TL;DR
This paper introduces a faster method for speaker identification by using Maximum Likelihood Linear Regression (MLLR) matrices and sufficient statistics to represent anchor speakers, significantly reducing computation time compared to traditional GMM-based anchor models. This approach achieves comparable performance, especially when cascaded with a GMM-UBM system, offering a substantial speedup for large-scale speaker indexing and identification.
What Problem Does This Paper Address?
Speaker identification and indexing in large audio databases is computationally intensive. Traditional "anchor modeling" techniques aim to reduce this complexity by representing speakers using a set of pre-defined "anchor models." However, calculating the likelihood of a test utterance against these anchor models, even with optimizations, remains a bottleneck. This paper addresses the need for a more computationally efficient method to generate speaker characterization vectors (SCVs) within the anchor model framework, enabling faster and more scalable speaker identification systems.
Methods
The core innovation lies in replacing traditional Gaussian Mixture Models (GMMs) for anchor speakers with Maximum Likelihood Linear Regression (MLLR) matrices. MLLR is a standard technique for adapting acoustic models in Automatic Speech Recognition (ASR). In this work, MLLR matrices are learned for each anchor speaker using their training data, with the GMM-Universal Background Model (GMM-UBM) serving as the speaker-independent (SI) model.
The key to the speedup is the efficient calculation of the likelihood of a test utterance with respect to an anchor speaker. Instead of directly computing GMM likelihoods, the method leverages two pre-computed statistics (K and G) derived from aligning the test utterance with the GMM-UBM. These statistics, along with the anchor speaker's MLLR matrix, are used in a matrix multiplication to rapidly estimate the likelihood. This avoids repeated GMM evaluations and significantly reduces computation. The paper also explores a cascade system where the proposed Fast MLLR anchor model system acts as a first stage to prune the search space, followed by a conventional GMM-UBM system for final identification.
Datasets
- NIST 2004 SRE (core condition): Used for evaluating the proposed method in a closed-set speaker identification task. The dataset details include 310 unique speakers in the training set and 306 in the test set, with 1163 test utterances.
- NIST 1999, 2001 SRE, and Switchboard-1 Release-2: Used for training the GMM-UBM and for providing a larger pool of speakers (1346) from which anchor models were selected.
Key Insights and Ideas
- MLLR as a Compact Anchor Representation: Using MLLR matrices as representations for anchor speakers is a novel and effective way to capture speaker-specific characteristics without the computational overhead of full GMMs.
- Sufficient Statistics for Speed: The clever use of pre-computed sufficient statistics (K and G) derived from the GMM-UBM allows for a dramatic reduction in likelihood computation for each anchor model. This is the primary driver of the speedup.
- Cascade System Synergy: Combining the Fast MLLR anchor model system with a conventional GMM-UBM system offers a compelling trade-off: it achieves comparable identification accuracy to a purely conventional cascade system while being significantly faster.
- Scalability Potential: The computational efficiency of the proposed method makes it highly suitable for large-scale speaker identification and indexing tasks where traditional methods would be prohibitively slow.
Key Results
- Speedup: The Fast MLLR anchor model system is 9.37x faster than the conventional anchor model system for a set of 346 anchor models.
- Cascade Performance: When used in a cascade system, the proposed method is 4.21x faster than the conventional cascade system while maintaining comparable performance. For instance, achieving a similar accuracy to the conventional system at Nf=40, but with Nf=65 and a 4.21x speedup.
- Accuracy: While the Fast MLLR anchor model system shows a slight degradation in performance compared to the conventional anchor model system when used alone, this difference becomes negligible when integrated into a cascade system. The accuracy of both systems converges as the "N-best" (number of top candidates considered) increases.
Limitations and Future Work
- Slight Accuracy Degradation (Standalone): The paper notes that the Fast MLLR anchor model system, when used independently, exhibits a slight performance degradation compared to the conventional anchor model system.
- Future Work: The authors plan to investigate the applicability of the proposed method for "online cohort models selection" in speaker verification tasks.
Summary updated