TL;DR
This paper introduces a novel, simple-to-compute feature called Low Frequency Spectral Density (LFSD) for distinguishing between breathy, modal, and pressed voice phonation types. LFSD shows promising results, outperforming some existing features, particularly for female voices, and offers a valuable tool for applications like speech style diarization and emotion recognition.
What Problem Does This Paper Address?
Accurately identifying different voice phonation types (breathy, modal, pressed) is crucial for various speech processing applications, including emotion recognition, speaker diarization, and even assessing mental states like depression. Existing methods often rely on complex features like Normalized Amplitude Quotient (NAQ) or spectral slope measures (e.g., H1-H2 difference). However, these features can be sensitive to noise, difficult to estimate accurately for all voices (especially high-pitched female voices), and may not always provide sufficient discrimination. This paper addresses the need for a simpler, robust feature that can effectively differentiate these phonation types.
Methods
The core contribution of this paper is the introduction and evaluation of the Low Frequency Spectral Density (LFSD) feature. LFSD is derived from the short-time spectrum of a speech segment and quantifies the energy in the low-frequency region below the first vocal tract resonance, specifically around the glottal formant. The hypothesis is that breathy phonation, characterized by a higher open quotient and greater subglottal cavity influence, will exhibit higher spectral density in this low-frequency band compared to modal and pressed phonation.
The LFSD is computed using the formula:
$$ LFSD = \frac{\sum_{\omega_c}^{\omega} |X(\omega)|}{\sum |X(\omega)|} $$
where $X(\omega)$ is the discrete-time Fourier transform of the speech segment, and $\omega_c$ is a chosen cut-off frequency. The authors empirically determined a window size of 10 ms and a cut-off frequency of 200 Hz to be effective.
The paper also compares LFSD against several established features:
- Normalized Amplitude Quotient (NAQ): Related to the glottal pulse's closed quotient.
- H1-H2: Difference in amplitude between the first two harmonics, a spectral slope measure.
- Band Harmonic-to-Noise Ratio (HNR3): Measures the ratio of harmonic to noise energy in a specific frequency band.
- Wavelet Peak Slope (PS) and Normalized Wavelet Peak Slope (NPS): Features derived from wavelet transform outputs, proposed for better performance than NAQ.
These features were analyzed using box plots to visualize their distributions across phonation types and Spearman correlation coefficients to assess their separability. Classification experiments were conducted using a linear discriminant analysis (LDA) classifier with a leave-one-speaker-out strategy.
Datasets
The study utilized a dataset of 8 Finnish vowels uttered in three phonation types (breathy, modal, pressed) by 6 female and 5 male speakers. Each vowel was recorded three times, resulting in a total of 792 isolated vowels. The original recording was at 44.1 kHz and downsampled to 16 kHz for analysis. This dataset appears to be a previously established benchmark for voice quality research, as referenced by [9, 10].
Key Insights and Ideas
- LFSD's Physical Intuition: The paper effectively links the acoustic phenomenon of increased low-frequency energy in breathy voices to the physiological characteristic of a higher open quotient and greater subglottal influence. This provides a strong theoretical basis for the proposed feature.
- Female Voice Advantage: LFSD shows a notable advantage over features like H1-H2 for female voices. This is attributed to the potential overlap of the second harmonic with the first vocal tract resonance in female speech, which can confound H1-H2 measurements.
- Complementary Information from HNR: While LFSD and NAQ showed similar information, the HNR feature, particularly in discriminating breathy from modal voices, provided complementary information. This suggests that combining features might yield further improvements.
- Speaker Variability: The study highlights that no single feature performs consistently best across all speakers, underscoring the challenge of robust voice quality analysis. However, LFSD demonstrated competitive performance and even outperformed NAQ for some individual speakers.
Key Results
- Distributional Separation: Box plots clearly show that LFSD provides a good monotonic separation between breathy, modal, and pressed phonation types, with breathy voices having the highest LFSD values and pressed voices the lowest.
- Classification Performance: In a three-class LDA classification task, LFSD achieved a correct classification rate of approximately 55% on the test set. While NAQ performed slightly better (around 65%), LFSD demonstrated a significant improvement over H1-H2 (around 45%) and NPS (around 35%).
- Individual Speaker Performance: For some individual speakers, LFSD achieved better scores than NAQ, suggesting its utility in scenarios where NAQ might fail.
- Two-Class Classification: When splitting the problem into two-class tasks (e.g., breathy vs. modal), classification performance generally improved. HNR was found to be particularly effective in distinguishing breathy from modal voices.
Limitations and Future Work
- Feature Range Variability: The authors acknowledge that the absolute range of LFSD values might vary across different recording setups, similar to other spectral slope features. However, they argue that the discrimination among phonation types would likely be preserved within a homogeneous recording environment.
- Noise Robustness: The paper mentions that performance of features typically deteriorates significantly at SNRs below 10 dB. The robustness of LFSD to different types and levels of noise was not extensively studied and is suggested as a direction for future work.
- Combination of Features: While combining LFSD with HNR showed some promise for specific discriminations, further exploration of feature combinations for improved overall classification is warranted.
- Real-world Applications: The authors suggest that LFSD could be useful for quick annotation of audio books, which generally have good signal quality, and further investigation into its performance in noisy conditions is needed.
Summary updated