An Exploration of Speech Conditioned Large Language Models (SLMs)
Keywords: Speech Conditioned Large Language Models
Abstract: Efforts to enable Large Language Models (LLMs) to understand human speech have spurred the development of an increasing number of Speech-Conditioned Large Language Models (SLMs). While these models have demonstrated success on various speech-related tasks, such as automatic speech recognition (ASR), the design space of SLMs has not been thoroughly explored. In this work, we revisit key design choices for SLMs, aiming to gain insights into how these choices impact the performance of SLMs and how we could optimize them for better results. Surprisingly, our experiments reveal that current SLMs struggle to follow speech instructions or respond to speech inputs, even for simple queries like ”who has been to the moon?”. Our experimental findings indicate that speech instruction following data is crucial for improving these capabilities. Leveraging this insight, we propose to use synthetic speech instruction following data to enhance speech instruction following capability. Combining the findings from our other experiments, we provide an effective recipe for developing SLMs. Our model, called SiM, not only achieves strong ASR performance, but also significantly outperforms existing SLMs in speech instruction following.
Primary Area: foundation or frontier models, including LLMs
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12789
Loading