voxint.

Voxint

Pre-alpha

Turn any recording into a speaker-labelled transcript, on your own hardware.

Open-source audio intelligence. Transcription, diarization, and speaker identification — all running locally. No cloud, no per-minute fees, nothing uploaded.

Try the Demo Read the Docs View on GitHub

Self-hosted Apache-2.0 Yours to run

voxint / transcript reviewLocal
Voxint transcript review with speaker-labelled segments, an audio waveform, and controls to verify and edit the words.
Your recording. Your speakers. Your final say.

From recording to readable

How it works

For researchers, journalists, educators, and small teams who need their recordings to stay local.

  1. 01

    Add your recording

    Upload, paste a URL, or point at a folder.

  2. 02

    Voxint does the heavy lifting

    Transcription (Whisper), speaker separation (pyannote), voice identity (TitaNet). All local.

  3. 03

    You review

    Confirm speakers, fix wording. Machine guesses stay separate from your decisions.

  4. 04

    Read or export

    On-screen read mode, or download as Markdown, plain text, subtitles, or structured data.

A closer look

Built for the review, too.

From the first setup to the last speaker decision, see what needs your attention.

Speaker review workbench showing possible voice matches and actions to confirm each speaker.
Review only the uncertain voices
Review queue listing completed recordings with speaker decisions still to make.
Your review queue
Browser setup wizard with dependency readiness checks and setup guidance.
Guided setup in the browser
Voxint home dashboard with items needing attention, quick actions, and recent activity.
Home: what needs your attention

Local by design

Your audio. Your infrastructure.

Runs on your hardware

No cloud account needed. CPU, NVIDIA, AMD, or Apple Silicon.

Vendored models, no tokens

All model weights are bundled. No Hugging Face account or download step.

Domain vocabulary

Load names, jargon, and prompts per project so specialist terms transcribe correctly.

Crash-resilient

Every run's progress lives in the database. Restart resumes where it left off.

Speaker paper trail

Machine proposals stay separate from your rulings. Full provenance.

Semantic search

Search across all transcripts by meaning, not just keywords. Also runs locally.

Make it yours

Up and running.
On your terms.

You need Docker with the Compose plugin (≥ 2.24). The installer asks a few questions, then handles the rest.

Read the installation docs
QuickstartTerminal
git clone https://github.com/bengizmo/voxint.git && cd voxint
./scripts/install.sh

No GPU? That's fine. Voxint runs the full pipeline on CPU with ~8 GB free memory.