TL;DR
This paper introduces two novel "glottal signatures" derived from the residual speech signal, outperforming existing glottal-based methods for speaker recognition. These signatures are robust to text and language variations and require only a few seconds of speech, offering a promising new avenue for more efficient and accurate speaker identification systems.
What Problem Does This Paper Address?
Speaker recognition systems typically rely on features extracted from the speech's spectral envelope, which primarily capture information about the vocal tract. However, the way our vocal folds vibrate (the glottis) also carries unique information about our identity. Previous attempts to leverage this glottal information have had limited success or required complex signal processing. This paper aims to explore the potential of simpler, yet informative, "glottal signatures" derived from the residual signal (what's left after removing the vocal tract's influence) for speaker recognition, seeking to improve accuracy and efficiency.
Methods
The core of this work lies in extracting two "glottal signatures" from the residual signal. First, the residual signal is analyzed using a Deterministic plus Stochastic Model (DSM). This model decomposes the voiced residual signal into a low-frequency deterministic component and a high-frequency stochastic component.
The two signatures are:
- First Eigenresidual: This is the primary waveform representing the deterministic part of the residual signal, derived from a Principal Component Analysis (PCA) of pitch-synchronous residual frames.
- Noise Energy Envelope: This signature captures the temporal modulation of the high-frequency stochastic component, essentially an average of the Hilbert envelope of the high-frequency residual content.
These signatures are then used in a speaker identification system. For a given speaker, training signatures are computed. When a new speech sample arrives, its signatures are extracted and compared to the training signatures using a Relative Time Squared Error (RTSE) measure. A combined confusion matrix, formed by multiplying the RTSE matrices for both signatures, is used to determine the closest match and identify the speaker.
Datasets
- TIMIT: A widely used speech corpus containing recordings from 630 speakers (438 male, 192 female) sampled at 16 kHz.
- YOHO: A corpus of speech from 138 speakers (108 male, 30 female) sampled at 8 kHz, collected over multiple sessions with a time gap between them.
Key Insights and Ideas
- Residual Signal is Rich: The residual signal, often overlooked in favor of spectral features, contains significant speaker-specific information, particularly related to glottal activity.
- Simplicity and Effectiveness: The proposed glottal signatures are relatively simple to extract and demonstrate strong performance, suggesting that complex glottal flow modeling might not always be necessary for speaker recognition.
- Text and Language Independence: The signatures are shown to be comparable across different phonetic classes and are expected to be language-independent, making them highly versatile.
- Data Efficiency: Reliable estimation of these signatures requires only a few seconds of voiced speech (around 7 seconds for males, 4 seconds for females), making them practical for real-world applications.
- Complementarity: Combining the two proposed glottal signatures significantly improves identification rates compared to using either one alone, highlighting the benefit of using multiple complementary features.
Key Results
- TIMIT Performance: Using both glottal signatures, the misidentification rate on the TIMIT database was 3.65% for 630 speakers, significantly outperforming previous glottal-based methods (e.g., 5.06% for 168 speakers in [4] and 28.64% in [3]).
- YOHO Performance: On the YOHO database, the proposed eigenresidual signature achieved a misclassification rate of 29.3%, which is better than the 36.3% reported in [4] for similar conditions.
- Speed of Convergence: The signatures converge quickly, with about 1000 residual frames (roughly 7 seconds of voiced speech) being sufficient for reliable estimation.
- Robustness: The signatures showed good performance even when training and testing sessions were separated by several months on the YOHO database, although performance did degrade with longer time gaps.
Limitations and Future Work
- Temporal Mismatch: Performance degrades when training and testing sessions are spaced over long periods, particularly on the YOHO database. This is attributed to variations in recording conditions and increased intra-speaker variability over time.
- Complementarity with Spectral Features: The authors acknowledge that combining their glottal signatures with conventional vocal tract-based features (like MFCCs) is expected to yield further improvements, as suggested by previous research.
- Channel Compensation: The paper suggests that applying channel compensation techniques could mitigate the impact of differing recording conditions on signature estimation, especially concerning phase response differences.
- Ongoing Work: The authors are actively pursuing research into combining these glottal signatures with existing systems and exploring channel compensation methods.
Summary updated