Morph & Beyond

We open-sourced Addis Scribe Streaming, realtime speech recognition for Amharic. It beats Google Chirp 3 on both accuracy and speed: Word error rate: 28.4% vs 32.5% First words on screen: 1.56s vs 6.67s Final text after you stop talking: 30ms vs 1,411ms On natural speech (WAXAL) we beat Google Chirp 3 in both streaming and batch mode. On FLEURS we went from 35.0% to 19.8% WER. Hohe remains the most accurate offline Amharic model in our WAXAL test, and on FLEURS the two are within a point of each other (19.8% vs 20.6%). We built Addis Scribe for the live and edge inference side, where text has to appear while people are still talking. Trained on 5,000+ hours of Amharic speech. Test sets were never used in training. Per-clip results and scripts are in the repo. Try it😊: https://docs.addisassistant.com/docs/playground/speech-to-text API and SDK docs: docs.addisassistant.com HuggingFace🤗: huggingface.co/addisai/addis-scribe-streaming @addisassistantai