Robust Pitch Estimation Using l1-regularized Maximum Likelihood Estimation

Tan Lee 李丹, Feng Huang, Tan LeeInterspeech 2012 · DOI

Introduction

Fundamental frequency ($F_0$) estimation, or pitch detection, is a critical component in speech processing applications ranging from speaker recognition to prosody analysis. While human listeners can easily extract pitch in noisy environments, computational methods often struggle when the signal-to-noise ratio (SNR) drops. Traditional time-domain methods like the Autocorrelation Function (ACF) are prone to octave errors and noise interference. This paper presents a robust pitch estimation framework, TAPS-$l_1$-ML, which leverages sparse representation and prior knowledge of harmonic structures to maintain accuracy in adverse acoustic conditions.

Methodology

The proposed method shifts the focus from simple periodicity detection to a sparse reconstruction problem in the spectral domain.

Temporally Accumulated Peak Spectrum (TAPS)

To enhance the robustness of spectral features, the authors introduce the Temporally Accumulated Peak Spectrum (TAPS). Instead of relying on a single frame, TAPS aggregates spectral peaks over several consecutive frames (typically four). This process emphasizes stable harmonic components while suppressing transient, irregular noise peaks, providing a cleaner representation of the underlying pitch.

Dictionary-Based Sparse Reconstruction

The core of the algorithm involves modeling the observed TAPS signal as a sparse linear combination of pre-learned exemplars. A dictionary $A$ is constructed from clean speech, where each column represents a prototypical harmonic structure for a specific pitch. The estimation problem is formulated as:

$$y = Ax + v$$

where $y$ is the observed TAPS, $x$ is the sparse weight vector, and $v$ represents noise. Because a single voice should only correspond to one or two pitch exemplars at a time, the system forces the weight vector $x$ to be sparse using $l_1$ regularization.

$l_1$-regularized Maximum Likelihood Estimation

Unlike standard sparse coding that assumes zero-mean Gaussian noise, the authors account for the non-negative nature of peak spectra. Because magnitude spectra are inherently biased, the noise $v$ is modeled with a non-zero mean. By applying $l_1$ regularization to $x$, the system identifies the most likely pitch exemplar that explains the observed signal. The final pitch is determined by the index of the highest-weighted exemplar in the sparse vector $x$.

Experimental Results

The performance of TAPS-$l_1$-ML was evaluated using the CSLU-VOICES corpus against established baselines such as RAPT and PEFAC.

Gross Pitch Error (GPE)

Gross Pitch Error measures the percentage of frames where the estimated pitch deviates significantly from the ground truth. In white noise conditions, TAPS-$l_1$-ML demonstrated superior robustness. At an SNR of -10 dB, the proposed method achieved a significantly lower GPE compared to traditional autocorrelation and spectral methods, reducing the error by 27.6% compared to weighted autocorrelation.

Figure 3: Gross Pitch Error (GPE) rate across different SNR levels in white noise.

Fine Pitch Error (FPE)

Fine Pitch Error assesses the precision of the estimate when the pitch is correctly identified. A key strength of the TAPS-$l_1$-ML approach is its stability; the RMS of the FPE remains relatively constant even as noise levels increase. This suggests that the dictionary-based approach effectively "locks onto" the correct harmonic structure once the noise is filtered through the sparse reconstruction process.

Figure 6: RMS of Fine Pitch Error (FPE) in white noise, showing the stability of the TAPS-l1-ML method.

Conclusion

By framing pitch estimation as a sparse reconstruction problem and incorporating non-zero mean noise modeling, the TAPS-$l_1$-ML method offers a significant leap in robustness for noisy speech processing. While computationally more demanding than simple ACF, its ability to maintain high accuracy at SNRs as low as -15 dB makes it a powerful tool for real-world applications. Future work may explore more complex noise models, such as Gaussian Mixture Models, to further enhance performance in non-stationary noise environments.

Summary updated

BibTeX
@inproceedings{2012robust,
  author = {Tan Lee 李丹 and Feng Huang and Tan Lee},
  title = {{Robust Pitch Estimation Using l1-regularized Maximum Likelihood Estimation}},
  booktitle = {Interspeech},
  year = {2012},
  doi = {10.21437/interspeech.2012-137},
}
Figure