TL;DR
This paper introduces a novel F0 estimation method that leverages autocorrelation functions on a log-time scale (LTACF). By directly estimating the change in log-F0 from shifts in LTACF peaks, the method achieves improved robustness against noise, particularly in Mandarin tone recognition tasks.
What Problem Does This Paper Address?
Accurate estimation of the fundamental frequency (F0) of speech is crucial for many speech processing applications, especially in tonal languages like Mandarin where F0 variations distinguish word meanings. However, existing F0 estimation methods often struggle in noisy environments, leading to degraded performance in real-world scenarios like mobile communication. This paper aims to develop a more noise-robust F0 estimation technique that can improve the accuracy of applications like Mandarin tone recognition.
Methods
The core innovation lies in operating on the autocorrelation function (ACF) in a log-time scale (LTACF). The authors observe that for periodic signals, the peaks of an ACF exhibit a consistent pattern on a log-time scale, with their positions shifting based on the period. This allows for the direct estimation of the change in log-F0 ($\Delta \log F0$) by analyzing the shift of these LTACF peaks between adjacent frames, without explicitly estimating F0 first.
The proposed method involves:
- Log-Time Scale Autocorrelation (LTACF): Calculating ACFs and then transforming them to a log-time scale.
- $\Delta \log F0$ Estimation: Estimating the shift in LTACF peaks between consecutive frames using cross-correlation functions. This shift directly corresponds to $\Delta \log F0$.
- F0 Estimation: Reconstructing the log-F0 by summing the LTACFs, shifted according to the estimated $\Delta \log F0$ values.
- Voiced/Unvoiced (V/UV) Classification: A thresholding mechanism is used to classify frames as voiced or unvoiced.
The method also incorporates a weighted summation of neighboring cross-correlation functions to enhance the robustness of $\Delta \log F0$ estimation against noise.
Datasets
The experiments were conducted using the keele pitch database. To simulate noisy conditions, in-car noise recorded under highway conditions was added to clean speech at various Signal-to-Noise Ratios (SNRs) including 0, 5, 10, 15, and 25dB.
Key Insights and Ideas
- Log-Time Scale Advantage: Operating on the log-time scale of ACFs reveals a consistent pattern of peaks whose shifts are directly related to changes in F0, offering a more direct path to estimating $\Delta \log F0$.
- Decoupling $\Delta \log F0$ Estimation: The ability to estimate $\Delta \log F0$ before F0 estimation is a key novelty, allowing for more robust estimation in noisy conditions where direct F0 estimation might fail.
- Summation of Shifted LTACFs: Enhancing the LTACF signal by summing shifted versions based on the estimated $\Delta \log F0$ helps to reinforce the periodic structure and improve accuracy.
- Robustness to SNR Changes: The proposed V/UV classification method demonstrates significantly less performance degradation when thresholds are fixed across different SNR conditions compared to the baseline method, indicating better inherent robustness.
Key Results
- F0 Estimation Accuracy: The proposed method significantly improved Gross Error Rates (GER) in F0 estimation compared to the baseline method, especially under lower SNR conditions (e.g., 5dB and 10dB). For instance, at 5dB SNR, the proposed method achieved a GER of 15.0% compared to the baseline's 26.2%.
- V/UV Classification: The proposed method showed better performance in V/UV classification, achieving lower False Alarm Rates (FAR) and False Rejection Rates (FRR) across various SNR levels, particularly when using a fixed threshold. Figure 4 visually demonstrates this improved accuracy.
- Mandarin Tone Recognition: The application to Mandarin tone recognition yielded substantial improvements. Tone error rates were consistently lower for the proposed method across all tested SNR conditions, with notable gains at lower SNRs. For example, at 5dB SNR, the proposed method achieved a tone error rate of 20.7% compared to the baseline's 33.3% (using the 0:5dB threshold).
Limitations and Future Work
- Limitations: While the paper demonstrates significant improvements, the authors note that the performance gap between the proposed and baseline methods was "much larger" for V/UV classification when thresholds were not tuned to specific SNR conditions. This implies that while the proposed method is more robust, some threshold tuning might still be beneficial for optimal performance.
- Future Work: The authors suggest further improvements to tone recognition accuracy and evaluation of the proposed method under a wider range of noise conditions.
Summary updated