TL;DR
This paper introduces a novel maximum likelihood (ML) approach for detecting "moments of maximum excitation" (MME) in speech, crucial for accurate speech analysis and synthesis. By modeling speech as a pulsed excitation convolved with a minimum-phase impulse response, their method automatically estimates MME locations and amplitudes, outperforming existing methods in certain reconstruction quality metrics.
What Problem Does This Paper Address?
Accurate detection of "moments of maximum excitation" (MME) in speech is fundamental for many speech processing tasks, including synthesis, coding, and even pathology detection. These MMEs, often closely related to glottal closure instants (GCIs), represent points of significant energy release in voiced speech. Existing methods for MME detection often rely on ad-hoc procedures that can be inconsistent across different speech styles and require manual adjustments. This lack of robustness and automaticity hinders the development of high-quality, general-purpose speech processing systems. The authors aim to provide a more consistent and automatic method for MME detection by leveraging a probabilistic model.
Methods
The core of the paper is a maximum likelihood (ML) estimation framework for MME detection. The authors model speech s(n) as the convolution of a pulsed excitation signal e(n) (where pulses occur at MMEs) with a time-varying minimum-phase impulse response h(n), plus additive noise w(n).
The proposed method involves several key steps:
- Speech Model: A probabilistic model is established where the likelihood of the speech signal
sis defined given the excitatione, impulse responseH, and noise covarianceR. - Maximum Likelihood Estimation: By assuming a specific form for the excitation signal (a train of pulses) and a normally distributed error, the authors derive a log-likelihood function. They then optimize this function to find the optimal pulse amplitudes and positions that best explain the observed speech signal. This involves iteratively estimating pulse amplitudes and then their positions.
- Pulse Elimination: To refine the detected MMEs, a pulse elimination step is introduced. This algorithm removes MMEs that are too close to each other (below a minimum distance threshold) or those with lower likelihood scores, especially if they are likely to be artifacts from unvoiced regions.
- Position Refinement: The detected MME positions are further refined by re-optimizing them within a local neighborhood to maximize the likelihood function.
- Fo-based Post-processing: Finally, fundamental frequency (Fo) information is used to remove MMEs from unvoiced regions and to detect and correct potential half-pitch errors.
The novelty lies in the application of ML to directly estimate MME locations and amplitudes within this speech production model, aiming for a more principled and automatic approach compared to heuristic methods.
Datasets
The experiments were conducted using 200 utterances from four speakers (two male, two female) selected from proprietary databases and the clean portion of the Hurricane Challenge dataset. The sampling frequency was 16 kHz. No new benchmarks were introduced; existing algorithms were used for comparison.
Key Insights and Ideas
- Probabilistic Foundation for MME: Framing MME detection as a maximum likelihood problem provides a more robust and theoretically grounded approach than ad-hoc methods.
- Energy Concentration at MME: The core assumption that speech concentrates more energy at MMEs is key. This allows the ML framework to prioritize locations with higher "impact" on the reconstructed signal.
- Unified Framework for Amplitude and Position: The ML approach naturally estimates both the amplitude and position of excitation pulses, offering a more complete representation of the excitation source.
- Automaticity: The algorithm is designed to be fully automatic, reducing the need for manual tuning, which is a significant advantage for practical applications.
- Importance of Post-processing: The inclusion of pulse elimination and Fo-based refinement steps highlights the necessity of multi-stage processing to achieve high-quality MME detection.
Key Results
The paper evaluates the proposed algorithm by using the detected MMEs for speech analysis and reconstruction via complex cepstrum. The primary metric used is the segmented signal-to-noise ratio (SNR-seg) of voiced regions.
- The proposed method achieved better SNR-seg results for speakers F1 and M2 compared to ESPS and DYPSA.
- For speakers F2 and M1, DYPSA performed slightly better, but overall, the quality of the reconstructed speech was reported as similar across methods.
- The authors emphasize that the proposed method's advantage lies in its simplicity, consistency, and fully automatic nature.
Limitations and Future Work
- Limitations: The paper acknowledges that the performance can vary across different speakers and voice types, as seen in the mixed results for F2 and M1. While the method is generally better than ESPS, it's not universally superior to all existing methods like DYPSA. The reliance on Fo information for post-processing means its effectiveness might be impacted in very noisy conditions where Fo estimation is poor.
- Future Work: The authors intend to perform a quantitative evaluation of the proposed method in the future, suggesting that the current evaluation, while informative, might be expanded. They also aim to further explore the quantitative aspects of their approach.
Summary updated