🤖 AI Summary
Addressing core challenges—including weak speech-text cross-modal alignment, low speech synthesis quality, and information density mismatch—this paper proposes the Speech-Language Modeling (SLM) framework for unified speech-language representation. Methodologically, it introduces fully decoupled speech and text tokenizers, adopts a multi-token prediction (MTP) paradigm to mitigate semantic density disparities across modalities, and constructs RoleTriviaQA—the first large-scale spoken question-answering benchmark incorporating speaker identity. The framework integrates a speaker-aware generation architecture with an LLM-centric training strategy. Experimental results demonstrate substantial improvements: word error rate drops significantly (6.07 → 3.01), speech synthesis quality and decoding speed improve concurrently (up to 12× acceleration), and on RoleTriviaQA, both knowledge comprehension accuracy and speaker consistency achieve marked gains.
📝 Abstract
Speech-language models (SLMs) offer a promising path toward unifying speech and text understanding and generation. However, challenges remain in achieving effective cross-modal alignment and high-quality speech generation. In this work, we systematically investigate the impact of key components (i.e., speech tokenizers, speech heads, and speaker modeling) on the performance of LLM-centric SLMs. We compare coupled, semi-decoupled, and fully decoupled speech tokenizers under a fair SLM framework and find that decoupled tokenization significantly improves alignment and synthesis quality. To address the information density mismatch between speech and text, we introduce multi-token prediction (MTP) into SLMs, enabling each hidden state to decode multiple speech tokens. This leads to up to 12$ imes$ faster decoding and a substantial drop in word error rate (from 6.07 to 3.01). Furthermore, we propose a speaker-aware generation paradigm and introduce RoleTriviaQA, a large-scale role-playing knowledge QA benchmark with diverse speaker identities. Experiments demonstrate that our methods enhance both knowledge understanding and speaker consistency.