STC Speaker Recognition Systems for the VOiCES from a Distance Challenge

Sergey Novoselov, Aleksei Gusev, Artem Ivanov, Timur Pekhovsky, Andrey Shulipa, Galina Lavrentyeva, Vladimir Volokhov, Alexandr KozlovInterspeech 2019

TL;DR

This paper presents STC's approach to the VOICES 2019 challenge, focusing on speaker recognition in noisy, far-field audio. They achieved strong results by exploring various deep neural network architectures for speaker embeddings, particularly those with residual connections and LSTM layers, and by carefully preparing training data with realistic reverberation and noise augmentation.

What Problem Does This Paper Address?

Speaker recognition, the task of identifying who is speaking, is notoriously difficult in real-world scenarios. This paper specifically tackles the challenge of far-field, noisy audio, where the speaker is distant from the microphone and the audio is corrupted by background noise and reverberation. This is a significant hurdle for traditional speaker recognition systems, which often perform best in clean, close-talking conditions. The VOICES 2019 challenge aimed to push the boundaries of speaker recognition in these difficult, "in-the-wild" conditions.

Methods

The STC team explored several key methodological components:

  • Deep Neural Network (DNN) Architectures for Speaker Embeddings:
    • x-vectors: A well-established DNN-based approach for extracting speaker embeddings. They experimented with standard x-vector architectures (TDNN) and extended versions incorporating LSTM layers for improved temporal modeling.
  • c-vectors: A newer architecture based on residual blocks (ResTDNN) and advanced activation functions (MFM, A-Softmax), aiming for more robust embeddings.
  • Data Augmentation and Preparation:
    • Reverberation Simulation: A crucial aspect was simulating realistic acoustic environments. They used an automatic room impulse response (RIR) generator to add reverberation to training data, making the models more robust to echoes.
  • Noise and Babble Addition: Standard augmentation techniques like adding babble, music, and noise were also employed.
  • Speech Activity Detection (SAD): Two types of SAD were investigated: a simple energy-based SAD and a more sophisticated ASR-based SAD, with the latter showing promise for improved performance.
  • Back-End Scoring:
    • Cosine Similarity Metric Learning (CSML): This discriminative approach was used to compare speaker embeddings and generate scores.
  • Score Normalization: Techniques like score s-normalization were applied to improve the reliability of the system's output.
  • Fusion and Calibration: Multiple individual systems were combined (fused) using linear regression to leverage their complementary strengths. Calibration was performed using logistic regression.

Datasets

  • Training Data:
    • VoxCeleb1 & VoxCeleb2: Large-scale datasets of celebrity speech, widely used for speaker recognition research.
  • SITW (Speakers in the Wild): A dataset designed to simulate more realistic, unconstrained speaking conditions.
  • NIST SREs (Speaker Recognition Evaluations): Data from previous NIST evaluations was incorporated for open training conditions.
  • MUSAN and RIR datasets: Used for generating augmented data (noise, music, reverberation).
  • Evaluation Data:
    • VOiCES 2019 Development and Evaluation Sets: The official datasets for the challenge, representing far-field, noisy audio.

Key Insights and Ideas

  • Deeper is Better (with caveats): Deeper DNN architectures, particularly those incorporating LSTM layers for better temporal modeling, generally outperformed shallower counterparts.
  • Realistic Augmentation Matters: The use of an automatic RIR generator for more realistic reverberation simulation significantly improved robustness. This highlights the importance of data that closely mimics real-world acoustic conditions.
  • ASR-based SAD Shows Promise: While energy-based SAD is simpler, ASR-based SAD demonstrated the potential to achieve better quality by more accurately identifying speech segments.
  • 16kHz Sampling Rate is Preferred: For this specific task, using a 16kHz sampling rate and 40-dimensional MFCC features yielded better results than downsampling to 8kHz, suggesting that preserving more audio information is beneficial.
  • Fusion is Key: Combining multiple well-performing individual systems through fusion significantly boosted overall performance, demonstrating the benefit of diverse system components.
  • Score Normalization is Crucial: Applying score normalization techniques provided an additional performance gain, making the system's outputs more reliable.

Key Results

The STC systems achieved competitive results in the VOICES 2019 challenge. Their best single system, Xvect-Ext-TDNN-LSTM-V3, achieved a minDCF of 0.194 on the development set and 0.349 on the evaluation set. Their fused systems further improved performance, with Open-2 achieving an EER of 1.62% and a minDCF of 0.172 on the development set and 4.53% EER and 0.315 minDCF on the evaluation set. These results indicate that their approach was effective in handling the challenging far-field, noisy conditions of the challenge.

Limitations and Future Work

  • Open Conditions Degradation: Systems trained for open conditions (using 8kHz downsampled data) showed a significant quality degradation compared to those trained on 16kHz data. This suggests that downsampling is detrimental for this task.
  • WPE Preprocessing: While WPE (Weighted Prediction Error) for dereverberation improved performance on evaluation data for open conditions, it was not applied during training. Future work could explore integrating WPE into the training pipeline.
  • Further Exploration of Architectures: The paper hints at the potential for further exploration of DNN architectures, particularly in how they handle temporal dependencies and residual connections.
  • Data Augmentation Strategies: While their augmentation was effective, continued research into more sophisticated and realistic data augmentation techniques could yield further improvements.

Summary updated

BibTeX
@inproceedings{novoselov2019stc,
  author = {Sergey Novoselov and Aleksei Gusev and Artem Ivanov and Timur Pekhovsky and Andrey Shulipa and Galina Lavrentyeva and Vladimir Volokhov and Alexandr Kozlov},
  title = {{STC Speaker Recognition Systems for the VOiCES from a Distance Challenge}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure