TL;DR
This paper introduces a novel all-pole modeling technique for wide-band speech called Weighted-Sum Line Spectrum Pair (WLSP). WLSP improves upon traditional linear prediction by better modeling the spectral envelope, particularly the formant structure of vowels, leading to more accurate speech representations.
What Problem Does This Paper Address?
Accurate modeling of the speech spectrum is crucial for many speech processing applications, especially in low bit-rate speech coding. Traditional linear prediction (LP) methods, particularly the autocorrelation method, are widely used due to their computational efficiency and ability to produce stable all-pole models. However, a known limitation of conventional LP is its tendency to poorly model certain formant structures in wide-band speech, especially when formants are closely spaced. This can lead to less perceptually accurate spectral representations. The authors aim to address this by developing a new all-pole modeling technique that offers improved spectral accuracy, particularly for the perceptually important formant regions.
Methods
The paper introduces the Weighted-Sum Line Spectrum Pair (WLSP) technique. Unlike conventional LP, which defines an all-pole filter by minimizing prediction error based on the autocorrelation function up to a certain lag, WLSP utilizes the Line Spectrum Pair (LSP) decomposition in a novel way.
Here's a breakdown of the WLSP approach:
- LSP Decomposition: The standard LP predictor polynomial is decomposed into two components: a symmetric polynomial $P(z)$ and an antisymmetric polynomial $Q(z)$. These polynomials are related to the roots of the predictor polynomial, known as LSPs, which are known to be well-behaved for interpolation.
- Weighted Sum: WLSP defines a new predictor polynomial $D(z, \lambda)$ as a weighted sum of $P(z)$ and $Q(z)$: $D(z, \lambda) = \lambda P(z) + (1 - \lambda) Q(z)$. The parameter $\lambda$ controls the weighting.
- Autocorrelation Matching: A key property of $D(z, \lambda)^{-1}$ is that its autocorrelation function matches the input signal's autocorrelation function exactly for lags 0 to $m-1$ (where $m$ is the prediction order). This is similar to conventional LP. However, WLSP sacrifices the exact match at lag $m$ to achieve a more accurate match in the higher autocorrelation range.
- Optimization: The WLSP algorithm involves:
- Computing an LP polynomial of order $m-1$ for the input signal.
- Constructing the LSP polynomials $P(z)$ and $Q(z)$ from this $m-1$ order predictor.
- Optimizing the weighting parameter $\lambda$ by minimizing the absolute difference between the normalized autocorrelations of the input signal and the inverse WLSP filter, over an adaptive window of lags (from $m$ up to the first peak of the autocorrelation function beyond $m$).
This approach allows WLSP to leverage information from the autocorrelation function beyond the prediction order $m$, which is not utilized by conventional LP.
Datasets
Not discussed in the paper. The experiments were conducted on two Finnish vowels (/a/ and /o/) produced by six speakers (four male, two female).
Key Insights and Ideas
- Beyond Lag $m$: The core insight is that by relaxing the strict autocorrelation matching at lag $m$, WLSP can achieve a better overall model of the speech spectrum. This is achieved by incorporating information from higher lags of the autocorrelation function.
- Improved Formant Modeling: The paper demonstrates that WLSP is particularly effective at modeling the formant structure of vowels. This is crucial because formants are key acoustic features that define vowel identity.
- Sharper Spectral Valleys: WLSP tends to create deeper spectral valleys between formant peaks compared to conventional LP. This suggests a more precise delineation of resonant frequencies.
- LSP as a Tool, Not Just for Quantization: The paper highlights a novel use of LSPs. Instead of solely using them for efficient quantization of LP coefficients (as in speech coding), WLSP employs them as a computational tool to construct a new type of all-pole filter with improved modeling properties.
Key Results
- Formant-Valley Difference: In experiments on wide-band speech (8 kHz bandwidth, 16 kHz sampling), WLSP consistently produced larger level differences between formant peaks and spectral valleys compared to conventional LP (with $m=16$). The average difference ($\Delta L_{diff}$) was 1.76 dB for the first formant and 1.84 dB for the second formant. This indicates WLSP better separates resonant peaks.
- Formant Detection: WLSP modeled a larger number of formants than LP in 7 out of 12 analyzed cases and modeled one fewer formant in only one case. This suggests WLSP can reveal more spectral detail.
- Qualitative Improvement: Figures 4-6 show that WLSP spectra more closely track the FFT spectrum of the input signal, particularly in the regions of formants and valleys, compared to the LP spectra.
Limitations and Future Work
- Adaptive Window Length: The paper notes that the procedure used in this study is slightly different from a previous work [8] in that it uses an adaptive length for the optimization window. This adaptation is not fully detailed but is presented as an improvement.
- Limited Data: The evaluation was performed on a small set of vowels and speakers. Further evaluation on a larger and more diverse dataset would be beneficial.
- Future Work: The authors suggest that WLSP could be beneficial in applications where accurate spectral modeling is critical, such as speech synthesis or advanced speech coding. The potential for using WLSP in these areas is implied but not explored in depth.
Summary updated