Abstract
We propose a pitch estimation method based on long frame harmonic model (LFHM) and short frame average correlation coefficient (SFACC). The long frame used here is three times as long as the short frame. Our approach consists of two main steps: pitch candidates extraction based on LFHM, and final pitch selection based on the SFACC. The long frame analysis for speech signal is able to achieve a higher harmonic resolution and ensures the inclusion of the true pitch in the pitch candidate pool. The SFACC based pitch selection method is robust to noise and keeps the short-time stationary features of speech. Experimental results show that our method outperforms several state-of-the-art algorithms under both clean and noisy acoustic environments.
Introduction
Pitch estimation, the process of identifying the fundamental frequency ($F_0$) of vocal cord vibration, is a critical component in speech processing applications such as speech synthesis, recognition, and hearing aids. While effective in quiet environments, traditional methods often struggle in noisy real-world conditions. Time-domain methods like autocorrelation can be misled by noise peaks, while frequency-domain harmonic models may lose the fundamental frequency peak in background noise. This paper presents a hybrid approach that leverages both long-frame frequency analysis and short-frame temporal correlation to achieve robust $F_0$ estimation.
Proposed Methodology
The proposed algorithm follows a multi-stage process to extract and refine pitch estimates, as illustrated in the system flowchart below.
1. Candidate Extraction via LFHM
The first stage employs a Long Frame Harmonic Model (LFHM). Unlike standard analysis windows of 25–30 ms, this method utilizes windows of 75–90 ms. This extended duration provides significantly higher harmonic resolution, allowing the true pitch peaks to remain distinguishable even in high-noise environments. To ensure the fundamental frequency is captured, higher-frequency peaks are "compacted" by dividing them by integer factors to populate a robust pool of potential pitch candidates.
2. Pitch Selection via SFACC
Once candidates are identified, the system switches to a Short Frame Average Correlation Coefficient (SFACC) for selection. Standard autocorrelation can be sensitive to amplitude fluctuations; SFACC instead measures the linear dependence between consecutive pitch periods. This approach maintains the short-time stationary features of speech while providing robustness against noise.
3. Birth-Death Pitch Tracking
To avoid jittery frame-by-frame decisions, a birth-death frequency tracking algorithm is implemented. This creates continuous pitch tracks over time. A final track is selected based on a confidence score that integrates track length and average correlation values, effectively filtering out isolated errors and pitch doubling artifacts.
The Advantage of Long Frame Analysis
The core innovation lies in the dual-window approach. Long windows are superior for resolving harmonics in noise, while short windows are necessary to capture the rapidly changing stationary features of speech.
As shown in the spectral analysis, the long frame (green) maintains distinct harmonic peaks in noisy conditions where the short frame (red) loses resolution. This ensures that the "true" pitch is always present in the initial candidate pool.
Experimental Results
The method was evaluated using the Keele Database (5.5 minutes of speech with laryngograph ground truth) and NOISEX-92 for realistic noise samples (white and babble noise).
- Noise Robustness: In challenging babble noise at -5 dB SNR, the proposed method achieved a Gross Pitch Error (GPE) of 30.41%, significantly outperforming industry standards like RAPT (58.40%) and YIN (57.35%).
- Precision: The Mean Fine Pitch Error (MFPE) remained consistently below 1 Hz across most conditions, demonstrating high accuracy once the correct pitch track is identified.
- Comparative Performance: While the SAFE algorithm showed slight advantages in extreme white noise, the proposed method proved superior in the more complex and common "cocktail party" scenario of babble noise.
Conclusion
By combining the high frequency resolution of long-frame analysis with the robust temporal correlation of short-frame selection, this hybrid method offers a significant advancement in pitch estimation for noisy environments. Future work may focus on reducing the computational overhead of the tracking stage and further refining performance in extreme white noise conditions.
Summary updated