Soft Counting Poisson Mixture Model-Based Polling Method for Speech/Nonspeech Classification

Youngjoo Suh, Hoirin Kim, Minsoo Hahn, Yong-Ju LeeIEICE Trans. Information and Systems 2006 · DOI

TL;DR

This paper introduces a novel method for classifying speech from nonspeech segments in audio, crucial for robust speech recognition in noisy environments. By enhancing the traditional Poisson Polling Method (PPM) with Poisson mixture models and a "soft counting" approach, the authors achieve significantly improved accuracy in distinguishing speech from various background noises.

What Problem Does This Paper Address?

Speech recognition systems often struggle in real-world scenarios like noisy homes or with voice-controlled robots. These environments are rife with nonspeech sounds (music, laughter, appliance noise) that can be mistakenly identified as speech by standard detectors. This leads to inefficient resource use and reduced recognition accuracy. While segment-level classification (analyzing longer audio chunks) is more practical than frame-by-frame analysis, existing methods often rely on simple accumulation of frame scores, lacking sophistication in modeling the diverse patterns of speech and nonspeech. The authors aim to bridge this gap by developing a more accurate and robust segment-level classification method.

Methods

The paper proposes two key enhancements to the baseline Poisson Polling Method (B-PPM):

  • Poisson Mixture Models: Instead of a single Poisson distribution to model observed feature frequencies, the authors use a mixture of Poisson distributions. This allows for a more flexible and accurate representation of the varied patterns found in speech segments. Think of it like using multiple different "templates" to describe speech, rather than just one.
  • Soft Counting: Traditional methods "hard count" features – a feature either belongs to a specific quantized level or it doesn't. This paper introduces "soft counting," where the contribution of a feature to a quantized level is weighted by its closeness to that level and its neighbors. This acknowledges that a feature might be "almost" in one level but still relevant to it, leading to a more reliable estimation of observed frequencies. Two variants are proposed:
  • Global Soft Counting (GSC) PPM: Centroids for quantization levels are computed globally from both speech and nonspeech data.
  • Class-based Soft Counting (CSC) PPM: Centroids are computed separately for speech and nonspeech classes, allowing for more class-specific modeling.

Datasets

The authors evaluated their methods using:

  • Speech Data: 13,988 distant-talking Korean utterances (approx. 8.2 hours) from 119 speakers.
  • Nonspeech Data: 2.5 hours of various home environment noises (music, laughter, appliance noise, etc.).

All data was sampled at 16kHz, blocked into 20ms frames with a 10ms interval. For feature extraction, a 26-dimensional vector (13-order mel-cepstral coefficients and their derivatives) was used, and likelihood ratios were computed from GMMs. Segments were fixed at 500ms with a 250ms interval for classification.

Key Insights and Ideas

  • Beyond Simple Accumulation: The core idea is that simply adding up frame-level scores isn't enough for robust segment-level classification. Capturing the distributional characteristics of features within a segment is crucial.
  • Mixture Models for Variability: Speech and nonspeech signals are not monolithic. Using mixture models acknowledges this inherent variability and allows the classifier to adapt better to different types of speech and noise.
  • Soft Counting for Robustness: Hard counting can be brittle. Soft counting provides a more nuanced way to estimate feature frequencies, making the system less sensitive to minor variations and quantization artifacts. It's like saying, "This sound is mostly like X, but it also has some characteristics of Y, so let's consider both."
  • Class-Specific Modeling Pays Off: The Class-based Soft Counting (CSC) approach generally outperformed the Global Soft Counting (GSC) approach, suggesting that tailoring the modeling to each class (speech vs. nonspeech) leads to better discrimination.

Key Results

  • The proposed methods (M-PPMs with soft counting) significantly outperformed the baseline B-PPM and a GMM-based segmentally accumulated log-likelihood ratio (GSALR) method.
  • Maximum error reduction:
    • 32% over B-PPM.
  • 39% over GSALR.
  • The CSC-PPM variant generally showed the best performance.
  • Increasing the number of mixture components in M-PPMs led to further error reduction.
  • Soft counting techniques were found to be very effective in improving both classification and speech detection accuracy.

Limitations and Future Work

  • Computational Cost: While the paper mentions efficiency, the introduction of mixture models and soft counting likely increases computational complexity compared to the baseline. The authors mention taking steps to relieve computational loads (e.g., limiting neighbor levels in soft counting), but this is an inherent trade-off.
  • Parameter Tuning: The performance of mixture models and soft counting depends on proper parameter tuning (number of components, quantization levels, distance metrics). The paper uses fixed values for experiments, but optimal tuning might be data-dependent.
  • Future Work: The authors suggest that incorporating mixture models and soft counting techniques are effective avenues for improving speech/nonspeech classification. Further exploration of these techniques and potentially other feature representations could lead to even better performance. The paper implicitly suggests that exploring different numbers of mixture components and quantization levels could yield further improvements.

Summary updated

BibTeX
@article{suh2006soft,
  author = {Youngjoo Suh and Hoirin Kim and Minsoo Hahn and Yong-Ju Lee},
  title = {{Soft Counting Poisson Mixture Model-Based Polling Method for Speech/Nonspeech Classification}},
  journal = {IEICE Trans. Information and Systems},
  year = {2006},
  doi = {10.1093/ietisy/e89-d.12.2994},
}
Figure