voicekit-team/T-one

# T-one Technical Summary Built on a Conformer acoustic model with CTC output, the pipeline processes audio in 300 ms chunks and maintains hidden state across segments to preserve acoustic context. A custom log-probability splitter detects phrase boundaries in real-time, then converts logprobs to text via greedy decoding or KenLM-based beam search. Containerized deployment via Docker, with Python inference APIs and Triton Inference Server support for production scaling.

249 stars.