Samsung Electronics Releases AnySimLite, a 700KB Ultra-Lightweight Voice Classification Encoder

Samsung Electronics' R&D India-Bangalore research team has released AnySimLite, a lightweight encoder for on-device voice proximity (SA) classification, based on a paper from Interspeech 2026. The model achieves a storage size of approximately 700KB and inference times of under 30ms on flagship smartphones, delivering results that approach or match SOTA in few-shot settings compared to 7B parameter LLMs.

Similarity-Based Classification Architecture

AnySimLite redefines classification problems as comparison problems, performing classification based on similarity scores between approximately 20 exemplars per class and the input. The core concept, Nuanced Text Similarity (NTS), is based on task-specific characteristic alignment rather than lexical or semantic similarity. The architecture independently encodes two texts to reduce inference complexity from O(n²) to O(n), combining a word channel (BiLSTM + Attention) and a character channel (Conv1D) to generate 16-dimensional L2-normalized embeddings. The character channel contributes to distinguishing spelling differences in out-of-vocabulary (OOV) names.

Performance and Deployment Efficiency

Across 9 datasets (Quora, IMDB, SNIPS, etc.), the model maintained SOTA or competitive performance with the smallest model size in most comparisons. It outperformed the strongest competitors on SMS Spam and TitleSimCurated, with an average accuracy drop of only 2.24% across all tasks. On the Samsung Galaxy S25 Ultra, disk usage is approximately 700KB with 8-bit quantization, and remains under 10MB even when supporting 12 tasks. While having approximately 100 times fewer parameters than BERT-scale models and over 100,000 times fewer than GPT-3 175B, the accuracy gap is maintained within a few points, lowering key barriers to on-device deployment.

Limitations and Future Directions

Performance gaps compared to large models occur in tasks relying on long document structures or requiring fine-grained multi-label classification. Additionally, due to the nature of the few-shot protocol, the model is sensitive to exemplar quality. The research team plans to explore the scalability of this 'problem reduction' recipe beyond classification.