Abstract
Laryngectomees often rely on an electrolarynx (EL) to produce speech, but the resulting audio is typically characterized by distracting mechanical noise leakage and a monotonic, robotic pitch. This paper proposes a hybrid speech enhancement system that integrates spectral subtraction (SS) with statistical voice conversion (VC). By utilizing SS to preserve the speaker's original articulation and VC to re-synthesize natural excitation parameters, the system significantly improves speech naturalness while maintaining high intelligibility. Experimental results demonstrate that the proposed method, particularly when using continuous $F_0$ modeling, outperforms traditional noise reduction and pure voice conversion techniques.
Introduction
Patients who undergo a laryngectomy lose their vocal folds, necessitating the use of assistive devices like the electrolarynx. While functional, EL speech suffers from two primary degradations: noise leakage from the device's mechanical vibration and a lack of natural prosody due to the device's constant, flat excitation. Previous attempts to improve EL speech have generally focused on either noise reduction or voice conversion. However, noise reduction alone fails to address the robotic quality, and pure voice conversion often introduces artifacts that severely degrade intelligibility. This research seeks a hybrid solution that cleans the signal while restoring natural human-like pitch variations.
Methodology
The proposed system employs a dual-path hybrid approach to address articulation and excitation separately:
1. Spectral Subtraction (SS) for Articulation
To maintain high intelligibility, the system preserves the spectral parameters of the original EL speech. Because laryngectomees typically retain their ability to articulate clearly with their mouth and tongue, these spectral shapes are highly informative. SS is used to estimate and subtract the mechanical leakage noise of the EL device, resulting in a cleaner version of the speaker's original articulation.
2. Statistical Voice Conversion (VC) for Excitation
To address the robotic prosody, the system replaces the flat mechanical buzz with predicted natural excitation. A Gaussian Mixture Model (GMM) is trained to map the speaker's spectral features to natural excitation parameters, specifically the fundamental frequency ($F_0$) and aperiodicity (breathiness).
3. Continuous $F_0$ Modeling
A key innovation in this work is the use of continuous $F_0$ patterns. Standard voice conversion often fails at the "Unvoiced/Voiced" (U/V) transition, leading to robotic glitches. By interpolating pitch through silent or unvoiced segments during training, the system produces smoother, more natural transitions in the converted speech.
Experimental Setup
The study utilized a parallel dataset of 50 phoneme-balanced sentences recorded by a male laryngectomee (source) and a non-disabled male speaker (target). Evaluation was conducted via 5-fold cross-validation, using 40 sentences for training and 10 for testing. The performance was assessed through objective metrics and subjective Mean Opinion Score (MOS) tests with five listeners.
Results and Discussion
The hybrid approach demonstrated a superior balance between clarity and naturalness compared to baseline methods:
- Intelligibility: The hybrid methods (SS+VC and SS+VC+CF0) maintained high MOS scores near 4.0, comparable to the original EL speech. In contrast, pure VC methods saw a significant drop in intelligibility (below 3.0) due to synthesis artifacts.
- Naturalness: While original EL speech and SS-only speech scored poorly (~1.0), the proposed hybrid method with continuous $F_0$ (SS+VC+CF0) achieved a score of nearly 3.5, a substantial improvement in prosodic quality.
- Error Reduction: The continuous $F_0$ model nearly eliminated Voiced-to-Unvoiced errors, reducing them from 4.42% in standard VC to just 0.09%.
Conclusion
By combining the strengths of spectral subtraction and statistical voice conversion, this hybrid framework effectively addresses the primary limitations of electrolaryngeal speech. The preservation of original spectral features ensures intelligibility, while the statistical prediction of continuous pitch contours restores naturalness. Future work will focus on improving the recreation of unvoiced phonemes and optimizing the system for real-time, portable applications.
Summary updated