Researchers probing the SLAM-ASR framework asked whether training a linear projector to map speech encoder features into a frozen LLM decoder for transcription also enables spoken language understanding. They trained monolingual projectors on 11 languages and then tested the models on yes/no and topic-selection questions about the input speech.

The results show a clear gap. When prompted with a yes/no question, the mean instruction following rate was only 49.8%; when the model ignored the instruction, it fell back to transcribing the audio. For topic questions phrased as a single-digit option, instruction following rose to 81.3%, but mean balanced accuracy was just 23.2% against a 14.3% chance baseline. The same decoder, given reference transcripts as text input, reached 71.5% balanced accuracy on the same topic task.

The authors conclude that the projector fails to transfer the information needed for spoken language understanding, even though transcription quality—measured by character error rate—was not a reliable predictor of instruction following. The work was accepted to the EMNLP 2026 IMPACT-SPEECH Workshop.