Use of harmonic phase information for polarity detection in speech signals

Ibon Saratxaga, Daniel Erro, Inmaculada Hernáez, Iñaki Sainz, Eva NavasInterspeech 2009

TL;DR

This paper introduces two novel methods, Phase Cut (PC) and Relative Phase Shift (RPS), that leverage harmonic phase information to accurately detect the polarity of speech signals. These methods offer a significant improvement over existing techniques, particularly in noisy conditions, and are crucial for robust speech synthesis and analysis.

What Problem Does This Paper Address?

Speech signals, especially voiced segments, have an inherent asymmetry in their waveform due to the glottal excitation pulses. This asymmetry, or polarity, is important for various speech processing tasks. For instance, in speech synthesis, concatenating units with inconsistent polarity can lead to audible artifacts. Similarly, pitch synchronous analysis techniques rely on accurate identification of pitch periods, which can be disrupted by polarity inversions. While visual inspection or simple thresholding can sometimes determine polarity, these methods are unreliable under noisy or reverberant conditions. Existing automatic methods also struggle with robustness across different recording conditions and voices. This paper aims to provide more reliable and robust automatic polarity detection methods.

Methods

The paper proposes two new algorithms based on the phase information of harmonic analysis:

  • Phase Cut (PC) Method: This method assumes that the vocal tract's phase contribution is negligible at lower frequencies. It focuses on the intersection point of the phase lines of the first two harmonics. The polarity is determined by whether this intersection occurs near phase 0 (positive peaks) or phase π (negative peaks). The formula derived is $P_{cut} = 2\phi_1 - \phi_2$.
  • Relative Phase Shift (RPS) Method: This method relies on the assumption that the vocal tract's frequency response is smooth. It analyzes the "ripple" or pattern of relative phase shifts between harmonics. When a signal is inverted, the relative phase shifts of even harmonics are shifted by π, leading to a distinct interleaved pattern in the "RPS phasegram" (a visualization of RPS over time and frequency). The algorithm calculates the smoothness of these RPS patterns to distinguish between original and inverted signals.

Both methods analyze voiced frames and utilize pitch estimation from the CDP algorithm.

Datasets

The paper evaluates its methods on nine databases with known polarity:

  • Karolina & Pello (Basque emotional speech)
  • Laura DB (TC-Star, English neutral speech)
  • CMU_ARCTIC_SLT & CMU_ARCTIC_BDL (US English speech)
  • Berlin DB of Emotional Speech (German emotional speech)
  • Subset of SPEECON Spanish DB (Spanish speech with varying SNR and recording conditions)

The paper also compares its algorithms against two established methods: GSGW (gradient of spurious glottal waveforms) and RAPT (robust algorithm for pitch tracking).

Key Insights and Ideas

  • Phase as a Robust Polarity Indicator: The core insight is that harmonic phase information, particularly the phase relationships between harmonics, is a highly reliable indicator of speech signal polarity, even in challenging acoustic conditions.
  • Sensitivity of RPS to Inversion: The RPS method reveals a striking sensitivity to polarity inversion, manifesting as a clear interleaved pattern in the RPS phasegram. This visual and algorithmic distinction is a key finding.
  • Robustness Across Conditions: The proposed phase-based methods demonstrate superior robustness compared to existing techniques, performing consistently well across different voices, languages, and recording environments (including noisy ones).
  • Harmonic Model Suitability: The methods are particularly well-suited for speech processing systems that already employ a harmonic model of speech.

Key Results

The experiments show that both the RPS and PC methods achieve very high accuracy, significantly outperforming GSGW and RAPT.

  • Overall Accuracy: For clean databases, RPS achieved an average accuracy of 99.89%, and PC achieved 99.19%. In contrast, RAPT averaged 81.93%, and GSGW averaged 86.17%.
  • Performance in Noise: On the noisy C2 subset of the SPEECON database, the phase-based methods (especially PC) performed better than methods using wider bandwidths, indicating their resilience to waveform distortion caused by noise.
  • Consistency: The authors report that they did not find any voice for which the detection performance of RPS and PC decayed noticeably, unlike the other methods.

Limitations and Future Work

  • Limitations: The paper notes that for very noisy conditions (like the C2 database), waveform distortion can affect phase information, impacting the RPS method more significantly than the PC method. The PC method's assumption of negligible vocal tract phase contribution at lower frequencies might be a limitation in certain scenarios.
  • Future Work: The authors suggest that these phase-based methods are highly suitable for speech processing systems that utilize harmonic models, implying their integration into such systems for improved performance. The paper doesn't explicitly detail future research directions beyond this application.

Summary updated

BibTeX
@inproceedings{saratxaga2009use,
  author = {Ibon Saratxaga and Daniel Erro and Inmaculada Hernáez and Iñaki Sainz and Eva Navas},
  title = {{Use of harmonic phase information for polarity detection in speech signals}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure