Deciphering Speech: a Zero-Resource Approach to Cross-Lingual Transfer in ASR

TL;DR

This paper introduces a zero-resource cross-lingual ASR method that uses a decipherment algorithm on unpaired speech and text data, enabling effective speech recognition without any target language transcriptions or phonetic knowledge.

Contribution

It presents the first practical zero-resource cross-lingual ASR approach that does not depend on hand-crafted phonetic information, using a novel decipherment technique on unpaired data.

Findings

Achieved near-supervised WERs with only 20 minutes of target language data.

Demonstrated the effectiveness of decipherment on unpaired speech and text data.

First practical approach to zero-resource cross-lingual ASR without phonetic knowledge.

Abstract

We present a method for cross-lingual training an ASR system using absolutely no transcribed training data from the target language, and with no phonetic knowledge of the language in question. Our approach uses a novel application of a decipherment algorithm, which operates given only unpaired speech and text data from the target language. We apply this decipherment to phone sequences generated by a universal phone recogniser trained on out-of-language speech corpora, which we follow with flat-start semi-supervised training to obtain an acoustic model for the new language. To the best of our knowledge, this is the first practical approach to zero-resource cross-lingual ASR which does not rely on any hand-crafted phonetic information. We carry out experiments on read speech from the GlobalPhone corpus, and show that it is possible to learn a decipherment model on just 20 minutes of data from the target language. When used to generate pseudo-labels for semi-supervised training, we obtain WERs that range from 32.5% to just 1.9% absolute worse than the equivalent fully supervised models trained on the same data.