Distilled Non-Semantic Speech Embeddings with Binary Neural Networks for Low-Resource Devices

TL;DR

This paper presents BRILLsson, a compact binary neural network for non-semantic speech tasks, achieving high performance on multiple benchmarks while being suitable for low-resource devices.

Contribution

Introduces BRILLsson, a binary neural network model trained via knowledge distillation, optimized for low-resource devices with minimal size and latency.

Findings

BRILLsson models are only 2MB in size.

Achieve comparable performance to larger models on eight benchmarks.

Latency is less than 8ms, suitable for real-time applications.

Abstract

This work introduces BRILLsson, a novel binary neural network-based representation learning model for a broad range of non-semantic speech tasks. We train the model with knowledge distillation from a large and real-valued TRILLsson model with only a fraction of the dataset used to train TRILLsson. The resulting BRILLsson models are only 2MB in size with a latency less than 8ms, making them suitable for deployment in low-resource devices such as wearables. We evaluate BRILLsson on eight benchmark tasks (including but not limited to spoken language identification, emotion recognition, health condition diagnosis, and keyword spotting), and demonstrate that our proposed ultra-light and low-latency models perform as well as large-scale models.