SignalOpen source, AGPL-3.0

Real-time PII detection, by the texture of speech.

An open-source baseline for flagging sensitive data in live voice, without transcribing it. Self-host it, and help improve it.

Get it on GitHub See integration

01Why texture

Transcription is too slow for live voice.

Redacting PII by transcribing first adds more than a second of latency. Signal reads the acoustic texture directly.

Status quo: transcription

Speech to text, +800ms
NLP analysis, +300ms
Redaction

~1.2s

Humsana Signal

Texture analysis, 10ms
Signal block, instant

~10ms

02Integration

Any platform that speaks WebSockets.

Point your media stream at a Signal engine and mute audio the moment PII is detected. Works with Vapi, Twilio, or a raw WebSocket client.

import websockets

async with websockets.connect("wss://your-host/stream?key=YOUR_KEY") as ws:
    await ws.send(json.dumps({"audio": chunk_b64, "format": "pcm16"}))
    if "PII_DETECTED" in await ws.recv():
        mute_audio()

Open source, like Swan.

A working baseline you can run and improve. For a managed deployment or a pilot, get in touch.

View the repository Contact us