Efficient multipulse approximation of speech excitation using the most singular manifold

Vahid Khanagha, Khalid DaoudiInterspeech 2012

TL;DR

This paper introduces a novel, computationally efficient method for approximating speech excitation signals by identifying "singular" points in the waveform. This "Most Singular Manifold" (MSM) approach directly extracts excitation locations from the audio signal, bypassing traditional predictor-based methods and achieving comparable perceptual quality to existing techniques while being significantly faster.

What Problem Does This Paper Address?

Efficiently representing the excitation signal (the "buzz" or "hiss" that drives speech production) is crucial for speech coding and synthesis. Traditional methods like MultiPulse Excitation (MPE) rely on complex iterative processes to find the optimal locations and amplitudes of these excitation pulses. These methods often depend on accurate models of the vocal tract (like Linear Predictive Coding - LPC), which can be computationally expensive and introduce their own complexities. The challenge is to find a way to identify these crucial excitation points directly from the speech signal itself, in a computationally efficient manner, without relying heavily on pre-existing vocal tract models.

Methods

The core innovation is the use of the "Most Singular Manifold" (MSM) concept, derived from the Microcanonical Multiscale Formalism (MMF). The MMF analyzes signals by looking at their local predictability across different scales. The MSM identifies points in the signal that are least predictable, which the authors hypothesize correspond to the impulsive excitations of speech (like glottal closure instances).

Here's a breakdown:

  • Multiscale Singularity Exponents (SE): The paper defines a measure of local predictability and estimates singularity exponents (SE) at each point in the speech waveform. These exponents quantify how predictable a signal is at a specific time.
  • Most Singular Manifold (MSM): The MSM is defined as the set of points with the lowest SE values. These are the "least predictable" points, believed to mark significant excitation events.
  • MSM-based Excitation Approximation: Instead of the iterative search in MPE, the authors directly use the MSM to find pulse locations. They select a fixed number of samples from the MSM (those with the lowest SE) as potential pulse locations.
  • Amplitude Optimization: Once the pulse locations are identified via the MSM, the amplitudes are optimized using the second stage of the standard MPE algorithm, which is a joint optimization process.
  • Efficiency Gain: The computational savings come from replacing the entire first stage of MPE (iterative pulse search) with a simpler sorting of SE values to identify the MSM.

Datasets

  • TIMIT Database: The evaluation was performed using approximately 1 hour of clean speech randomly selected from the TIMIT database. This dataset is known for its diversity in speakers (gender, accent, age).

Key Insights and Ideas

  • Direct Waveform Analysis: The most significant insight is that excitation pulse locations can be reliably identified directly from the speech waveform using a geometric, multiscale approach, independent of explicit vocal tract models.
  • Singularity as a Proxy for Excitation: The paper establishes a strong link between "least predictable" points (MSM) and significant speech excitation events (like Glottal Closure Instances - GCI). This is a novel way to frame the problem of excitation modeling.
  • Computational Efficiency: By replacing the iterative search of MPE with a sorting operation, the MSM method offers a substantial speedup, making it more practical for real-time applications or scenarios with limited computational resources.
  • Perceptual Quality Equivalence: Despite the computational gains, the perceptual quality of speech reconstructed using the MSM method is shown to be comparable to that of the established MPE method.

Key Results

  • Efficiency: The MSM-based algorithm is significantly more efficient than MPE. For example, when comparing computational processing times, the MSM method with 10 pulses per 20ms frame has a Relative Computation Time (RCT) of 9.6%, while MPE has an RCT of 71.9%.
  • Reconstruction Quality (MSE): The MSM method achieves a satisfactory performance in terms of mean squared error (MSE), though slightly less accurate than MPE. However, it outperforms a recent Compressed Sensing (CS) based method when the number of pulses (K) is low (e.g., K < 10). For K=8, the MSM method achieved an average normalized error of 0.55, compared to 0.68 for the CS method.
  • Perceptual Quality (CMOS): Subjective listening tests and the composite measure of speech quality (CMOS) indicate that the perceptual quality of speech reconstructed by the MSM method is very close to that of MPE, and roughly the same for K=20.

Limitations and Future Work

  • MSE vs. Perceptual Quality: The authors acknowledge that the MSM method is penalized in terms of MSE because MPE's iterative process directly minimizes this metric. However, they argue that MSE is not the sole determinant of perceptual quality, where their method performs well.
  • GCI Identification: While the MSM points are shown to be located around GCIs, the authors explicitly state that their primary goal is not GCI identification itself, and they have not yet investigated this aspect thoroughly.
  • Sparse Predictor Integration: The authors suggest exploring the use of sparse predictors (like those used in CS) in conjunction with their MSM approach, as their method is independent of the predictor.
  • Automatic GCI Identification: A promising future direction is to develop an automatic GCI identification method based on their findings, which could unlock applications in areas like closed-phase LPC.

Summary updated

BibTeX
@inproceedings{khanagha2012efficient,
  author = {Vahid Khanagha and Khalid Daoudi},
  title = {{Efficient multipulse approximation of speech excitation using the most singular manifold}},
  booktitle = {Interspeech},
  year = {2012},
}

Loading related content...

Figure