Improving Code-Switching Speech Recognition with TTS Data Augmentation

TL;DR

The paper tackles the data scarcity problem in conversational code-switching ASR by leveraging multilingual TTS augmentation. By fine-tuning the CosyVoice TTS on SEAME and pairing it with Whisper-based ASR, the authors synthesize diverse, speaker-rich CS speech to augment real data. They demonstrate substantial MER reductions on SEAME benchmarks, reveal the importance of speaker diversity over mere data quantity, and show cross-domain transfer to ASCEND. The approach offers a cost-efficient path to improve ASR robustness in low-resource CS scenarios, with future work aimed at expanding text diversity via multilingual language models.

Abstract

Automatic speech recognition (ASR) for conversational code-switching speech remains challenging due to the scarcity of realistic, high-quality labeled speech data. This paper explores multilingual text-to-speech (TTS) models as an effective data augmentation technique to address this shortage. Specifically, we fine-tune the multilingual CosyVoice2 TTS model on the SEAME dataset to generate synthetic conversational Chinese-English code-switching speech, significantly increasing the quantity and speaker diversity of available training data. Our experiments demonstrate that augmenting real speech with synthetic speech reduces the mixed error rate (MER) from 12.1 percent to 10.1 percent on DevMan and from 17.8 percent to 16.0 percent on DevSGE, indicating consistent performance gains. These results confirm that multilingual TTS is an effective and practical tool for enhancing ASR robustness in low-resource conversational code-switching scenarios.

Figures (1)

• Figure 1: End-to-end synthetic-data pipeline. Ground-truth text and speech are tokenised, passed through the Qwen-2 language model, a flow-matching decoder and a HiFT vocoder to yield synthetic audio, which is later used to fine-tune Whisper.