Offline model, re-run on the current utterance several times a second (grey text); an utterance is committed on a pause or when it reaches the length cap, and its audio is dropped. The model file is fetched once and kept in this browser.
Models: NVIDIA Parakeet (parakeet-tdt_ctc-110m, parakeet-tdt-0.6b-v2), licensed CC BY 4.0, converted to GGUF by handy-computer; the 4-bit (Q4_0) files were quantized from those conversions. Runtime: transcribe.cpp and ggml (MIT), modified and compiled to WebAssembly; licence texts.
Runs the model back to back over a test clip cut to a different length each time (0.3 s up to the utterance cap), without the microphone, and shows what the GPU backend holds after every pass. Buffers made and working memory should stop changing after the first few passes.
passes