LASER: Language-Queried Speech Enhancer

The performance of speech enhancement systems may be improved by the introduction of prior signal and noise information into their operation. Including this information as a textual query is desirable, since it is a flexible, intuitive and interpretable interface for human-machine interaction. This has recently gained attention in the field of Language-queried Audio Source Separation (LASS). In this work, we propose LASER, a LAnguage-queried Speech EnhanceR that extends the LASS framework to the speech enhancement task and operates on commands including speech transcription and noise information. Experiments on the VoiceBank-DEMAND dataset confirm that an appropriate approach to include the aforementioned information results in increased perceptual quality in comparison to baseline enhancement approaches, with and without text conditioning.