TL;DR
This paper introduces a novel approach to speaker recognition by analyzing speech signals synchronously with vocal fold vibrations (pitch). By separating excitation source and vocal tract characteristics and processing them pitch-by-pitch, they achieve significantly better speaker identification than traditional methods that analyze speech in fixed-size blocks.
What Problem Does This Paper Address?
Speaker recognition systems aim to identify individuals based on their unique vocal characteristics. Existing methods often struggle to effectively isolate speaker-specific information from the complex speech signal, which contains both the source of sound (vocal folds) and the filter that shapes it (vocal tract). Many techniques rely on analyzing speech in fixed-length frames (e.g., 10-50 ms), which can introduce artifacts and blur speaker-specific cues, especially when dealing with limited training data. This paper addresses the challenge of extracting more robust and discriminative speaker features by proposing a more precise analysis aligned with the natural rhythm of speech production – the vocal fold vibration cycle.
Methods
The core innovation lies in pitch-synchronous analysis. Instead of using fixed-size frames, the researchers identify Glottal Closure Instants (GCIs), which mark the moments of significant vocal fold closure and are strong indicators of vocal fold vibration.
- GCI Detection: A method involving differencing the speech signal, passing it through a cascade of zero-frequency resonators, and then removing the trend is used to accurately locate GCIs.
- Feature Extraction:
- Excitation Source: The linear prediction (LP) residual, representing the raw excitation signal, is extracted in 4 ms segments centered around each GCI.
- Vocal Tract System: Weighted Linear Prediction Cepstral Coefficients (wLPCCs), capturing vocal tract characteristics, are extracted from segments spanning two pitch periods around each GCI.
- Speaker Modeling: Autoassociative Neural Networks (AANNs) are employed to learn the speaker-specific patterns within these extracted features. AANNs are trained to reconstruct their input, effectively learning a compressed representation of the data. Separate AANNs are trained for excitation source features and vocal tract features.
The key novelty is performing this feature extraction and AANN modeling pitch-synchronously, focusing analysis on the most informative parts of the speech signal.
Datasets
The study utilizes speech data from the TIMIT database.
- A Universal Background Model (UBM) was trained using utterances from 100 speakers (50 male, 50 female).
- For speaker recognition experiments, data from 10 speakers (5 male, 5 female) was used.
- Eight utterances (approximately 15 seconds) per speaker were used for training AANN models.
- Testing was performed using other utterances from these speakers.
Key Insights and Ideas
- Pitch Synchronicity Matters: Analyzing speech features in sync with vocal fold vibrations (GCIs) significantly improves the ability to capture speaker-specific information compared to traditional fixed-frame analysis.
- Complementary Information: Both the excitation source (LP residual) and vocal tract system (wLPCCs) carry distinct, complementary information about a speaker.
- AANNs for Feature Learning: AANNs are effective in learning the complex, non-linear relationships within speaker-specific features, even with limited training data.
- Optimal Compression Layer Size: The number of units in the AANN's compression layer is a critical parameter, and its optimal value appears to be speaker-specific, suggesting potential for personalized model tuning.
- Excitation Source is Highly Discriminative: The pitch-synchronous analysis of the LP residual showed particularly strong discrimination between genuine and imposter speakers.
Key Results
- Improved Discrimination: Figures 4(a) and 6(a) (pitch-synchronous) show much clearer separation between genuine and imposter curves than Figures 4(b) and 6(b) (pitch asynchronous) for both excitation and vocal tract features.
- Cumulative Error Reduction: Cumulative sum of normalized errors (Figs. 4(a) and 6(a)) consistently showed lower total error for the genuine speaker when using pitch-synchronous features.
- Pitch Synchronous Superiority: The paper concludes that features derived from pitch-synchronous analysis yield significantly better speaker recognition performance than those from pitch-asynchronous analysis. The difference was more pronounced for excitation source information.
Limitations and Future Work
- Limited Scale: The experiments were conducted on a subset of speakers from TIMIT. Large-scale experiments are needed to fully evaluate the significance of the proposed method.
- Speaker-Specific AANNs: The authors suggest exploring methods to develop speaker-specific AANN models, rather than relying on a universal model.
- Combining Evidence: Future work should investigate optimal methods for combining the evidence from both excitation source and vocal tract system features for even better performance.
- Optimal Parameter Tuning: Further research is needed to systematically determine the optimal number of units in the AANN compression layer, potentially through adaptive methods.
Summary updated