Kolluru Sai Venkata Ravi Sastry supervised by Prof. Anil Kumar Vuppala received his Master of Science by Research in Computer Science Engineering (CSE). Here’s a summary of his research work on Exploration of Hierarchical Representations and Perceptual Decomposition for TTS Quality Prediction

Text-to-speech systems are advancing rapidly, yet reliably evaluating perceptual quality remains a significant challenge. Human listening tests are expensive, time-consuming, and subject to substantial inter-rater variability, motivating automatic MOS prediction. Existing approaches face two persistent limitations: most underexploit the hierarchical layer structure of self-supervised representations, and models trained on one corpus frequently fail to generalise across acoustic domains.

This thesis presents two complementary frameworks addressing these limitations. The first, QMOS, extracts hierarchical quality features from a frozen Qwen2-Audio encoder via Grouped Acoustic Pool-ing (GAP), stratifying its 32 layers into early, mid, and deep groups to capture phonetic, prosodic, and semantic signals simultaneously. These representations are fused with layer-weighted WavLM-Large features through cross-attention, with a rater disagreement encoder and a composite objective of het-eroscedastic negative log-likelihood and linear concordance correlation providing supplementary super-vision. Without system-ID conditioning or listener embeddings, QMOS achieves system-level SRCC of 0.923 and 0.921 on SOMOS and BVCC respectively, with a cross-domain SRCC of 0.702.

The second framework, MDMOS, decomposes perceived quality into four perceptual dimensions acoustic fidelity, prosodic naturalness, phonetic intelligibility, and semantic coherence each modelled by a dedicated expert head combining layer-grouped WavLM-Large representations with hand-crafted signal-processing features. A quality-adaptive fusion module dynamically weights expert outputs, while an ordinal KL loss provides independent per-head supervision preventing head collapse. MDMOS achieves system-level SRCC of 0.925 and 0.944 on SOMOS and BVCC, matching state-of-the-art single-model baselines without ensembling or listener-dependent modelling.

Ablation studies confirm that explicit correlation alignment is the most critical training component and that perceptual decomposition captures complementary information not recoverable from a single scalar predictor. Together, these results demonstrate that exploiting representational hierarchies and decomposing quality into perceptual dimensions are effective and complementary inductive biases for robust, generalisable MOS prediction.

July 2026