CJST: CTC Compressor based Joint Speech and Text Training for Decoder-Only ASR

TL;DR

This work addresses the challenge of integrating audio encoders into decoder-only ASR models without external language models and without explicit duration handling. It introduces CJST, a CTC compressor-based joint speech-text training framework that uses a simple modality adaptor, on-the-fly forced peaky alignment, and CTC class embeddings to match speech and text representations in both directions. A comprehensive study identifies the blank-probability removal mode with a high threshold (0.95) as the most robust compressor configuration, and CJST consistently improves both in-domain and cross-domain WER over LM-like text injection and adaptor baselines on LibriSpeech and TED-LIUM2. The approach leverages paired and text-only data, demonstrates effective text injection without duration handling, and provides guidance on robust compressor settings and edge-case handling for decoder-only ASR systems. Overall, CJST offers a practical path to deploy high-performing decoder-only ASR without external LMs, enhancing text integration and cross-domain adaptability.

Abstract

CTC compressor can be an effective approach to integrate audio encoders to decoder-only models, which has gained growing interest for different speech applications. In this work, we propose a novel CTC compressor based joint speech and text training (CJST) framework for decoder-only ASR. CJST matches speech and text modalities from both directions by exploring a simple modality adaptor and several features of the CTC compressor, including sequence compression, on-the-fly forced peaky alignment and CTC class embeddings. Experimental results on the Librispeech and TED-LIUM2 corpora show that the proposed CJST achieves an effective text injection without the need of duration handling, leading to the best performance for both in-domain and cross-domain scenarios. We also provide a comprehensive study on CTC compressor, covering various compression modes, edge case handling and behavior under both clean and noisy data conditions, which reveals the most robust setting to use CTC compressor for decoder-only models.

Figures (2)

• Figure 1: Overview of decoder-only model structure and CTC compressor.
• Figure 2: Proposed CJST framework