A Modulation Property of Time-Frequency Derivatives of Filtered Phase and its Application to Aperiodicity and fo Estimation

Hideki Kawahara, Ken-Ichi Sakakibara, Masanori Morise, Hideki Banno, Tomoki TodaInterspeech 2017

TL;DR

This paper introduces a novel, linear Signal-to-Noise Ratio (SNR) estimator for speech that operates reliably across a wide dynamic range (0-80 dB) without requiring additional linearization. By leveraging the modulation properties of time-frequency derivatives of filtered phase, the method offers a robust way to quantify aperiodicity in speech, paving the way for improved speech synthesis and analysis tools.

What Problem Does This Paper Address?

Estimating aperiodicity in speech – the non-periodic, noisy components – is crucial for realistic speech synthesis and analysis. Existing methods often struggle because aperiodic components are significantly weaker than the dominant periodic (voiced) components. This makes it difficult to accurately measure the aperiodic signal without being overwhelmed by the periodic signal or being highly sensitive to errors in fundamental frequency (f0) estimation. The paper aims to provide a more robust and reliable method for estimating the level of these aperiodic components, effectively measuring the "noise" relative to the periodic signal.

Methods

The core of the proposed method lies in analyzing the modulation property of the time-frequency derivatives of the filtered phase. Instead of directly estimating the aperiodic component, the authors propose to estimate the ratio of the periodic component to the background noise. This is achieved by:

  • Instantaneous Frequency and Group Delay: They leverage the relationship between the time derivative of the phase (instantaneous frequency) and the frequency derivative of the phase (group delay).
  • Modulation Magnitude: The magnitude of the modulation in the instantaneous frequency and its time derivative, when mapped from filter center frequency, is shown to be proportional to the amplitude ratio of the dominant periodic component and the interfering noise.
  • Singularity Removal: To handle potential singularities in the instantaneous frequency calculation, they employ a smoothing kernel (a weighted average) to stabilize the estimates.
  • Windowing Function Design: A critical aspect is the choice of windowing function. The paper proposes and evaluates a six-term cosine series window known for its excellent sidelobe attenuation (-114.24 dB) and steep decay rate (54 dB/oct). This is crucial for isolating individual harmonic components and rejecting leakage from adjacent frequencies.
  • Scaling: The time and frequency derivatives are scaled to minimize the variance of a combined index, effectively balancing their contributions.

Datasets

Not discussed in the paper. The experiments were conducted using synthesized signals (pulse trains with added Gaussian white noise) and real speech data (a Japanese vowel sequence).

Key Insights and Ideas

  • Indirect SNR Estimation: The most significant insight is that instead of directly measuring the weak aperiodic component, one can reliably estimate the SNR by analyzing the modulation of the periodic signal's time-frequency characteristics. This circumvents the problem of the aperiodic signal being masked by the periodic one.
  • Windowing Function is Key: The performance of the SNR estimator is highly dependent on the windowing function. The proposed six-term cosine series window demonstrates superior performance due to its exceptional sidelobe suppression, which is critical for accurate spectral isolation.
  • Linearity Across Dynamic Range: The proposed method achieves a linear SNR estimation from 0 dB to 80 dB without needing complex linearization steps, making it highly practical.
  • Aperiodicity as a Modulation Phenomenon: The paper reframes aperiodicity estimation as a problem of quantifying the modulation effects caused by noise on the underlying periodic signal.

Key Results

  • The proposed SNR estimator is linear and accurate across a wide SNR range (0-80 dB).
  • The six-term cosine series window significantly outperforms other tested windows (Hanning, Blackman, Nuttall, Kaiser, DPSS) in terms of estimation accuracy and linearity, particularly at higher SNRs.
  • The method is shown to be effective on real speech data, producing a detailed time-frequency SNR map.

Limitations and Future Work

  • Temporal Distribution of Noise: The paper acknowledges that it does not discuss the temporal distribution of the random component within a single pitch period, which is an important factor for perceived noise level. Future work could explore this using group delay analysis.
  • Non-linear Interactions: The authors note that non-linear interactions between the vocal tract and vocal fold vibration, along with temporal fluctuations, might introduce apparent aperiodicity around resonant frequencies. Further analysis and tuning are needed for applications like speech synthesis.
  • Application to Real Speech: While the method works for constant f0 signals, adapting it to the continuously varying f0 of real speech requires an adaptive time-axis warping technique, which is introduced but noted as an area for further refinement.
  • Excitation Source Design: The estimated SNR map is useful for annotation but requires further processing (e.g., compensating for background noise bias, equalizing vocal tract resonances) to be directly used for designing excitation source signals for speech synthesis.

Summary updated

BibTeX
@inproceedings{kawahara2017modulation,
  author = {Hideki Kawahara and Ken-Ichi Sakakibara and Masanori Morise and Hideki Banno and Tomoki Toda},
  title = {{A Modulation Property of Time-Frequency Derivatives of Filtered Phase and its Application to Aperiodicity and fo Estimation}},
  booktitle = {Interspeech},
  year = {2017},
}
Figure