The QUT-NOISE-TIMIT corpus for the evaluation of voice activity detection algorithms

David Dean, Sridha Sridharan, Robert Vogt, Michael MasonInterspeech 2010

TL;DR

This paper introduces the QUT-NOISE-TIMIT corpus, a massive 600-hour dataset of noisy speech designed to rigorously evaluate voice activity detection (VAD) algorithms across diverse real-world noise conditions. By mixing clean TIMIT speech with extensive background noise recordings from various environments, they provide a much-needed benchmark for VAD systems that goes beyond existing limited datasets.

What Problem Does This Paper Address?

Voice Activity Detection (VAD) is a fundamental component in many speech processing systems, from enhancement to recognition. Its job is to distinguish between speech and non-speech segments in audio. A major hurdle in developing and comparing VAD algorithms has been the lack of comprehensive, realistic datasets. Existing noisy speech corpora are often limited in duration, variety of noise types, and the number of speakers. This makes it difficult to truly assess how well VAD algorithms perform in the wild, where they encounter a wide spectrum of background noises at varying levels and in different acoustic environments. The authors aim to bridge this gap by creating a large-scale, diverse, and well-characterized noisy speech corpus.

Methods

The core of this work is the creation of the QUT-NOISE-TIMIT corpus. This involved two main stages:

  1. QUT-NOISE Corpus Construction: The researchers collected over 10 hours of background noise recordings from 10 distinct locations across 5 common noise scenarios (CAFE, HOME, STREET, CAR, REVERB). For reverberant environments (CAR and REVERB), they also captured impulse responses to simulate realistic acoustic reflections. The recording equipment used was a prosumer-grade Zoom H2, chosen for its quality.
  2. QUT-NOISE-TIMIT Corpus Generation: The collected QUT-NOISE background recordings were then mixed with clean speech segments from the widely-used TIMIT corpus. This mixing process was carefully controlled to generate 600 hours of noisy speech across 24,000 files. Key parameters controlled during mixing included:
  3. Signal-to-Noise Ratios (SNRs): Six different SNRs were used, ranging from -10 dB to 15 dB.
  4. Speech Proportions: Files were generated with less than 25%, 25-75%, and more than 75% speech content.
  5. Noise Lengths: Two durations for background noise segments were used (60 and 120 seconds).
  6. Reverberation: For CAR and REVERB scenarios, speech was convolved with the captured impulse responses.
  7. Speech Event Combination: To create longer speech events, TIMIT utterances were combined with a 50% probability, allowing for overlap and varied silence durations.

The paper also details an evaluation protocol where five baseline VAD systems (G729B, ETSI, LTSD, Sohn, and GMM-MFCC) were trained and tested. The evaluation was designed to be robust by training on one location group (A or B) and testing on the other, simulating a scenario where the VAD system knows the general noise level but not the specific environment.

Datasets

  • TIMIT Corpus: Used as the source of clean speech for generating the noisy speech sequences.
  • QUT-NOISE Corpus: A newly created corpus of over 10 hours of background noise recordings from 10 locations and 5 scenarios.
  • QUT-NOISE-TIMIT Corpus: The primary contribution of this paper. A 600-hour benchmark dataset of noisy speech, created by mixing QUT-NOISE with TIMIT. This dataset is designed for evaluating VAD algorithms.

Key Insights and Ideas

  • The critical need for diverse, large-scale noisy speech data: The paper strongly argues that existing datasets are insufficient for robust VAD evaluation, highlighting the limitations of short noise recordings.
  • Systematic construction of a realistic noisy speech corpus: The detailed methodology for mixing clean speech with varied background noises, SNRs, and speech proportions provides a blueprint for creating similar datasets.
  • Reverberation simulation is crucial: Capturing and applying impulse responses for specific environments adds a significant layer of realism to the generated noisy speech.
  • GMM-MFCC VAD systems show superior performance: The experimental results suggest that model-based approaches, particularly those using GMMs with MFCC features, can achieve better performance by learning cross-scenario characteristics.

Key Results

The paper evaluated five baseline VAD systems on the QUT-NOISE-TIMIT corpus. The key finding was that the GMM-MFCC, LTSD, and Sohn systems generally outperformed the standards-based ETSI and G729B systems. Specifically, the GMM-MFCC system demonstrated the best performance across all noise levels, attributed to its ability to learn cross-scenario characteristics of speech and non-speech. The paper presents results in terms of Half-Total-Error Rate (HTER), which combines Miss Rate (MR) and False Alarm Rate (FAR). While specific HTER values are presented graphically, the qualitative finding is that the more sophisticated, learning-based systems leverage the extensive data to achieve better segmentation.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the mixing process, particularly at low SNRs, can lead to some clipping due to high noise energy. However, they prioritized maintaining consistent reference energy levels.
  • While the CAR scenario used only one car, the authors attempted to mitigate this by varying driving conditions (highway vs. city).
  • Future Work:
    • The authors believe the QUT-NOISE-TIMIT corpus will serve as a solid foundation for developing and evaluating more robust VAD algorithms that can operate across a wide variety of noise scenarios.
  • The detailed metadata provided with the corpus (locations of TIMIT files and speech events) is intended to facilitate ground truth for speech detection experiments.

Summary updated

BibTeX
@inproceedings{dean2010qut,
  author = {David Dean and Sridha Sridharan and Robert Vogt and Michael Mason},
  title = {{The QUT-NOISE-TIMIT corpus for the evaluation of voice activity detection algorithms}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure