Efficient Compression of Multitask Multilingual Speech Models

TL;DR

This paper introduces DistilWhisper, a novel compression method for multilingual speech models that improves recognition accuracy for under-represented languages while maintaining model robustness and efficiency.

Contribution

It proposes a dual strategy of lightweight fine-tuning with language-specific experts and knowledge distillation to enhance multilingual speech model performance.

Findings

DistilWhisper outperforms standard fine-tuning and LoRA adapters.

It improves ASR accuracy for low-resource languages.

The approach introduces negligible parameter overhead.

Abstract

Whisper is a multitask and multilingual speech model covering 99 languages. It yields commendable automatic speech recognition (ASR) results in a subset of its covered languages, but the model still underperforms on a non-negligible number of under-represented languages, a problem exacerbated in smaller model versions. In this work, we examine its limitations, demonstrating the presence of speaker-related (gender, age) and model-related (resourcefulness and model size) bias. Despite that, we show that only model-related bias are amplified by quantization, impacting more low-resource languages and smaller models. Searching for a better compression approach, we propose DistilWhisper, an approach that is able to bridge the performance gap in ASR for these languages while retaining the advantages of multitask and multilingual capabilities. Our approach involves two key strategies: lightweight modular ASR fine-tuning of whisper-small using language-specific experts, and knowledge distillation from whisper-large-v2. This dual approach allows us to effectively boost ASR performance while keeping the robustness inherited from the multitask and multilingual pre-training. Results demonstrate that our approach is more effective than standard fine-tuning or LoRA adapters, boosting performance in the targeted languages for both in- and out-of-domain test sets, while introducing only a negligible parameter overhead at inference.