Abstract
Pitch estimation, or the identification of the fundamental frequency ($F_0$), is essential for various speech processing applications but remains challenging in high-noise environments. This paper introduces Temporally Accumulated Peak Spectrum (TAPS), a robust method designed to extract pitch from heavily corrupted speech. By focusing on spectral peaks and accumulating them over consecutive frames, TAPS exploits the temporal persistence of speech harmonics to suppress transient noise. Experimental results demonstrate that TAPS significantly outperforms conventional methods, particularly at low signal-to-noise ratios (SNR), reducing Gross Pitch Errors (GPE) by over 20% in extreme conditions.
Introduction
Identifying the fundamental frequency of a voice is a cornerstone of speech technology, from automated recognition systems to assistive hearing devices. While human listeners can isolate a speaker's pitch in a noisy room with ease, algorithmic pitch tracking often fails when background noise is introduced. Standard techniques, such as time-domain autocorrelation or frequency-domain cepstral analysis, are highly sensitive to noise-induced distortions. In environments with low SNRs, noise peaks can be mistaken for speech harmonics, leading to "Gross Pitch Errors" (GPE), where the estimated pitch is off by an octave or more.
The TAPS Approach
The proposed Temporally Accumulated Peak Spectrum (TAPS) method addresses these challenges through a two-stage process designed to make the speech signal's harmonic structure "emerge" from the noise.
1. Peak Spectrum Extraction
Rather than processing the entire magnitude spectrum, the algorithm identifies local maxima (peaks) and discards the rest of the spectral information. This non-linear filtering step effectively removes a large portion of low-level, broadband noise that does not correspond to harmonic peaks.
2. Temporal Accumulation
The core innovation of TAPS is the accumulation of these peak spectra over a short window of consecutive frames (typically $K=4$, or approximately 60ms). This strategy relies on the observation that speech harmonics are relatively stable over short durations, whereas noise peaks are typically random and transient. By summing the peak spectra, the consistent harmonic components of speech reinforce one another, while the inconsistent noise peaks remain scattered and weak.
Experimental Evaluation
The TAPS method was evaluated using the CSLU VOICES dataset, featuring 100 utterances from a balanced group of male and female speakers. To simulate real-world conditions, the researchers utilized the NOISEX-92 database (including white, speech, and babble noise) and a custom Artificial Non-Stationary Noise (NS-White) that varied in variance every 8ms.
Performance was measured against standard Cepstrum (CEP) and Weighted Autocorrelation (WAutoC) methods, focusing on the Gross Pitch Error (GPE) rate across various SNR levels.
Results and Discussion
The TAPS-AutoC method demonstrated superior robustness across all tested noise types. A key finding was the "Goldilocks" window for accumulation: while too few frames fail to suppress noise, too many frames can blur the natural pitch variations (prosody) of the speaker. A window of $K=4$ was found to be optimal.
At a challenging SNR of -7.5 dB, TAPS reduced the average GPE by 22.8% compared to the Weighted Autocorrelation method. Unlike conventional algorithms whose error rates spike sharply as noise increases, the TAPS error rate grows gracefully, indicating high reliability for practical applications.
Conclusion
The TAPS method provides a powerful framework for pitch estimation in adverse acoustic conditions. By leveraging the temporal "stickiness" of speech harmonics, it effectively filters out "slippery" background noise. While the accumulation process introduces a slight smoothing effect on rapid pitch dynamics, the gains in robustness at low SNRs are substantial. Future work may involve integrating TAPS with statistical models like Hidden Markov Models (HMMs) to further refine pitch tracking in dynamic environments.
Summary updated