TL;DR
This paper tackles the challenge of automatically tuning a crucial parameter, ro, in time-frequency signal analysis methods that leverage spectrogram zeros. By comparing different statistical tests and "summary functions" for estimating ro on real-world data, they provide practical recommendations for signal detection and estimation, particularly for audio processing.
What Problem Does This Paper Address?
Time-frequency (TF) representations are essential for analyzing signals whose frequency content changes over time. A recent class of methods uses the "zeros" of a spectrogram (a TF representation) to detect and reconstruct signals. These zeros, when a signal is present, form "holes" in the otherwise regular pattern observed in pure noise. Identifying these holes helps delineate the signal's TF support.
However, these methods rely on a hyperparameter, ro, which acts as a scale or radius to distinguish signal-induced holes from random noise fluctuations. Choosing an appropriate ro is critical: too large, and you lose signal information; too small, and noise contaminates the estimated signal. The problem is that this parameter is difficult to tune, especially when the noise characteristics are unknown. This paper addresses the lack of systematic studies on how to adaptively tune ro using real data and investigates its impact on signal detection and estimation quality, particularly for audio signals.
Methods
The core of the paper involves comparing different statistical approaches to estimate the ro parameter. They focus on three "summary functions" that describe the spatial distribution of spectrogram zeros:
- Empty Space Function (F): A classic measure of how likely it is to find no points within a certain radius.
- Variance-Stabilized Empty Space Function (F): A modified version of F to improve statistical properties.
- Accumulated Persistence Function (APF): A novel approach drawing from topological data analysis, which captures more complex structural information about the zero patterns.
These summary functions are used within two statistical tests for signal detection:
- Global MAD Test: Compares the observed summary function to simulated ones using a maximum absolute deviation.
- Global Rank Envelope Test: A more recent test that assigns a rank to the observed function based on its extremality compared to simulations.
Once a signal is detected (i.e., the null hypothesis of pure noise is rejected), ro is estimated. The signal is then reconstructed using a "Delaunay triangulation" (DT) approach, where triangles formed by spectrogram zeros are filtered based on the estimated ro. The paper also compares this to a standard "Hard Thresholding" (HT) method for audio denoising.
Datasets
The paper uses synthesized signals and real-world audio excerpts for evaluation. Specifically:
- Synthetic Linear Chirp: Used for initial detection power analysis.
- Speech Signals: Synthesized using Google Cloud realistic text-to-speech API.
- Cello and Trumpet Excerpts: Real musical instrument recordings.
- Babble Noise: From the TIMIT database, used to contaminate the audio signals at various Signal-to-Noise Ratios (SNRs).
The authors also contribute mcsm-benchs, a Python-based package for collaborative benchmarking in signal processing, which is of independent interest.
Key Insights and Ideas
- Adaptive Tuning is Crucial: The study strongly suggests that hand-tuning
rois suboptimal. Adaptive estimation using statistical tests on the data itself is effective and can lead to near-optimal results. - Summary Function Matters: The choice of summary function (F, F, or APF) impacts the estimation of
roand, consequently, the signal estimation quality. APF, being a novel approach, shows promise for stableroestimation across different signals. - Detection Test Choice is Important: For low SNRs, the global rank envelope test generally outperforms the global MAD test in terms of detection power.
- Zero-Based Methods vs. Traditional Denoising: While zero-based methods can achieve comparable general audio quality (PESQ scores) to traditional methods like Hard Thresholding, they may introduce more audible artifacts ("musical noise"), especially without post-processing. This is a critical consideration for audio applications.
- APF's Trade-off: APF tends to yield a larger
ro, which can improve signal-noise separation but might leave more small, isolated regions that contribute to artifacts.
Key Results
- Signal Detection: For SNRs above 10 dB, most test/summary function pairs perform well. At lower SNRs, the global rank envelope test is the most effective, regardless of the summary function. The global MAD test is computationally less expensive due to requiring fewer simulations.
- Signal Estimation: The adaptive estimation of
rousing the proposed methods is effective, leading to comparable or better results than a "Best DT" baseline. The APF summary function provides more stableroestimates across different signals compared to F and F. - Audio Quality Metrics:
- PESQ: Zero-based methods and Hard Thresholding show similar performance for general audio quality.
- APS: APF-based
roestimation can lead to lower APS scores (more musical noise) compared to F and F, indicating a trade-off between signal separation and artifact introduction.
Limitations and Future Work
- Artifacts in Audio: A key limitation acknowledged is that zero-based denoising methods, without post-processing, might introduce more musical noise artifacts than traditional thresholding methods.
- Global vs. Local Scale: For signals with complex TF structures, using a single global
romight be insufficient. Future work could explore local scale parameters dependent on TF location. - Connections to Topological Data Analysis: The authors highlight the potential of further exploring the connection between spectrogram zeros and topological data analysis, particularly inspired by the APF results.
- Benchmarking: The paper promotes the use of
mcsm-benchsfor collaborative benchmarking, suggesting future work could involve comparing more methods on this platform.
Summary updated