Audio8 ASR Infinite
A native streaming speech-recognition model that transcribes audio of unlimited length continuously, with a rolling key-value cache that keeps memory and latency bounded even in 24-hour operation.
Audio8 ASR Infinite decodes 12.5 times per second with a selectable audio clock of 80, 120 or 160 milliseconds and a configurable transcription delay from 240 to 560 milliseconds, trading latency for accuracy. A semantic voice-activity detector distinguishes a thinking pause or stuttering from a real end of turn, which the project says traditional acoustic voice-activity detection fails to do, and the model transcribes Chinese and English.
This preview release delivers the streaming transcription base, built on a causal audio tower initialized from Voxtral's realtime audio architecture and a Qwen2.5-3B-Instruct decoder, with the audio projector and frame-length embedding trained from random initialization. A planned formal release adds frame-level semantic perception beyond transcription on the same frame grid and acoustic forward pass. It is licensed Apache-2.0 and written in Python.