Mel-Frequency Cepstral Coefficients of Voice Source Waveforms for Classification of Phonation Types in Speech

Sudarsana Reddy Kadiri, Paavo AlkuInterspeech 2019

TL;DR

This paper introduces a novel approach to classifying different voice phonation types (breathy, modal, tense) by extracting Mel-Frequency Cepstral Coefficients (MFCCs) directly from the voice source waveform rather than the speech signal itself. This method, particularly using the Zero Frequency Filtering (ZFF) derived source waveform, achieves state-of-the-art classification accuracy and offers complementary information to existing features.

What Problem Does This Paper Address?

The way our vocal folds vibrate – our "phonation type" – significantly impacts the quality and expressiveness of our voice. Distinguishing between phonation types like breathy, modal, and tense is crucial for various applications, including speech synthesis, emotion recognition, and even detecting speech pathology. While existing methods use features derived from the speech signal or directly from the glottal flow, these often struggle with consistency across speakers or fail to capture all the nuances. This paper aims to improve phonation type classification by leveraging the rich information contained within the voice source waveform, which is the fundamental excitation signal produced by the vocal folds before it's shaped by the vocal tract.

Methods

The core innovation lies in deriving MFCCs from two different estimates of the voice source waveform:

  • Quasi-Closed Phase (QCP) Glottal Inverse Filtering: This method estimates the glottal flow waveform by modeling the vocal tract and then "removing" its effect from the speech signal. It assumes a source-filter model of speech production.
  • Zero Frequency Filtering (ZFF): This method provides an approximate voice source waveform without explicitly modeling the vocal tract. It leverages the property that glottal closure instants create impulse-like excitations that manifest across the entire spectrum, including at 0 Hz. The ZFF signal's zero-crossings correspond to these glottal closure events.

Once these source waveforms are obtained, standard MFCC extraction procedures are applied: the spectrum is computed, warped to the mel scale, logarithmically scaled, and then transformed using the Discrete Cosine Transform (DCT). The first 13 cepstral coefficients, along with their delta and double-delta derivatives, form the feature vectors.

Datasets

The study utilizes a Finnish speech database consisting of 8 different vowels uttered by 11 speakers (6 female, 5 male) across three phonation types: breathy, modal, and tense. Each vowel was spoken three times per speaker, resulting in a total of 792 vowel instances. The speech was recorded at 44.1 kHz and downsampled to 16 kHz for analysis.

Key Insights and Ideas

  • Voice Source is Key: The paper strongly suggests that the voice source waveform itself contains highly discriminative information about phonation type, often more so than the processed speech signal.
  • ZFF's Advantage: The ZFF method, which doesn't require explicit source-filter modeling, proves particularly effective. MFCCs derived from ZFF-based source waveforms achieve the highest individual classification accuracy.
  • Complementary Information: The proposed MFCC features (both QCP and ZFF based) offer information that complements existing voice quality features (VQ), conventional MFCCs, and Zero-Time Windowing Cepstral Coefficients (ZTWCCs). Combining these features leads to significant improvements in classification accuracy.
  • All Features are Better: The best performance is achieved when all considered feature sets (VQ, conventional MFCCs, ZTWCCs, MFCC-QCP, and MFCC-ZFF) are combined, indicating that each contributes unique and valuable information.

Key Results

  • Highest Individual Accuracy: MFCCs derived from the ZFF-based voice source waveform (MFCC-ZFF) achieved the highest individual classification accuracy of 70.54%.
  • Combined Performance: Combining all feature sets (VQ, MFCC, ZTWCC, MFCC-QCP, MFCC-ZFF) resulted in the highest overall classification accuracy of 76.58%.
  • Improvement over References: The proposed MFCC-ZFF features outperformed existing reference features (VQ, MFCC, ZTWCC) and MFCC-QCP individually.
  • Confusion Matrix Insights: The confusion matrices highlight that modal phonation is often confused with breathy and tense voices. While improvements were seen with the proposed features, especially for breathy voice detection, confusion between modal and tense voices persists, suggesting a need for features that specifically capture tense phonation characteristics.

Limitations and Future Work

  • Tense Phonation Confusion: The paper acknowledges that confusion between modal and tense phonation types still exists, even with the combined features. This suggests that further research is needed to develop features that more effectively capture the production variations associated with tense phonation.
  • Database Specificity: The study used a Finnish speech database. While the findings are promising, further validation on larger and more diverse multilingual databases would be beneficial.
  • Feature Combination Strategy: While combining all features yielded the best results, exploring more sophisticated feature selection or fusion techniques could potentially lead to even better performance or more compact feature sets.

Summary updated

BibTeX
@inproceedings{kadiri2019mel,
  author = {Sudarsana Reddy Kadiri and Paavo Alku},
  title = {{Mel-Frequency Cepstral Coefficients of Voice Source Waveforms for Classification of Phonation Types in Speech}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure