All-pole model estimation of vocal tract on the frequency domain

Luis Weruaga, Amar Al-KhayatInterspeech 2006

TL;DR

This paper introduces a novel method for estimating vocal tract spectral characteristics directly from frequency domain data, bypassing traditional time-domain analysis. This frequency-domain approach offers a more direct and potentially robust way to model speech, especially for challenging scenarios like non-stationary vocal tracts.

What Problem Does This Paper Address?

Speech analysis often relies on linear prediction (LP) or autoregressive (AR) models to represent the vocal tract's spectral envelope. The dominant method, Yule-Walker, operates in the time domain and estimates AR coefficients. While effective, this indirect approach has limitations. For instance, parameters derived from time-domain analysis (like AR coefficients) are often translated to frequency-domain representations (like Line Spectral Frequencies or poles) for stability checks and better interpolation. This paper argues that directly estimating these frequency-domain parameters, specifically the poles of the all-pole model, from spectral samples could offer advantages. The challenge is to develop a robust and accurate method for this direct frequency-domain estimation, especially when dealing with voiced speech where harmonic structures are prominent.

Methods

The core of the proposed method is an "analysis-by-synthesis" (ABS) framework. Instead of estimating time-domain coefficients and then deriving poles, this approach directly fits an all-pole model to the spectral envelope of the harmonics in voiced speech. The process involves:

  1. Harmonic Sampling: Exploiting the fact that voiced speech can be approximated as a periodic impulse train exciting the vocal tract. The spectrum of such a signal consists of samples of the vocal tract's frequency response at harmonic frequencies.
  2. Logarithmic Spectral Distance: The paper defines a risk functional based on the difference between the logarithm of the squared spectral samples (from the speech signal) and the logarithm of the squared spectral response of an all-pole model. This log-based measure is considered more perceptually relevant than linear spectral differences.
  3. Numerical Optimization: The problem is framed as minimizing this risk functional with respect to the all-pole model parameters (log-power level and poles). A Newton-Raphson-like gradient algorithm is employed for this minimization. The update rule for the poles is intuitively described as "pulling" them towards the unit circle based on the error and their distance from the harmonic frequencies.

Datasets

The paper primarily uses synthetic signals generated to mimic voiced speech with periodic pulse excitation. The fundamental frequency (pitch) is varied to test the method's robustness. Experiments are also mentioned with real speech segments, but specific datasets are not detailed.

Key Insights and Ideas

  • Direct Frequency-Domain Estimation: The most significant insight is the feasibility and potential benefits of estimating vocal tract poles directly from spectral samples, rather than relying on indirect time-domain methods.
  • Perceptual Relevance of Log-Spectral Distance: Using the logarithm of spectral energy as the error metric is highlighted as being more aligned with human perception of spectral differences compared to linear spectral distance.
  • Robustness to Pitch Variation: The proposed method demonstrates robustness to changes in fundamental frequency, a key advantage over some existing harmonic coding techniques.
  • Potential for Non-Stationary Scenarios: The ABS framework is suggested to be adaptable to non-stationary vocal tracts or excitations, a significant limitation of many current methods.

Key Results

  • Synthetic Signal Performance: Comparative results on synthetic signals show the proposed method's excellent performance. It accurately estimates the all-pole spectral envelope, outperforming a spline-based interpolation method used in harmonic coding, especially at higher fundamental frequencies.
  • Pitch Independence: The estimated spectral envelope is not significantly affected by the fundamental frequency of the excitation signal.
  • Model Order Indication: The evolution of the poles during the analysis suggests that the actual all-pole model order might be lower than initially assumed, as some poles tend towards zero.
  • Convergence: The algorithm exhibits rapid convergence to one of the possible pole permutations, and the risk functional is shown to be convex around the minimum.

Limitations and Future Work

  • Real Speech Challenges: While experiments with real speech were conducted, the all-pole fitting was not always satisfactory. The authors attribute this to potential non-stationarity within the speech segment (both pitch and vocal tract) and the limitations of the harmonic spectrum in strictly delineating an all-pole system.
  • Future Research Directions:
    • Modifying the ABS technique to handle non-stationary speech segments.
  • Investigating the integration of this frequency-domain estimation method into Harmonic Coding.
  • Further exploration of the non-stationary nature of real speech and how to address it with the proposed technique.

Summary updated

BibTeX
@inproceedings{weruaga2006all,
  author = {Luis Weruaga and Amar Al-Khayat},
  title = {{All-pole model estimation of vocal tract on the frequency domain}},
  booktitle = {Interspeech},
  year = {2006},
}
Figure