TL;DR
This paper introduces a novel method for glottal inverse filtering that accurately estimates vocal tract and glottal source parameters, even when the phase information of the speech signal is corrupted. By leveraging the glottal formant, the technique offers a robust alternative to traditional methods that rely on precise timing information, paving the way for more reliable speech analysis and synthesis.
What Problem Does This Paper Address?
Estimating the vocal tract filter and the glottal source waveform is fundamental to many speech processing tasks, including speaker identification and speech synthesis. Glottal inverse filtering aims to achieve this separation. Traditional methods often rely on assumptions about the glottal source's time-domain shape and require precise knowledge of key timing instants, such as the glottal closure instant. However, real-world recording conditions can corrupt phase information, making these timing estimates unreliable. This paper addresses the challenge of performing accurate glottal inverse filtering and parameterization when this crucial timing information is unavailable or compromised.
Methods
The proposed method exploits the "glottal formant," a spectral characteristic related to the opening phase of the glottis, rather than relying on precise timing information. The core of the approach involves:
- Estimating Glottal Formant Parameters: The fundamental frequency ($f_0$) and the glottal formant parameters (center frequency $w_g$ and bandwidth $a$) are estimated from the speech signal.
- Generating Glottal Shape Candidates: Using the estimated glottal formant parameters and the principle of area balance (ensuring no baseline drift in the glottal flow), a discrete set of candidate glottal source shapes (based on the Liljencrants-Fant (LF) model) are generated. This is achieved by varying the pole position of an IIR filter approximation for the return and closed phases.
- Joint Source-Filter Estimation: Each candidate glottal shape is then used to perform inverse filtering on the speech frame. An autoregressive analysis is applied to the resulting signal to estimate the vocal tract filter parameters. The glottal shape candidate that yields the "best fit" (measured by minimizing the residual energy) is selected, and its parameters are refined using optimization.
This approach combines the flexibility of frequency-domain methods with the robustness of a well-defined glottal source model, while being resilient to phase corruption.
Datasets
The paper describes experiments using synthetic speech signals. These signals were generated by:
- Creating LF model source signals for male and female voices with specific fundamental frequencies.
- Modulating Gaussian noise to the source signal to simulate aspiration noise.
- Filtering the source signal using a vocal tract transfer function for the vowel /æ/.
- Introducing phase distortion using an all-pass filter to simulate the effect of a PC sound card.
- A more realistic scenario involved a modulated first formant during the open phase of the signal.
No specific publicly available datasets are mentioned; the experiments rely on synthetically generated data designed to mimic real-world challenges.
Key Insights and Ideas
- Glottal Formant as a Robust Cue: The glottal formant, a spectral feature, proves to be a more robust indicator for glottal source parameterization than precise time-domain timing information, especially in the presence of phase corruption.
- Bridging Time and Frequency Domains: The method effectively bridges the gap between time-domain glottal models and frequency-domain inverse filtering techniques, offering the benefits of both.
- LF Model Versatility: The Liljencrants-Fant (LF) model is highlighted as a versatile glottal flow model, capable of representing a wide range of glottal waveforms.
- Area Balance is Crucial: The principle of area balance is essential for generating valid LF model pulses, ensuring the physical plausibility of the estimated glottal source.
Key Results
The paper demonstrates that the proposed method can accurately parameterize the glottal source and vocal tract filter, even for phase-corrupted signals.
- Parameter Accuracy: Deviations of the estimated glottal shape parameters (open quotient $O_q$, asymmetry coefficient $a_m$, and return phase coefficient $Q_a$) from their true values were generally small, with mean deviations typically in the range of $10^{-3}$ to $10^{-2}$ and standard deviations in a similar range.
- Correlation: The correlation coefficient between the original and estimated normalized LF model pulse shapes was high, often above 0.99, indicating strong similarity.
- Robustness to Phase Corruption: The method showed good performance even when phase corruption was introduced, as evidenced by the results in Table 1 and Table 2.
- Impact of Realistic Scenarios: While the method remained effective, a slight degradation in accuracy was observed in the more realistic scenario involving a modulated first formant during the open phase, suggesting that unmodeled interaction effects can impact performance.
Limitations and Future Work
- Glottal Formant Estimation: The paper acknowledges that the robust estimation of glottal formant parameters from speech frames is critical and requires further research.
- Source Amplitude Estimation: The current method focuses on shape parameters and does not explicitly estimate the glottal source amplitude ($E_e$). Expanding the algorithm to include this would provide a more complete source description.
- Parameter Continuity: For running speech, ensuring the continuity of estimated parameters across frames is important. Smoothing parameter trajectories using algorithms like the Viterbi algorithm is suggested.
- Higher Fundamental Frequencies: The accuracy of the method was noted to decrease slightly at higher fundamental frequencies.
- Interactive Speech Model: The performance degraded somewhat in the more realistic interactive speech model, indicating that further work is needed to account for more complex speech production phenomena.
Summary updated