Introduction
Automatic Speech Recognition (ASR) systems rely heavily on the representation of the speech signal. Formants—the spectral peaks corresponding to vocal tract resonances—are highly discriminant for phonetic classification. However, their automatic extraction is notoriously difficult due to peak merging or tracking errors. Consequently, Mel-Frequency Cepstral Coefficients (MFCCs) have become the standard, despite being less physically intuitive. This paper investigates whether automated "formant-like" features can achieve the performance levels of hand-labeled formants (HLF) in vowel classification tasks.
Methodology and Feature Extraction
The study compares five distinct feature sets using Hidden Markov Models (HMMs):
- Hand-Labeled Formants (HLF): The expert-curated baseline provided by phoneticians.
- Robust Formants (RF): Derived from the Split Levinson Algorithm (SLA), ensuring a fixed number of spectral peaks and continuous tracks.
- HMM2 Feature Extractor: A novel 2D-HMM approach that segments the speech spectrum into high-energy regions rather than pinpointing exact peaks.
- MFCC13: A standard 13-dimensional state-of-the-art baseline.
- MFCC3: A 3-dimensional MFCC baseline for direct comparison with 3D formant features.
The experiments utilized the American English Vowels (AEV) database, containing recordings from men, women, and children, including rare hand-edited formant tracks for validation.
Key Insights: The HMM2 Approach
The HMM2 approach represents a shift from pinpointing exact resonance peaks to identifying broader spectral energy distributions. As shown in Figure 2, the model uses a frequency-based HMM to segment the spectrum. While the resulting tracks appear "blocky" compared to smooth hand-labeled curves, they capture the essential discriminant information needed for classification. This suggests that for ASR, the general region of spectral energy is often more critical than the precise mathematical peak.
Experimental Results
The evaluation focused on vowel classification accuracy across gender-independent and gender-dependent models.
Key findings include:
- HLF Superiority: Hand-labeled formants achieved over 80% accuracy, confirming their high discriminant power even in standard HMM recognizers.
- HMM2 Performance: In gender-dependent scenarios, HMM2 features performed comparably to hand-labeled formants, demonstrating the viability of automated statistical extraction.
- Dimensionality Gap: While HLF outperformed the 3D MFCC3 baseline, it could not match the 13D MFCC13. This indicates that higher-dimensional spectral envelopes retain useful information discarded by compact formant representations.
Conclusion
While hand-labeled formants remain the gold standard for low-dimensional features, automated methods like HMM2 provide a robust alternative for vowel classification. The study highlights that identifying high-energy regions is a powerful proxy for traditional formant tracking. Future work should explore these features in noisy environments, where formant-based representations are hypothesized to be more resilient than standard MFCCs.
Summary updated