Desert Ant Labs introduced the Uhm model for detecting parasitic words

Company Desert Ant Labs released Uhm — a specialized model for local audio stream processing. The tool is designed for automatic detection of filler words such as uh and um in real time. Architecture of the model is optimized to work directly on the device, which eliminates the need to transfer audio data to the server. The system provides detection accuracy with a delay of only 20 ms. Developers claim that the model was trained using acoustic transfer methods, which allows it to effectively recognize pauses and filler sounds in English and a number of other languages. Technical parameters of Uhm allow integrating it into transcription systems and tools for editing podcasts or video content. The model demonstrates high performance, reaching a score of 8 points when processing audio for 1 hour. Weight and computational resource requirements allow running the solution on consumer equipment without specialized accelerators. The emergence of such narrowly specialized models simplifies post-processing of audio recordings. Using local solutions for cleaning speech from filler words reduces the load on cloud APIs and increases data confidentiality when working with personal or corporate audio materials.