TL;DR
This paper introduces a novel blind signal-to-noise ratio (SNR) estimation method specifically designed for speech corrupted by high levels of noise (SNR < 0 dB). By combining a robust pitch tracking algorithm with a 2D image processing approach on a "segmented speech matrix," the method achieves accurate SNR estimation even in challenging noisy environments where traditional methods fail.
What Problem Does This Paper Address?
Estimating the Signal-to-Noise Ratio (SNR) is a crucial preprocessing step for many speech processing tasks, including noise reduction and speech recognition. While existing methods perform well in low to moderate noise conditions (SNR > 0 dB), they often struggle significantly in high noise environments (SNR < 0 dB). This limitation hinders the performance of speech processing systems in real-world scenarios where speech is frequently masked by substantial background noise, such as in noisy vehicles or industrial settings. The authors aim to bridge this gap by developing an SNR estimator that is specifically robust to these challenging high-noise conditions.
Methods
The proposed system operates in two main stages:
- Robust Fundamental Period Analysis: The core of the method lies in a novel pitch tracking algorithm designed to work in high noise. The hypothesis is that even in noisy speech, the short-term autocorrelation and long-term periodicity of voiced segments are somewhat preserved. This algorithm computes a "score function" based on the variance of samples at candidate fundamental period intervals. To improve robustness, especially against periodic noise like car noise, a high-pass filter (around 300 Hz) is applied as a pre-processing step. This filter helps the algorithm track speech periodicity rather than noise periodicity.
- Two-Dimensional SNR Estimation: The output of the pitch tracker is used to construct a "segmented speech matrix" (SSM). This matrix is a 2D representation where each row corresponds to a single period of the speech signal. The SSM is then treated as an image, and a 2D autocorrelation-based SNR estimation technique, adapted from image processing, is applied. This approach leverages both the short-term (horizontal axis of the SSM) and long-term (vertical axis, related to pitch) properties of speech to estimate the SNR. The final SNR estimation formula is derived from the autocorrelation of the mixture signal, using estimations of the clean speech and noise components.
Datasets
- NOISEX-92 database: Used for car noise samples.
- NTT Advanced Technology Corporation, Multi-Lingual Speech Database for Telephonometry 1994: Used for real speech data, consisting of phonetically balanced sentences from male and female speakers.
No new benchmarks were introduced; existing datasets were used for evaluation.
Key Insights and Ideas
- Exploiting Periodicity in Noise: The fundamental insight is that even in very noisy speech, the underlying periodicity of voiced segments can be a strong cue for both pitch tracking and SNR estimation.
- 2D Representation for SNR: Treating the segmented speech signal as a 2D image allows for the application of powerful image processing techniques to extract SNR information, effectively combining temporal and pitch-related features.
- High-Pass Filtering for Noise Discrimination: The simple addition of a high-pass filter significantly improves the pitch tracker's ability to distinguish speech periodicity from periodic noise, a critical step for robust performance in real-world noisy environments.
- Robustness to High Noise: The method demonstrates a significant improvement over traditional techniques, maintaining reasonable accuracy even at SNR levels below 0 dB.
Key Results
- Fundamental Period Tracking: The proposed pitch tracking method shows good performance even at SNR of -10 dB. The standard deviation of the tracking error is significantly reduced when using the high-pass filter, especially for car noise (e.g., standard deviation of 0.0004 with HP filter vs. 0.0830 without for -5 dB SNR car noise).
- SNR Estimation: The SNR estimation method performs accurately for white noise down to approximately -10 dB. For car noise, the standard deviation of the estimation error remains relatively constant across a wide SNR range, although the average error gradually increases. The method is particularly effective in high energy sections of speech.
Limitations and Future Work
- Low Energy Sections: The paper notes that performance degrades in low energy sections of speech where periodicity is absent, as this negates the benefits of the 2D representation.
- Adaptive Pre-filtering: The authors suggest that further investigation into making the high-pass filter's parameters adaptive to the input signal could lead to improved performance.
- Voice Activity Detection (VAD): Incorporating a VAD could enhance accuracy by preventing the system from processing noise-only segments, which can introduce bias.
- Current Application: The system is currently being used as part of a speech intelligibility enhancement application under development.
Summary updated