Abstract
Speech coding is a critical technique for transforming high-bandwidth speech data into low-bandwidth digital signals suitable for narrowband underwater acoustic communication. This paper explores the Mixed-Excitation Linear Predictive (MELP) coding algorithm and applies the Perceptual Evaluation of Speech Quality (PESQ) metric for objective quality assessment. To mitigate stationary noise components in decoded speech, a spectral subtraction method is employed. Experimental results indicate that this approach significantly improves PESQ scores and enhances auditory perception, even in high-error environments.
Introduction
Underwater acoustic communication is characterized by extremely limited bandwidth, making high-quality voice transmission a significant engineering challenge. To overcome these constraints, low-bitrate speech coding is typically employed. However, these codecs often introduce "steady noise" or artifacts that degrade signal clarity.
Traditionally, speech quality is measured using the Mean Opinion Score (MOS), which relies on human panels. In underwater operations, subjective testing is often impractical due to small sample sizes and the confidential nature of the communications. This paper addresses these challenges by implementing an objective evaluation framework and a computationally efficient enhancement technique.
Methodology
The researchers utilized a three-part technical approach to improve and evaluate speech quality:
- MELP Coding: The 2.4kbps Mixed-Excitation Linear Predictive algorithm was selected for its high efficiency in narrowband channels and its ability to preserve the speaker's unique vocal characteristics.
- Spectral Subtraction: This core enhancement technique identifies "quiet sections" in the audio to estimate the power spectrum of background noise. This estimate is then subtracted from the power spectrum of the decoded speech. Its low computational overhead makes it ideal for real-time systems.
- PESQ Evaluation: The ITU-T P.862 PESQ standard was used to provide an objective score that mimics human auditory perception.
Results and Discussion
The study evaluated the performance of the enhancement method using standard speech samples (Test1.wav, Test2.wav, Test3.wav) across various simulated channel error rates (0% to 3%).
Quantitative Improvements
The application of spectral subtraction yielded a clear quantitative boost. For the primary test sample, the PESQ score increased from 3.046 to 3.625, representing a 19.01% improvement in objective quality.
Robustness to Channel Errors
A significant finding was the system's resilience to transmission errors. The enhanced signal at a 3% bit error rate (score: 2.378) outperformed the non-enhanced signal at a lower 2% error rate (score: 2.338). This suggests that spectral subtraction can effectively compensate for quality loss caused by volatile underwater channel conditions.
Visual Confirmation
Waveform analysis confirms the reduction of the "noise floor" in the intervals between spoken words, where stationary noise is most prominent. The identification of these quiet intervals is crucial for accurate noise estimation.
Conclusion
This research demonstrates that combining MELP coding with spectral subtraction provides a robust solution for underwater voice communication. By utilizing PESQ for objective evaluation, engineers can reliably measure improvements without the need for subjective human testing. While effective against stationary noise, future work should investigate adaptive techniques to handle the dynamic, non-stationary noise common in real-world maritime environments.
Summary updated