Speaker recognition experiments using connectionist transformation network features

Alberto Abad, Isabel TrancosoInterspeech 2010

TL;DR

This paper explores using adaptation transforms from speech recognition systems as features for speaker recognition. They introduce and evaluate a "Transformation Network features with SVM modelling" (TN-SVM) approach, showing it can achieve good performance even with imperfect speech transcriptions and can be enhanced by session variability compensation techniques.

What Problem Does This Paper Address?

Speaker recognition systems typically rely on modeling short-term acoustic features like cepstral coefficients. However, these features might not fully capture the higher-order structures in speech that characterize a speaker. Researchers are exploring alternative feature extraction methods. One promising avenue is to leverage speaker adaptation techniques, commonly used in Automatic Speech Recognition (ASR), to derive features that represent speaker-specific information. This paper specifically investigates using the weights of "Transformation Networks" (TN), a connectionist adaptation method, as features for speaker recognition, and how to optimize this process by exploring different transcription sources and compensation methods.

Methods

The core of the paper is the Transformation Network features with SVM modelling (TN-SVM) approach. Here's a breakdown:

  • Connectionist Adaptation (Transformation Network - TN): Instead of directly modeling acoustic features, they use a linear input network (TN) to map speaker-dependent (SD) input vectors to a speaker-independent (SI) connectionist system. The weights of this TN, which represent the "difference" between a specific speaker and a generic model, are extracted.
  • Feature Extraction: These TN weights are concatenated with segment mean and variance statistics to form a high-dimensional feature vector for each speech segment.
  • Speaker Modelling (SVM): Support Vector Machines (SVM) with a linear kernel are then trained using these TN feature vectors to build speaker models.
  • Phonetic Alignment Exploration: The paper investigates the impact of the source of phonetic alignment on TN feature extraction. They compare using:
  • NIST-provided transcripts (Baseline).
  • Weak transcriptions from their own ASR system (Word model).
  • Phonetic sequences generated by a simple phonetic model (Phone model).
  • Session Variability Compensation (NAP): They apply Nuisance Attribute Projection (NAP), a technique designed to remove unwanted variations (like channel differences) from high-dimensional feature spaces, to the TN features.

Datasets

  • Training Data:
    • Additional training data from previous NIST SRE evaluations (SRE2004, SRE2005, SRE2006) for background modeling and NAP projection matrix training.
  • Broadcast News (BN) data for training the baseline ASR system.
  • Evaluation Data:
    • A subset of the NIST Speaker Recognition Evaluation 2008 (NIST SRE 2008) telephone-telephone training and test condition. This condition consists of two-channel telephone conversational excerpts.

Key Insights and Ideas

  • Adaptation Transforms as Rich Features: The paper reinforces the idea that speaker adaptation transforms, particularly those derived from connectionist models, can capture rich speaker-specific information beyond traditional short-term features.
  • Robustness to Imperfect Transcriptions: While high-quality transcriptions are ideal, the TN-SVM approach shows a surprising degree of robustness. Even with weak transcriptions from their own ASR system or a simple phonetic model, performance degradation is observed but remains significant, suggesting the TN features are inherently informative.
  • Phonetic Model as a Viable Alternative: The phonetic model approach, while slightly worse than the word model, demonstrates that even a simplified phonetic representation can be useful for generating alignments when high-quality ASR is unavailable.
  • NAP Enhances Performance: Applying NAP for session variability compensation leads to improved speaker recognition performance, particularly in reducing the impact of channel differences, even in a relatively controlled test condition.

Key Results

  • Baseline TN-SVM: Achieved a minimum detection cost function (minDCF) of 6.59 and an Equal Error Rate (EER) of 17.33% using NIST transcriptions.
  • Impact of Transcriptions:
    • Using their own ASR system (Word model) resulted in a performance drop to minDCF of 7.82 and EER of 20.76%.
  • Using a phonetic decoder (Phone model) yielded minDCF of 7.97 and EER of 21.41%.
  • NAP Compensation:
    • With NAP compensation (e.g., 32 dimensions), performance improved to minDCF of 6.11 and EER of 16.23%. This represents a relative reduction of 7.3% in minDCF and 6.4% in EER compared to the baseline.

Limitations and Future Work

  • ASR Mismatch: The significant performance drop when using their own ASR system highlights a strong mismatch between the BN training data and the target CTS-like data. This suggests that a better-matched ASR system would yield even better TN features.
  • Limited Cross-Channel Variability: The study was conducted on a telephone-telephone condition, which has less cross-channel variability than other SRE conditions. The authors suggest that evaluating TN-SVM with NAP in scenarios with stronger cross-channel variability would be beneficial.
  • Language Independence: Future work aims to use the phonetic decoder approach to obtain speaker-dependent mappings for any phonetic network, regardless of the language of the utterance or the phone classifier. This could lead to language-independent TN features.

Summary updated

BibTeX
@inproceedings{abad2010speaker,
  author = {Alberto Abad and Isabel Trancoso},
  title = {{Speaker recognition experiments using connectionist transformation network features}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure