Unsupervised NAP training data design for speaker recognition

Hanwu Sun, Bin MaInterspeech 2012

TL;DR

This paper introduces a novel unsupervised clustering method to automatically generate training data for Nuisance Attribute Projection (NAP), a technique that significantly improves speaker recognition by removing unwanted variations. This approach bypasses the need for manually labeled data, making NAP more practical for real-world applications, and achieves performance comparable to systems using labeled data.

What Problem Does This Paper Address?

Speaker recognition systems are highly susceptible to variations in recording conditions, such as different microphones, background noise, and communication channels. Nuisance Attribute Projection (NAP) is a powerful technique that mitigates these issues by learning to remove these "nuisance attributes" from speech signals. However, NAP typically requires a large amount of labeled training data, where each utterance is tagged with specific speaker and channel information. This labeling process is time-consuming and expensive, limiting NAP's applicability in scenarios where such data is unavailable or difficult to obtain. This paper addresses this gap by proposing an unsupervised method to create effective NAP training data without any prior labeling.

Methods

The core of this paper is an unsupervised clustering strategy for designing NAP training data. The authors leverage techniques from speaker diarization to group unlabeled utterances into speaker-dependent clusters. The process involves:

  1. Supervector Generation: Extracting speaker supervectors (concatenated mean vectors from Gaussian Mixture Models) from speech segments.
  2. Initial Clustering: Randomly dividing these supervectors into an initial set of clusters.
  3. Purification and Iterative Refinement: A key contribution is a fast purification process. This involves iteratively assigning supervectors to clusters based on their similarity (measured by dot product) to existing cluster centroids. The algorithm refines these assignments until no further changes occur.
  4. Cluster Pruning: Small clusters (e.g., containing fewer than 3 supervectors) are discarded and their members are re-assigned to the most similar larger clusters.
  5. NAP Matrix Generation: Once speaker-dependent clusters are formed, they are used to drive the NAP matrix, which is then applied to compensate for channel variations.

The speaker recognition system itself uses a standard GMM-SVM classifier with MFCC features, including their first and second derivatives. Spectral subtraction is used for noise reduction, and RASTA filtering and cepstral mean and variance normalization (MVN) are applied to the features.

Datasets

  • NIST 2004 Speaker Recognition Evaluation (SRE04) corpus: Used for evaluating the unsupervised clustering method itself, as NAP labels are available for this dataset.
  • NIST 2006 Speaker Recognition Evaluation (SRE06) 1conv-1conv all English trials: Used for evaluating the speaker recognition system with different NAP approaches.
  • NIST 2008 Speaker Recognition Evaluation (SRE08) short2-short3 Condition 7 Tel-Tel all English trials: Also used for evaluating the speaker recognition system.

No new benchmarks were introduced; the paper focuses on applying the unsupervised method to existing standard evaluation datasets.

Key Insights and Ideas

  • Unsupervised NAP is Viable: The most significant insight is that effective NAP training data can be generated without manual labeling, achieving performance very close to that obtained with meticulously labeled data. This dramatically lowers the barrier to using NAP in practical systems.
  • Leveraging Diarization Techniques: The paper demonstrates the power of adapting speaker diarization algorithms for tasks beyond just segmentation, showing their utility in feature engineering for speaker recognition.
  • Clustering Quality Matters: The performance gains are directly tied to the quality of the unsupervised clustering. The proposed purification and iterative refinement process is crucial for creating meaningful speaker-dependent groups.
  • Robustness Across Datasets: The unsupervised NAP approach shows consistent improvements on both the SRE06 and SRE08 Tel-Tel subtasks, indicating its generalizability.

Key Results

  • The unsupervised NAP training data design achieved an Equal Error Rate (EER) of 2.89% and a minimum Detection Cost Function (DCF) of 1.62% on the SRE06 1conv-1conv trials.
  • This represents a significant improvement: a 41.7% relative reduction in EER and a 32.2% relative reduction in minimum DCF compared to systems without any NAP compensation.
  • The performance of the unsupervised method was very close to that achieved using ground-truth labeled NAP data, with only a small difference in EER and DCF.
  • For the SRE08 Tel-Tel subtask, the unsupervised NAP design achieved a 36.8% EER reduction over the "No NAP" condition and a 25.5% EER reduction over the "Random NAP" condition.
  • The unsupervised clustering achieved approximately 93% speaker cluster rate and 90% speaker purification rate on the SRE04 dataset.

Limitations and Future Work

  • Dataset Specificity: The current unsupervised clustering experiment was conducted on the SRE04 telephone-based dataset.
  • Future Work: The authors suggest extending the algorithm to handle more complex NAP training conditions, specifically mentioning the NIST SRE microphone channel datasets and interviewer channel datasets. This implies a need to test the robustness of the unsupervised method against a wider range of acoustic variations.

Summary updated

BibTeX
@inproceedings{sun2012unsupervised,
  author = {Hanwu Sun and Bin Ma},
  title = {{Unsupervised NAP training data design for speaker recognition}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure