Figure