Sota សោតៈ · Khmer speech recognition
Sota (សោតៈ, "ear") is a small Khmer + English speech recognition model, trained from scratch: 26M parameters, ≈33 MB int8 ONNX, runs on an ordinary CPU. v1 is the released model (test CER 3.40%). v2 is an experiment with more and harder data (test CER 4.26%, so v1 is still better). Pick one, or compare both. Website
Press Record and speak. Text appears while you talk (the line ending in … is still listening) and is finished with the more accurate beam search after each short pause.
Click an example below, then Transcribe. Clips: FLEURS Khmer test (Google) and LibriSpeech, both CC BY 4.0.