Estimating the voice source in noise

Gang Chen, Yen-Liang Shue, Jody Kreiman, Abeer AlwanInterspeech 2012

TL;DR

This paper introduces a novel, noise-robust algorithm for estimating the "voice source" – the fundamental excitation signal of speech. By converting high-speed glottal area recordings into glottal flow and using a codebook search approach, their method significantly improves accuracy over existing techniques, especially in noisy conditions, paving the way for more reliable speech analysis.

What Problem Does This Paper Address?

Understanding the voice source is crucial for many speech processing applications, from synthesis to clinical voice analysis. The voice source is the initial puff of air from the vocal folds that gets shaped by the vocal tract. Estimating this source signal accurately is challenging, particularly when the speech is contaminated with noise. Traditional methods often rely on inverse filtering the speech signal, which requires accurately estimating the vocal tract's characteristics (like formant frequencies). However, these formant estimation techniques are notoriously unreliable in noisy environments, leading to inaccurate voice source estimates. This paper aims to bridge this gap by proposing a method that doesn't heavily depend on precise formant estimation, making it more robust to noise.

Methods

The core of this work lies in a noise-robust automatic source estimation algorithm that builds upon a previous approach [2]. The key innovation is the use of a codebook search strategy. Here's a breakdown:

  1. Area-to-Flow Conversion: High-speed video recordings of the vocal folds provide glottal area waveforms. These are converted into glottal flow signals using the LeTalker software, which implements a three-mass vocal fold model. This conversion is important because glottal flow, not just area, is the true excitation signal.
  2. Modified Source Model: A flexible five-parameter model is used to represent the glottal source waveform. This model is a refinement of a previous one, simplifying calculations by redefining parameters like "speed of opening phase" (Sop) and "speed of closing phase" (Scp) to be amplitude-based rather than time-based. This optimization significantly speeds up computation.
  3. Codebook Search: A codebook of potential voice source waveforms is generated using the modified source model. The algorithm then searches this codebook to find the source entry that best matches the input acoustic signal. This search involves minimizing an error function that considers the difference between the input signal's harmonics and the model's harmonics, constrained by vocal tract characteristics.
  4. Noise Robustness: Crucially, the algorithm does not explicitly estimate formant frequencies beforehand. Instead, it jointly searches for the best combination of source and vocal tract parameters, making it inherently more resilient to noise that would otherwise corrupt formant estimation. The process involves two iterations of the search to refine the results.

Datasets

The study utilizes data collected from six subjects (three male, three female) who sustained the vowel /i/ for approximately 10 seconds. This data includes:

  • Synchronous audio and high-speed video recordings of vocal folds: This is the primary source for generating the reference glottal flow signals.
  • Synthesized speech with added Gaussian white noise: To evaluate the algorithm's robustness, the audio signals were corrupted with Gaussian white noise at three Signal-to-Noise Ratio (SNR) levels: 20 dB, 10 dB, and 5 dB.

The authors state that the data used is the same as in a previous study [2], allowing for direct comparison and evaluation of generalization.

Key Insights and Ideas

  • Glottal Area vs. Glottal Flow: The paper highlights that glottal area, while observable, is not a perfect proxy for glottal flow due to factors like air inertia. Converting area to flow using a vocal fold model is a critical step for accurate source estimation.
  • Codebook Search for Robustness: The codebook search approach, by jointly estimating source and vocal tract parameters, bypasses the problematic step of explicit formant estimation in noisy conditions. This is a significant departure from traditional inverse filtering methods.
  • Model Simplification for Speed: Redefining parameters in the source model (Sop and Scp) led to a substantial computational speedup (90% reduction), making the algorithm more practical.
  • High-Pitch and Pressed Phonation Challenges: The results indicate that OQ estimation error tends to be higher for high-pitched voices and pressed phonations, suggesting these are still challenging areas for accurate source modeling.

Key Results

The proposed algorithm demonstrates improved accuracy in estimating the voice source waveform, as measured by Mean Square Error (MSE), compared to previous methods and the commercial toolkit Aparat.

  • Overall Performance: In clean and noisy conditions, the proposed algorithm consistently achieved lower average MSE.
  • Noise Robustness: Under 5 dB SNR, the proposed approach showed a statistically significant improvement of 4.1% over Aparat and 2.1% over the previous method, confirming its robustness to noise.
  • Phonation Type and Pitch: While generally robust, the algorithm showed some limitations in accurately estimating OQ for high-pitched and pressed phonations, with higher errors observed for females.

Limitations and Future Work

  • Incomplete Glottal Closure: The current model lacks an effective way to model incomplete glottal closure, a phenomenon known to be physiologically important for voice production.
  • Noise Types: The study focused on Gaussian white noise. The authors suggest exploring the algorithm's performance with other noise types, such as babble noise.
  • Vowel Specificity: While the algorithm is presented as suitable for other vowels, the current evaluation data primarily consists of the vowel /i/. Further validation across a wider range of speech sounds would be beneficial.

Future work will focus on incorporating incomplete glottal closure into the source modeling and investigating the impact of different noise types.

Summary updated

BibTeX
@inproceedings{chen2012estimating,
  author = {Gang Chen and Yen-Liang Shue and Jody Kreiman and Abeer Alwan},
  title = {{Estimating the voice source in noise}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure