End-to-end speech-to-text retrieval is emerging as a key feature for voice agents, allowing spoken queries to find relevant documents without first converting speech to text. The new paper introduces SSEmb, a model designed for this task. Its central claim is that such a retriever can be trained using only automatic speech recognition (ASR) data.
The abstract notes that training these retrievers typically depends on something—the text is cut off in the source—but SSEmb's approach is to use ASR data alone. The goal is a generalizable model that can handle spoken queries directly.
Because the source is only an abstract, details about the training procedure, architecture, and evaluation results are not available. The significance lies in the suggestion that ASR data, which is more abundant than paired speech-to-text retrieval data, might be sufficient for building effective retrieval models.