The IIR NIST SRE 2008 and 2010 summed channel speaker recognition systems

Hanwu Sun, Bin Ma, Chien-Lin Huang, Trung Hieu Nguyen, Haizhou Li

This paper, presented at INTERSPEECH 2010 by researchers from the Institute for Infocomm Research (I2R) in Singapore, details their high-performing submission to the NIST Speaker Recognition Evaluations (SRE). It focuses on the "summed channel" problem—a scenario where two speakers are recorded on a single audio track, making it difficult to isolate and identify the target individual.

Summary

The researchers developed a robust speaker recognition system for "summed channel" (2-wire) telephone recordings by integrating a sophisticated speaker diarization front-end with a GMM-SVM backend. By effectively separating "who spoke when" before attempting recognition, they achieved a massive ~50% relative improvement in error rates compared to systems without diarization.

The Summed Channel Challenge

In ideal speaker recognition scenarios, a "4-wire" recording is used where each speaker is on their own dedicated channel. However, in many real-world analogue telephone systems, these channels are summed into a single "2-wire" track. This creates significant interference and requires the system to determine which audio segments belong to the target speaker without prior knowledge of either voice.

Proposed Methodology

The authors propose a three-stage pipeline that explicitly separates the speakers before recognition. The architecture, shown in Figure 1, handles both enrollment and testing phases.

System architecture for enrollment and test phases, highlighting the integration of speaker diarization.
  1. Voice Activity Detection (VAD): A spectral subtraction-based VAD identifies and selects the most "useful" speech frames while discarding silence and non-speech noise.
  2. Hybrid Speaker Diarization: This core innovation handles the two-speaker telephone call scenario through purification (using EM and MAP adaptation) and refinement (applying Viterbi decoding and BIC) to iteratively merge segments until only two distinct speaker clusters remain.
  3. Speaker Recognition (GMM-SVM): The system extracts both MFCC and LPCC features. A GMM-SVM framework is used, where Gaussian Mixture Model (GMM) supervectors are fed into a Support Vector Machine (SVM). Nuisance Attribute Projection (NAP) and TZnorm are applied to handle channel variability and normalize scores.

Key Insights

  • Diarization is Essential: Speaker recognition on summed channels is significantly degraded without diarization. Separating speakers first is more critical than the specific recognition architecture.
  • Feature Fusion: MFCCs and LPCCs capture complementary acoustic information. Fusing scores from both models consistently outperformed using either alone.
  • The "Higher Score" Strategy: After diarization splits the audio, the system runs the model against both tracks and selects the one yielding the higher match score.

Experimental Results

The system was evaluated using NIST SRE 2008 and 2010 corpora (specifically the 1conv-summed and 8conv-summed tasks). The impact of the diarization step was dramatic. As shown in the performance table below for the 8conv-summed task, the system achieved significant error reductions across all metrics.

Comparison of EER and DCF metrics before and after diarization for different feature sets in the 8conv-summed task.

On the NIST SRE 2008 English trials:

  • 1conv-summed (1 conversation): The Equal Error Rate (EER) dropped from 6.68% to 3.46% (a 48.2% relative improvement) after applying diarization.
  • 8conv-summed (8 conversations): The EER dropped from 4.08% to 1.87% (a 54.2% relative improvement).
  • Minimum Detection Cost (DCF): The system saw a 43.1% relative improvement in the 1conv task, indicating much higher reliability for decision-making.

Limitations and Future Work

While effective, the system assumes a two-speaker scenario, which may struggle in multi-party environments. Additionally, handling "double-talk" (overlapping speech) remains a significant hurdle, as diarization error rates increase substantially in these segments. Future work aims to explore scenarios where training data is also summed-channel to better simulate forensic conditions.

Summary updated

BibTeX
@inproceedings{sun2010iir,
  author = {Hanwu Sun and Bin Ma and Chien-Lin Huang and Trung Hieu Nguyen and Haizhou Li},
  title = {{The IIR NIST SRE 2008 and 2010 summed channel speaker recognition systems}},
  booktitle = {Interspeech},
  year = {2010},
  doi = {10.21437/interspeech.2010-132},
}
Figure