Voxtract

Audio to speaker-attributed transcript — extract who said what, when.

Installation

uv pip install -e ".[all]"

Usage

# Full pipeline: transcribe + diarize
voxtract process audio.m4a --json

# Transcribe only (no speaker diarization)
voxtract transcribe audio.m4a --json

# With context hints for better accuracy
voxtract process audio.m4a --context "교합력 센서, 울산대학교" --json

Configuration

Environment variables (prefix VOXTRACT_):

Variable	Default	Description
VOXTRACT_DEVICE	auto	auto, cuda, cuda:0, cuda:1, cpu
VOXTRACT_STT_CONTEXT	""	Contextual hints for ASR
VOXTRACT_CHUNK_MINUTES	25	Audio chunk size for long files