On enhancing feature sequence filtering with filter-bank energy transformation in speaker verification with telephone speech

Claudio Garretón, Néstor Becerra YomaInterspeech 2010

TL;DR

This paper introduces a novel method for improving speaker verification accuracy over telephone lines by filtering features in the "filter-bank energy" domain. This approach effectively reduces time-varying channel distortions, leading to significant reductions in error rates, especially when combined with existing techniques.

What Problem Does This Paper Address?

Speaker verification systems, particularly those used in real-world scenarios like telephone calls, struggle with "channel mismatch." This refers to the difference between the acoustic conditions under which a system is trained and the conditions it encounters during testing. Telephone lines introduce various distortions (e.g., varying microphones, transmission lines) that can significantly degrade performance. Traditional methods often assume channel distortions are static or slow-varying, but real-world telephone speech exhibits time-varying distortions that are harder to compensate for. This paper aims to address this by developing a feature enhancement technique that specifically targets these dynamic distortions without relying on strong assumptions about the channel.

Methods

The core of the proposed method is a filter-bank energy (LFBE) spectrum domain transformation. Instead of directly filtering the speech signal or its traditional features (like MFCCs), this technique first transforms the LFBE features into the frequency domain using a Discrete Fourier Transform (DFT). Then, a band-pass filter is applied in this spectral domain on a frame-by-frame basis. This process aims to attenuate spectral components that are more susceptible to channel distortion while preserving those that carry more speaker-specific information.

The parameters of this spectral filter (specifically, the cut-off frequencies) are determined using a Relative Importance Analysis (RIA). RIA leverages a discriminative function that measures the separability between different speakers and within the same speaker. By analyzing how this discriminative function changes when different spectral components are included or excluded, the method identifies which components are most crucial for distinguishing speakers and which are most affected by channel noise. This data-driven approach allows for a more targeted and effective filtering.

The paper also explores combining this novel spectral filtering with existing techniques like Cepstral Mean Normalization (CMN) and RASTA (RelAtive SpecTrAl Audio) processing, which are standard methods for dealing with channel distortions.

Datasets

The experiments were conducted using a subset of the YOHO database, adapted for telephone speech. This involved:

  • Training: A background speaker model was trained using 40 background impostor speakers.
  • Evaluation: An evaluation database (Y1) was used to estimate the relative importance of spectral components. This database consisted of enrolling utterances from 40 background speakers recorded with different telephone handsets (hset2, hset3, hset4).
  • Testing: Client models were trained using enrolling utterances from 30 target speakers recorded with a reference handset (hset1). Verification attempts used test utterances recorded with different handsets (hset5, hset6, hset7) to simulate channel mismatch.

Key Insights and Ideas

  • Spectral Domain Filtering is Powerful: Applying filters in the LFBE spectral domain, rather than directly on the time-domain features, offers a new perspective for tackling channel distortion. This domain seems to capture channel effects in a more structured way.
  • Time-Varying Distortion is Key: The paper highlights that static or slow-varying channel models are insufficient. The time-varying component of telephone channel distortion is a significant challenge that requires frame-by-frame compensation.
  • RIA for Targeted Compensation: Using a discriminative function within RIA provides a principled way to determine the optimal filter parameters, focusing on maximizing speaker separability while minimizing distortion.
  • Synergy with Existing Methods: The proposed spectral filtering method not only improves performance on its own but also complements and enhances the effectiveness of established techniques like CMN and RASTA.

Key Results

The proposed feature enhancing method demonstrated significant improvements in speaker verification performance, measured by the Equal Error Rate (EER):

  • Standalone Improvement: The feature transform alone reduced the EER by 10.8% compared to the baseline system without trajectory filtering.
  • Combined Improvements:
    • When combined with RASTA, the EER was further reduced by 23.5% compared to the baseline without trajectory filtering.
  • When combined with CMN, the EER was reduced by 40% compared to the baseline without trajectory filtering.
  • Comparison to CMVN: The paper notes that CMVN, another normalization technique, actually increased the error rate by 37.2% in this context, likely due to inaccurate variance estimation with limited data.

Limitations and Future Work

  • Limited Data Sensitivity: The paper notes that traditional methods like CMN and EM-based bias removal can be sensitive to utterance length and limited data availability. While the proposed method aims to mitigate this, its performance with extremely short utterances might still be a concern.
  • Future Research: The authors suggest that combining the filter-bank energy transformation with other channel removal approaches could be a promising direction for future research. They also imply that further exploration of the spectral domain for channel compensation is warranted.

Summary updated

BibTeX
@inproceedings{garreton2010enhancing,
  author = {Claudio Garretón and Néstor Becerra Yoma},
  title = {{On enhancing feature sequence filtering with filter-bank energy transformation in speaker verification with telephone speech}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure