Voxint documentation¶
New to Voxint? Start here:
| Doc | Contents |
|---|---|
setup.md |
Install Voxint on your operating system and hardware: prerequisites, guided vs. manual install, and every compute tier (CPU / NVIDIA / AMD / Apple) |
onboarding.md |
First-run path once it's installed: guided installer, setup wizard, and the bundled guided tutorial |
how-to/ |
Day-to-day task guides for non-technical operators: add media & manage runs, review & adjudicate, manage speakers & export, translate transcripts, check for AI-generated speech, settings & troubleshooting, and (advanced) changing pipeline models |
Reference & internals:
| Doc | Contents |
|---|---|
architecture.md |
Pipeline stages, state machine, data model |
adr/ |
Architecture decision records: the load-bearing decisions behind significant changes (media identity, project membership, editor and claim contracts, speaker provenance, plugin scope) |
semantic-search.md |
Transcript semantic search: what the embedding index is, building/refreshing it with voxint embed backfill, and the weights requirement on native installs |
annotations.md |
Operator annotations: the anchor contract (kinds, coordinate mapping, hashing, staleness, refresh, API taxonomy) |
domain-packs.md |
Domain packs: manifest, resolution, the per-run frozen snapshot, and which fields shape which stages |
plugins.md |
Plugin author guide: writing, registering, and shipping a plugin |
gpu-contracts.md |
Versioned HTTP contracts for the ASR / diarizer / embedder services |
gpu-smoke.md |
Build + real-inference smoke procedure for the GPU service images |
quality-gates.md |
Enhancement failure semantics, matching eligibility + grounding gates, confidence semantics |
enrichment-triage.md |
Draft triage: the read-time, explainable review-priority score, its components, and the source-authority allowlist |
interpreting-diarization.md |
Reading the output: segment labels vs the turn ledger, short-clip over-splitting, and giving the pipeline a speaker count when it over-segments |
timeouts-and-leases.md |
COMPUTE_TIER timing profiles: stage timeouts and worker leases per tier |
harness.md |
Offline scoring harness: voxint score file contracts, verdict vocabularies, the cross-space invariant, feeding it from live runs (voxint.harness_export + the export_match_evidence driver), and the attribution evaluation layer (AMI recurrence, protocol manifests, gold-to-slot alignment, the eval_attribution CLI driver) |
benchmark.md |
Shipped benchmark corpus, CLI runner, WER scorer, and cross-run comparison |
operations.md |
Deployment, migrations, pipeline operations, recovery, adjudication workflow, multi-user authentication |
service-controls.md |
Enable model service restarts from Settings > Status, manual restart commands, and troubleshooting |
native-macos-preview.md |
Technical preview: run the whole stack on macOS/arm64 without Docker, under launchd (scripts/native/voxint-native.sh) |
testing.md |
Test layers and how to run them; the manual browser-verification procedure for the review console; the offline eval-quality harness (DER/JER/WER/cpWER against AMI/VoxConverse); the attribution eval harness (FAR/FRR/coverage, frozen regression pack) |
release-process.md |
Maintainers: how a release is cut (tag → GHCR images → PyPI → GitHub Release) |
security/ |
Security audits calibrated to the single-operator threat model (predates multi-user mode): the whole-repo audit, its threat model, and the standing findings plus their remediation status |
reports/ |
Dated measurement reports (parity screens, bakeoff diagnostics, negative results): the evidence behind verdict blocks in gpu-contracts.md |
../examples/ |
End-to-end voxint score walkthrough on a synthetic dataset |
Writing or editing docs? Follow the two-lane house style described in
the Documentation section of CONTRIBUTING.md.
Agents should load the voxint-docs skill first.