Spectral intersections for non-stationary signal separation

Trausti Kristjansson, Thad HughesInterspeech 2012

TL;DR

This paper introduces "Spectral Intersections," a novel and computationally efficient method for separating speech from non-stationary noise like music. It achieves performance comparable to more complex algorithms, offering a significant improvement over traditional methods and a practical alternative for real-world applications.

What Problem Does This Paper Address?

In today's world of ubiquitous mobile devices and voice-activated assistants, robust speech recognition is paramount. However, real-world audio is often contaminated by non-stationary noise, such as music, which significantly degrades recognition accuracy. Traditional noise suppression techniques like spectral subtraction are effective for stationary noise but struggle with dynamic sounds. While more advanced model-based methods like Algonquin can handle non-stationary noise, they are often computationally expensive and complex to implement. This paper aims to bridge this gap by proposing a simpler yet highly effective method for non-stationary noise suppression.

Methods

The core of the "Spectral Intersections" method lies in its novel approach to approximating the Minimum Mean Squared Error (MMSE) signal separation. Instead of directly modeling signals in the time domain, the authors work in the log-spectrum domain, representing signals as Gaussian Mixture Models (GMMs).

The key innovation is how they approximate the complex expectation integral required for MMSE estimation. They leverage the observation that acoustic mixing, in the magnitude spectrum, can be approximated by either constructive (signals in phase) or destructive (signals out of phase) combinations. By focusing on the "intersections" of these idealized mixing lines with lines representing the prior GMM modes, they derive two specific points (constructive and destructive intersections) in the log-spectrum domain. These points are then used as sparse samples for the expectation integral, a technique they liken to "unscented importance sampling." This significantly simplifies the computation while retaining good performance.

Datasets

The authors evaluated their method on the Google Voice Search dataset, which comprises approximately 38,000 manually transcribed English-language spoken queries. To create realistic noisy conditions, they artificially mixed music into these utterances at various Signal-to-Noise Ratios (SNRs). The music used for mixing was sourced from a database of 500 popular songs.

Key Insights and Ideas

  • The power of simplification: The paper demonstrates that by focusing on idealized constructive and destructive mixing scenarios, a significant computational burden can be lifted without sacrificing performance.
  • Log-spectrum domain is key: Working in the log-spectrum domain, combined with GMMs, provides a flexible framework for modeling speech and noise.
  • "Unscented importance sampling" as a practical approximation: The novel sampling strategy, inspired by unscented Kalman filters, offers a computationally efficient way to approximate complex integrals in MMSE estimation.
  • Performance rivaling complexity: The fact that a simpler method can achieve results comparable to much more complex algorithms is a significant finding, suggesting potential for wider adoption.

Key Results

The Spectral Intersections method achieved a 4% relative reduction in Word Error Rate (WER) on the Google Voice Search task, approaching the performance of the more complex Algonquin algorithm. Notably, it substantially outperformed the traditional Max method. At a 10 dB SNR, Spectral Intersections provided a 7% relative reduction in WER, outperforming Algonquin in this specific condition. The paper also highlights that the computational complexity of Spectral Intersections is comparable to that of the Max and Algonquin methods, making it a practical choice.

Limitations and Future Work

The authors acknowledge that their method is an approximation and that the "unscented importance sampling" approach is a simplification. While it performs well, it might not capture all the nuances of complex acoustic mixing. The paper doesn't explicitly detail limitations beyond the inherent approximations of their model. Future work could explore further refinements of the sampling strategy or investigate its application to other types of non-stationary noise beyond music.

Summary updated

BibTeX
@inproceedings{kristjansson2012spectral,
  author = {Trausti Kristjansson and Thad Hughes},
  title = {{Spectral intersections for non-stationary signal separation}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure