SEAM: Bridging the Temporal-Semantic Granularity Gap for LLM-based Speech Recognition

AbstractSpeech-LLM integration faces a temporal-semantic granularity gap: speech representations scale with temporal duration while text tokens scale with semantic content. Existing duration-based methods generate embeddings at fixed rates, creating distributional mismatch with LLM pre-training. We propose SEAM (Speech Encoder-Decoder Alignment Module), an encoder-decoder architecture employing variable-rate generation through cross-attention between speech features and text embeddings. SEAM produces embeddings at adaptive rates that closely match natural text distributions while preserving pre-trained knowledge by freezing both speech encoder and LLM. We introduce a multi-stage training strategy and First Token Guidance to improve initial token prediction. SEAM achieves competitive performance on LibriSpeech (2.6%/5.2% WER). More significantly, trained only on LibriSpeech (960h), SEAM achieves 4.7% WER on cross-domain TED-LIUM-v2, demonstrating that integrating LLM’s linguistic knowledge enables effective generalization beyond limited speech training data.