Linear Script Representations in Speech Foundation Models Enable Zero-Shot Transliteration

TL;DR

This paper demonstrates that script variations in multilingual speech models are linearly encoded in activations, allowing zero-shot control over output script through activation modification, improving speech recognition flexibility.

Contribution

It reveals that script information is linearly embedded in model activations and introduces a method to control output script at inference time by modifying these activations.

Findings

Activation modification enables script control in speech recognition.

Effective script change even for unconventional language-script pairs.

Performance remains competitive across all Whisper model sizes.

Abstract

Multilingual speech foundation models such as Whisper are trained on web-scale data, where data for each language consists of a myriad of regional varieties. However, different regional varieties often employ different scripts to write the same language, rendering speech recognition output also subject to non-determinism in the output script. To mitigate this problem, we show that script is linearly encoded in the activation space of multilingual speech models, and that modifying activations at inference time enables direct control over output script. We find the addition of such script vectors to activations at test time can induce script change even in unconventional language-script pairings (e.g. Italian in Cyrillic and Japanese in Latin script). We apply this approach to inducing post-hoc control over the script of speech recognition output, where we observe competitive performance across all model sizes of Whisper.