TL;DR
This paper introduces a novel method for distinguishing voiced from unvoiced speech segments using the analytic signal, achieving high accuracy and robustness to noise. The proposed method outperforms traditional energy and Zero Crossing Rate (ZCR) measures, particularly in noisy conditions, and demonstrates the potential for reducing frame size without significant performance degradation.
What Problem Does This Paper Address?
Accurately segmenting speech into voiced and unvoiced segments is a fundamental task in speech processing, crucial for applications like automatic speech recognition (ASR), speaker identification, and speech synthesis. Existing methods, such as those based on energy, zero-crossing rate (ZCR), or autocorrelation, often rely on the assumption of signal stationarity over short frames (milliseconds), requiring a trade-off between temporal resolution and accuracy. This paper addresses the challenge of developing a robust and accurate voiced/unvoiced (V/UV) segmentation method that can potentially operate with shorter frames and maintain performance in the presence of noise.
Methods
The core of the proposed method lies in the computation of the analytic signal (AS). The process involves:
- Band-pass filtering: The audio signal is filtered to isolate frequencies relevant to speech, specifically between 50 Hz and 300 Hz. This step aims to enhance the prominence of the first harmonic in voiced segments.
- Analytic Signal Computation: A Hilbert transform is applied to the filtered signal to obtain the analytic signal. For voiced segments, this signal approximates a complex exponential representing the fundamental frequency and its phase. For unvoiced segments, it primarily contains noise.
- Magnitude Estimation: The magnitude of the analytic signal is computed. In voiced regions, this magnitude serves as an approximation of the amplitude of the first harmonic, while in unvoiced regions, it's significantly lower due to noise.
- Thresholding: The estimated magnitude is automatically thresholded to perform the V/UV segmentation. The paper compares two thresholding techniques: an "ad-hoc" method and Otsu's method, finding Otsu's method to be more robust.
The paper also investigates the impact of frame size on performance, exploring shorter frames than typically used in traditional methods.
Datasets
The primary dataset used for evaluation is the TIMIT Acoustic-Phonetic Continuous Speech Corpus. This dataset is manually segmented into phonemes, providing ground truth for V/UV segmentation. The authors note that TIMIT is relatively balanced in terms of voiced (56%) and unvoiced (44%) samples.
Key Insights and Ideas
- Analytic Signal Superiority: The analytic signal provides a more discriminative feature for V/UV segmentation compared to energy and ZCR, especially in challenging conditions.
- Robustness to Noise: The AS-based method exhibits significantly better robustness to additive noise than energy and ZCR. This is a critical advantage for real-world applications where clean speech is rare.
- Frame Size Flexibility: The method demonstrates that it's possible to reduce frame sizes (e.g., to 14 ms) without a substantial loss in V/UV detection accuracy, offering potential for improved temporal resolution.
- Otsu's Method Effectiveness: Otsu's method proves to be a reliable and robust thresholding technique for V/UV segmentation, outperforming a simpler "ad-hoc" method and the original Otsu method in certain aspects when adapted.
Key Results
- High Detection Accuracy: The AS-based method achieved a 92.599% true detection rate (TBD) with a frame size of 14 ms, significantly outperforming energy (81.767%) and ZCR (89.066%) under similar conditions.
- Noise Robustness: Even at a low SNR of 5 dB, the AS method maintained a TBD of 89.718%, while energy dropped to 81.065% and ZCR to 74.023%.
- Temporal Precision: The method also demonstrated good temporal precision in locating segmentation boundaries, with the best results achieved for a frame size of 18 ms.
- Parameter Stability: The learned thresholding parameter (C3) for the AS method showed less dependency on SNR compared to the energy method, indicating greater inherent robustness.
Limitations and Future Work
- Limitations:
- The current method cannot distinguish boundaries between successive voiced or unvoiced phonemes within longer segments.
- The method's performance is evaluated on TIMIT, which has specific dialectal characteristics.
- Future Work:
- Leveraging the pitch estimation capability of the AS to incorporate periodic information of voiced signals.
- Testing the method with different types of noise, particularly non-white noise and in realistic acoustic environments (e.g., car interiors).
- Comparing the method with other existing V/UV measures.
- Applying the V/UV information to improve existing ASR and speech alignment systems.
Summary updated