Abstract
In this paper we present a statistical framework to demodulate speech resonances, which are modeled as AM-FM signals. The first approach utilizes bandpass filtering and a standard demodulation algorithm which regularizes instantaneous amplitude and frequency estimates. The second employs particle filtering techniques to allow temporal variations of the parameters that are connected with spectral characteristics of the analyzed signal. Results are presented on both synthetic and real speech signals and improved performance is demonstrated. Both approaches appear to cope quite satisfactorily with the nonstationarity of speech signals.
Introduction
Traditional speech processing often relies on linear models that assume speech is stationary over short windows. However, speech is inherently non-stationary; its resonances (formants) change rapidly in both amplitude and frequency. To capture this, researchers use AM-FM models, where speech is seen as a sum of modulated components. While the Model Based Demodulation Algorithm (MBDA) was a step forward, it relied on Extended Kalman Filtering (EKF) which is notoriously sensitive to initialization and abrupt spectral shifts. This paper introduces two powerful extensions to the statistical framework to make these trackers more resilient.
Methodology
Gabor-ESA + MBDA Hybrid
The first proposed improvement integrates the Energy Separation Algorithm (ESA)—a deterministic method for finding amplitude and frequency—into the state-space model. Gabor bandpass filters are used to isolate resonances, and the ESA output serves as an additional observation or constraint for the EKF. This prevents the statistical tracker from wandering too far from the actual signal characteristics, combining the speed of deterministic algorithms with the tracking capabilities of stochastic ones.
Particle Filtering for Time-Varying Parameters
Standard models often treat parameters like formant bandwidth and power as constants. In reality, these change during an utterance. The authors allow these parameters to vary over time according to a first-order Markov process. Because this makes the model highly non-linear and high-dimensional, the authors employ Particle Filtering (Sequential Importance Sampling). This allows for the joint estimation of both the state of the speech signal and the parameters governing its evolution, overcoming the limitations of the EKF in highly non-linear environments.
Results and Discussion
The researchers evaluated their methods using synthetic AM-FM signals and real speech analysis (e.g., the word "yell"). The enhanced algorithms demonstrated significant performance gains over the standard MBDA.
Quantitative Performance on Synthetic Signals
In synthetic tests, the proposed hybrid algorithm achieved a substantial reduction in error. As shown in Figure 1, the Normalized Mean Square Error (NMSE) for amplitude estimation dropped from 1.4337 (standard MBDA) to 0.2537. Similarly, for frequency estimation, the NMSE was reduced from 0.067 to 0.038.
Robustness in Real-World Speech
When analyzing the word "yell," the standard MBDA produced erroneous variations and even impossible negative frequency estimates for the first formant. In contrast, the Particle Filtering approach successfully tracked all four resonances smoothly, even in areas of high spectral density.
Conclusion
This work demonstrates that combining deterministic signal processing with stochastic tracking significantly improves the robustness of AM-FM speech demodulation. While Particle Filtering is more computationally demanding, it provides a much more accurate representation of the probability distributions involved in speech tracking. Future work will focus on optimizing these methods for real-time applications and integrating them into automatic speech recognition (ASR) systems.
Summary updated