v0.2.0

What's new

vLLM backend for transcription

CohereX can now offload the ASR step to a vLLM server via its
OpenAI-compatible /v1/audio/transcriptions endpoint. VAD, forced alignment, and
diarization still run locally, so only the model inference moves.

Two ways to use it:

• Connect to a server you already run:

  coherex audio.mp3 --language en --backend vllm --vllm_url http://localhost:8000

• Let CohereX manage the server — it starts vllm serve, waits until ready, uses it,
  and shuts it down automatically when the run finishes (including on errors and Ctrl-C):

  coherex audio.mp3 --language en --backend vllm

New CLI options

• --backend {local,vllm} — default local (in-process, unchanged behavior).
• --vllm_url — URL of a running vLLM server; omit to auto-launch one.
• --vllm_api_key — API key if the server requires one (falls back to VLLM_API_KEY).
• --vllm_args — extra flags forwarded to vllm serve, e.g. --vllm_args "--gpu-memory-utilization 0.8".

Installation extras

• pip install "coherex[vllm]" — installs the vLLM backend dependencies.
• pip install "coherex[all]" — language detection and vLLM together.

Notes

• The default backend stays local; existing commands and the Python API are unchanged.
• --language auto works with the vLLM backend as well.

v0.1.0

Prepare v0.1.0 for PyPI