// 100% local
Is it really 100% local?
All processing happens on your machine, with no telemetry. Internet is needed only on first run to download Python, deps and models; after that it runs fully offline.
no telemetry · offline after first run
PLYNTE LIVEAUDIO v1.2.0 · WINDOWS & LINUX
Real-time local captions for your OBS.No cloud, no subscription, no API key.
LiveAudio is a free, open-source (MIT) app that generates real-time Whisper speech captions 100% locally and streams them to OBS over a local WebSocket. It runs on Windows and Linux. No cloud, no API key, no per-minute cost.
- 100% local
- MIT open-source
- Windows + Linux
subtitulos_obs.html
ws://127.0.0.1:8765
- id1046text"and that is how the local…"latency1.08stotal_delay1.27sis_replayfalse
- id1045text"Whisper turbo is decoding…"latency1.12stotal_delay1.32sis_replayfalse
- id1044text"Silero VAD just trimmed t…"latency1.16stotal_delay1.37sis_replayfalse
- id1043text"captions are streaming st…"latency1.20stotal_delay1.42sis_replaytrue
Live preview · illustrative
- 100% local
- MIT open-source
- Windows + Linux
- No cloud
- No API key
// why LiveAudio
Straight answers to the real objections.
// linux
Does it work on Linux?
Yes. Linux x86_64 captures from the microphone and needs libportaudio2. System-audio loopback (WASAPI) is Windows-only.
linux x86_64 · sudo apt install libportaudio2
loopback Windows-only// gpu
Do I need a GPU?
No. The CPU works. NVIDIA CUDA is optional but recommended: it is auto-detected and needs driver ≥ 525 and VRAM ≥ 4 GiB.
cpu works · cuda driver ≥ 525 · vram ≥ 4 GiB
// the signal path
From microphone to OBS in four stages.
In plain terms: LiveAudio listens to your voice, turns speech into text on the spot, and sends those captions to OBS — all on your own machine. That speech-to-text engine (ASR, automatic speech recognition) runs locally in four stages, kept steady by isolated processes, an audio ring buffer and automatic reconnection.
-
Capture
Grab a physical microphone or system audio (WASAPI loopback on Windows; mic on Linux). The ring buffer keeps audio flowing even under load.
mic · system loopback (windows) -
Gate
Silero VAD trims silence with a configurable onset pre-roll and VAD threshold, so Whisper only decodes real speech.
silero-vad · onset pre-roll + threshold -
Decode
This is where speech becomes text. Whisper transcribes in real time — tiny, base, small or turbo — on CPU or optional CUDA. A hallucination blacklist filters junk text.
whisper · tiny / base / small / turbo -
Broadcast
The captions go straight to OBS over a local connection — clean subtitle JSON that OBS or any app on your computer can read. Sub-second to ~1 s, tunable per profile. (The exact address is in small print below.)
ws://127.0.0.1:8765
LIVE DEMO
See LiveAudio running live.
A real screen recording of the app generating captions — not a mockup.
real capture
// integration
Works with OBS — or any WebSocket client.
LiveAudio broadcasts clean subtitle JSON over a local WebSocket (ws://127.0.0.1:8765). OBS is the built-in target via the included subtitulos_obs.html overlay, but any HTML or WebSocket client on localhost can connect and receive the same broadcast — so you can build your own consumer or wire it into other local tools. Connections are accepted only from localhost; no auth needed.
{
"id": 482,
"text": "real-time local captions",
"style": "live",
"latency": 1.1,
"total_delay": 1.3,
"is_replay": false
}// in the box
Seven built-in tools: adaptive ribbon overlay, hot-swap device & model, FPS-aware profiles, hallucination blacklist, session files, local-first diagnostics, in-app updates.
// obs overlay new in v1.2.0
Adaptive ribbon overlay
The OBS overlay adds an adaptive vertical “ribbon” subtitle buffer, with improved subtitle legibility and capped reveal-animation timing.
// hot-swap
Hot-swap device & model
Change the audio device or Whisper model without restarting. Apply changes and the engine reconfigures live.
// profiles
FPS-aware profiles
Fast, Balanced, Quality and Stable Streaming presets balance latency against GPU load while you game.
// blacklist
Hallucination blacklist
An editable blacklist filters out the junk phrases Whisper invents on silence — and you control the list.
// session files
Session files
Every run saves transcript.jsonl and subtitles.vtt (plus session.json) so nothing valid is ever lost.
// diagnostics
Local-first diagnostics
Export sanitizes secrets and paths and excludes raw audio and full transcripts. No telemetry, ever.
// updates
In-app updates
One-click in-app update, or run the launcher with --update. New versions land without a reinstall.
// interactive sandboxes
Six controls, six live sandboxes.
Every LiveAudio setting maps to a knob you can feel before you download. Profiles, Whisper models, the hallucination blacklist, OBS backlog policy, hot-swap devices, and Silero VAD — try each one below.
sandbox active
// profiles
FPS-aware profiles
Fast, Balanced, Quality, and Stable Streaming retune latency, model, and silence gate together. Editing a built-in makes a Custom profile; apply changes to activate.
The recommended balance of speed and accuracy.
Latency~0.6 s
VRAM~480 MB
VAD cutoff0.8 s
statusready
FPS-aware presets: Fast / Balanced / Quality / Stable Streaming.
// models
Whisper model footprint
tiny, base, small, and turbo trade resource footprint for accuracy. Documented install footprint is ~400 MB CPU / ~2.5 GB CUDA; per-model VRAM below is illustrative.
Sizing tip
small / base is the streamer sweet spot: fast transcription, modest footprint. GPU optional — CPU works fine.
// blacklist
Hallucination blacklist
Whisper sometimes emits filler from training data. An editable, comma-separated blacklist strips those phrases before captions reach OBS.
Load the demo and run the filter to see blacklisted phrases removed.
illustrative
// backlog
OBS backlog policy
When the engine falls behind, the backlog policy controls only what shows live in OBS — everything valid is always saved to your session files.
ASR bufferBuffer empty
OBS output
Waiting for audio…
Live policy only changes the OBS view — every valid line is still saved.
// hot-swap
Available inputsHot-swap devices
Switch microphone or system audio without restarting. Pick the active input; the engine swaps capture live with no dropped captions.
Active inputMicrophone (Realtek Audio)
illustrative
// silero vad
Silero VAD gate
Silero VAD trims silence and non-speech so Whisper only runs on real voice. v1.2.0 adds a configurable onset pre-roll and VAD threshold.
Send a signal to see the VAD gate pass speech and block noise.
Powered by the local Silero VAD model.
// the short version
How LiveAudio compares at a glance.
| LiveAudio | Cloud ASR / plugins | |
|---|---|---|
| 100% local | Yes: | cloud-dependent |
| $0 per-minute cost | Yes: | billed per minute |
| No API key required | Yes: | usually required |
| System-audio loopback | Conditional: Windows-only | varies |
// answers
The questions people ask first.
Is it really 100% local?
Yes. All processing happens on your machine, with no telemetry. Internet is only needed on first run to download Python, deps and models; after that it runs fully offline.
Does it work on Linux?
Yes, on Linux x86_64 with microphone capture (needs libportaudio2). System-audio loopback is Windows-only.
Is it free and open-source?
Yes. LiveAudio is free and open-source under the MIT license. No subscription and no API key.
What's the latency?
Low, tunable latency — well under a second on a typical setup. Profiles trade latency against accuracy and GPU load.
// download
Local captions in your OBS in minutes.
Free & open-source (MIT). No subscription, no API key. You only pay your own electricity — hardware not included.