Voxint
Turn any recording into a speaker-labelled transcript, on your own computer. Voxint transcribes your audio or video, works out who spoke when, and gives you a simple review screen to confirm the speakers and fix the wording before you export.
It is built for individuals and small teams (researchers, journalists, educators) who need their recordings to stay local. Your audio is transcribed on your own hardware by default: no cloud account, no per-minute fees, nothing uploaded. (Two optional features do reach the network when you turn them on: fetching a recording from a URL, and sending transcript text to an outside AI model to polish it. Both are off or opt-in, and clearly labelled.)
Early days (pre-alpha). Voxint works end to end, but it is young software. The interface, database, and settings can still change between releases through the 0.x series. Great for hands-on use and feedback; not yet for mission-critical archives.
See it in action
(All screenshots use small synthetic sample recordings, never real audio.)
 |
 |
| Your review queue: completed recordings waiting for your decisions. | Review only the uncertain voices: strong matches are shown automatically; you confirm the possible ones and rule on the rest. |
 |
 |
| Guided setup in the browser: honest readiness checks, plain-language fixes. | Home shows what needs your attention and how to add a recording. |
What it does
Voxint takes a recording and walks it through four steps:
- Add your recording: upload it in the browser, paste a URL, or point Voxint at a file it can already see.
- Voxint does the heavy lifting: it transcribes the words and works out who spoke when, then suggests who each voice is.
- You review: confirm each speaker, and fix any wording, in a screen built for exactly this. Machine guesses stay separate from your decisions; you always have the final say.
- Read or export: read a finished transcript on screen, or download a clean, speaker-labelled copy (plain text, Markdown, subtitles, or structured data).
Once you have a few transcripts, you can also search across all of them by meaning, not only by the exact words. Type what you are looking for and Voxint finds the closest passages from every recording, each with a link straight to that spot. This search runs on your own hardware too, with no outside AI.
The models all run locally: transcription (Whisper), speaker separation (pyannote), and voice identity (TitaNet). Everything they need is bundled into Voxint, so there is no Hugging Face account or token to set up.
Quickstart
You need Docker with the Compose plugin (≥ 2.24). One command takes a fresh copy to a running console:
git clone https://github.com/bengizmo/voxint.git && cd voxint
./scripts/install.sh
The installer asks only for what it can't invent (an admin password, a folder for your media, and which hardware runs the models), then generates everything else, starts Voxint, waits until it is healthy, and prints the console address.
On an Apple Silicon Mac and would rather not install Docker Desktop? A docker-free native preview runs the whole stack under macOS's own service manager instead. It is a macOS-only technical preview (a few shell commands, not the one-command install above), so read docs/native-macos-preview.md if that is you. Every other install path, on any operating system, needs Docker.
No graphics card? That's fine. Voxint runs the whole pipeline on an ordinary computer's CPU (needs roughly 8 GB of memory free). It is slower, and a long recording can take hours rather than minutes, but it works anywhere. A GPU (NVIDIA, AMD, or an Apple Silicon Mac) just makes it faster.
Have a GPU? How much VRAM you need. The transcription suite (Whisper + pyannote + TitaNet) shares one card and fits comfortably on 8 GB (e.g. RTX 3050/3060 Ti/4060). Turning on the optional bundled local LLM adds ~5 GB, so running everything on one card wants 12 GB (e.g. RTX 3060 12 GB) or more. AMD cards work via the ROCm tier. Full breakdown and card examples → docs/setup.md.
When it finishes, open the console at http://127.0.0.1:8080/ and sign in
with the username and password you set. On a fresh install Voxint walks you
through a short in-browser setup wizard and an optional guided tutorial
on the bundled sample, so you see the whole review loop before pointing it at
your own audio.
Full setup for your operating system and hardware → docs/setup.md. First-run walkthrough → docs/onboarding.md.
Using Voxint
Once it is running, these short guides cover the day-to-day tasks:
- Add media & manage runs: upload a file, paste a URL, or watch a folder; follow a run and requeue, cancel, or archive it.
- Review & adjudicate: confirm speakers, correct the transcript, keyboard shortcuts, the waveform, splitting and reassigning segments.
- Manage speakers & export: the speaker roster and the export formats.
- Settings & troubleshooting: configure everything from the browser, and fix common problems.
A little more depth
You don't need any of this to use Voxint, but if you're curious or evaluating it:
- Nothing is lost to a crash. Every run's progress lives in a database, so a restart resumes where it left off, and pausing for review is just a saved row.
- Speaker identity keeps a paper trail. Voxint matches voices against a roster that grows as you use it, and keeps machine proposals strictly separate from your rulings.
- Optional AI polish. Voxint can send transcript text to any OpenAI-compatible model to tidy it up and suggest names. It is off by default, and a slow or failing model never blocks a run.
- Swappable vocabulary. Names, jargon, and prompts load from a domain pack you can pick per folder, so specialist terms transcribe correctly. See docs/domain-packs.md.
- Held to measured gates. The models are pinned and their outputs are held to measured-equivalence gates, so an upgrade can't quietly change results. See docs/gpu-contracts.md.
- Optional multi-user mode. By default a single admin password protects the
console. Turn on
VOXINT_MULTI_USERand each reviewer gets their own login, with decisions attributed to the person who made them. See docs/operations.md.
For developers
The console is server-rendered (FastAPI + Jinja + htmx) with a few small React "islands"; the model services are separate containers behind versioned HTTP contracts. To run the code you checked out instead of the release images, layer the build overlays; to work without Docker at all:
uv sync --extra dev
uv run pytest tests/unit
uv run uvicorn voxint.api.app:app --reload
There is also a standalone, database-free scoring harness. pip install voxint
gives you the voxint score CLI for speaker-attribution metrics (see
examples/). Architecture, contracts, operations, and the
release process are documented under docs/.
License
Apache-2.0. See LICENSE and NOTICE. Vendored model weights
are redistributed under their own licenses with attribution (titanet:
CC-BY-4.0; pyannote segmentation: MIT; WeSpeaker embedding: CC-BY-4.0). See the
provenance files under services/*/models/ and the model-asset releases.
1 maintainer
Metadata
1 maintainer
Release files for voxint 0.41.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| voxint-0.41.0-py3-none-any.whl | Python 3 | none | any | Details |
Release files / voxint-0.41.0-py3-none-any.whl
| Download URL | voxint-0.41.0-py3-none-any.whl |
|---|---|
| Size | 4.6 MB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
097631ab4b14527bb6732133f7b3c2284b4dce6625c01a91bd62e691dfd4cea7
|
|
BLAKE2b-256 checksum How to use checksums |
7583b4b86271f0e4113a208c3ac1f4fdab010a18f033c17e41102eba791249a7
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
uv/0.11.28 {"installer":{"name":"uv","version":"0.11.28","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"26.04","id":"resolute","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":null}
|
Security for voxint 0.41.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
