Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper introduces Multimodal Depth Upscaling, a method for adapting pre-trained text LLMs to speech by inserting and training new transformer layers (including E-Branchformer) on speech data while keeping the original text LLM frozen. This approach mitigates the text performance degradation typically observed in full fine-tuning or LoRA. Experiments on SmolLM2 models show that depth up-scaling achieves comparable or better ASR performance with significantly less text degradation and fewer trainable parameters compared to full fine-tuning.
Forget catastrophic forgetting: inserting speech-specific layers into frozen text LLMs lets you build strong speech models without trashing your original text smarts.
Adapting pre-trained text Large Language Models (LLMs) into Speech Language Models (Speech LMs) via continual pretraining on speech data is promising, but often degrades the original text capabilities. We propose Multimodal Depth Upscaling, an extension of an emerging strategy in continual LLM pre-training, where new transformer layers are inserted into a frozen text LLM and only the added layers are trained on speech data. Experiments with SmolLM2-360M and SmolLM2-1.7B on 48k hours of English Automatic Speech Recognition (ASR) data show that depth up-scaling achieves ASR comparable to full fine-tuning while causing far less text degradation than both full fine-tuning and Low-Rank Adaptation (LoRA). We further show that incorporating E-Branchformer, an architecture designed for speech recognition, as the inserted layers achieves ASR that matches or surpasses full fine-tuning on the larger model while reducing text degradation by over 75% with 60% fewer trainable parameters.