Robust and efficient pitch estimation using an iterative ARMA technique

Jung Ook Hong, Patrick J. WolfeInterspeech 2010

TL;DR

This paper introduces a novel iterative ARMA (AutoRegressive Moving Average) technique for robust and efficient pitch estimation in speech. By incorporating the harmonic structure of voiced speech and assuming smooth transitions in pitch over time, their method achieves higher accuracy and lower computational cost compared to existing state-of-the-art algorithms, especially in noisy conditions.

What Problem Does This Paper Address?

Accurately estimating the pitch (fundamental frequency) of voiced speech is crucial for many speech processing tasks, including speech synthesis, recognition, and speaker identification. Traditional methods often struggle with noisy speech, requiring careful initialization and potentially suffering from multiple local minima in their optimization processes. Existing iterative techniques, while improving robustness, can be computationally expensive. This paper aims to bridge this gap by developing a pitch estimation method that is both accurate and computationally efficient, particularly in challenging, noisy environments.

Methods

The core of this paper is an iterative ARMA technique adapted for pitch estimation. The authors build upon Quinn's generalized algorithm for estimating frequencies in time series. Their key innovations include:

  • Iterative ARMA for Single Frequency: They model a single sinusoidal component with additive noise as an ARMA(2,2) process. An iterative algorithm is then used to estimate the frequency parameter by minimizing a penalized likelihood function.
  • Multi-frame Time Series and Penalized Likelihood: To handle the sequential nature of speech, they extend the single-frame estimation to multi-frame analysis. Crucially, they incorporate a Markovian assumption, meaning the pitch at the current frame is influenced by the pitch in the previous frame. This is achieved by adding a penalty term to the likelihood function that favors smooth pitch transitions.
  • Incorporating Harmonic Structure: Recognizing that voiced speech is composed of harmonic partials (multiples of the fundamental frequency), they extend the ARMA model to account for multiple sinusoidal components (ARMA(2K, 2K)). This allows for a more direct estimation of the fundamental frequency by leveraging the harmonic relationships.
  • Empirical Parameter Selection: They propose an empirical method for setting the smoothing parameter ($\lambda$) based on the signal-to-noise ratio (SNR), further enhancing robustness.

Datasets

The KED-TIMIT database was used for evaluation. This database contains 452 phonetically rich utterances from a male American English speaker, with simultaneous electroglottograph (EGG) recordings and hand-specified phoneme labels. The "ground truth" pitch was derived from the EGG signals. The utterances were corrupted with Gaussian noise to create various SNR levels (-5 dB to 15 dB).

Key Insights and Ideas

  • Harmonicity is Key: Explicitly modeling the harmonic structure of voiced speech, rather than just treating it as a single sinusoid, significantly improves pitch estimation accuracy.
  • Smoothness Pays Off: The Markovian assumption, enforcing smooth pitch transitions between frames, acts as a powerful regularizer, preventing erratic estimates, especially in noisy conditions.
  • ARMA's Efficiency: The iterative ARMA framework provides a computationally efficient way to estimate multiple frequency components, making it practical for real-time applications.
  • SNR-dependent Regularization: Adapting the smoothing parameter based on the SNR allows the algorithm to be more flexible and robust across different noise levels.

Key Results

The proposed algorithm, named i-ARMAP, demonstrated superior performance compared to existing methods like RAPT and MAP (Maximum A Posteriori) in terms of both Root Mean Square Error (RMSE) and Gross Error Rate (GE).

  • Accuracy: i-ARMAP achieved lower RMSE and GE, particularly in low SNR environments. For instance, at -5 dB SNR, i-ARMAP had a gross error rate of 16.9%, significantly better than RAPT (40.1%) and MAP (22.3%).
  • Efficiency: The computational cost was dramatically lower. i-ARMAP required approximately 22 minutes to process the entire dataset across five SNR levels, while MAP and MAP* took 149 and 518 hours, respectively. This represents a substantial speedup, making it a practical choice for real-time applications.
  • Qualitative Findings: Visualizations showed i-ARMAP successfully tracking the true pitch, even when other algorithms struggled or produced erratic estimates in noisy segments.

Limitations and Future Work

  • Computational Cost vs. MAP: While i-ARMAP is significantly faster than MAP, the paper notes that MAP can achieve slightly better accuracy if its frequency grids are very finely tuned, at a much higher computational cost.
  • Unvoiced/Voiced Detection: The experiments removed unvoiced/voiced detection for RAPT to ensure fair comparison. Future work could integrate robust unvoiced/voiced detection into i-ARMAP.
  • Parameter Tuning: While an empirical method for setting $\lambda$ is proposed, further research into optimal parameter selection strategies could be beneficial.
  • Generalization: The paper focuses on voiced speech. Extending the technique to handle unvoiced segments and transitions would be a valuable next step.

Summary updated

BibTeX
@inproceedings{hong2010robust,
  author = {Jung Ook Hong and Patrick J. Wolfe},
  title = {{Robust and efficient pitch estimation using an iterative ARMA technique}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure