User repository on GitHub
Low-latency audio censorship POC using Jev typed decisions and ffmpeg.
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Tools & CLIs
Low-latency audio decisions POC
Private TypeScript proof of concept that flags Spanish profanity with Jev Boolean and Score questions and replaces the matching word intervals with a beep via ffmpeg, keeping the original duration.
- Stars
- 2
- Language
- TypeScript
- Reading time
- 4 min
- Published by
- Community
- Author
- santos-sanz
- Last verified
- Sep 2026
About this project
Private TypeScript proof of concept: identify Spanish profanity with Jev's typed probabilistic decisions and replace the matching word intervals with a beep without changing the audio duration. The checked-in demo uses a timestamp sidecar so the Jev and DSP path is reproducible without tying the POC to an ASR vendor. AsrProvider is the seam for plugging in a streaming ASR that emits partial word timestamps. A TypeSafe AI API key with Jev access.
Quoted from the project's README. Read the full README.
Highlights
- Each word is evaluated in its local context with two typed questions: isInsult (Boolean) and severity (Score with four ordered levels).
- The pipeline goes ASR timestamps to parallel Jev evaluations, threshold padding and merging, then an ffmpeg beep overlay.
- The ASR provider is a seam for streaming; the README sketches a live version with a 300-600 ms PCM ring buffer.
- The default 0.72 threshold is explicitly a POC starting point, not a claim of production calibration.
- Uses the official @ai-sdk/typesafe-ai provider with the System One endpoint and jev-latest, overridable with JEV_MODEL.
Quickstart
bash
npm install
export TYPESAFE_AI_API_KEY='YOUR_API_KEY'
npm run check
npm run demoWatch out
No license file, so reuse terms are unclear. Needs Node.js 22+, ffmpeg on PATH and a TypeSafe API key; audio and transcript text are sent to providers, and the demo relies on a word-timestamp sidecar.
Questions about this entry
Short answers drawn from the linked project or page.
No keyword list appears in the pipeline. Every word is evaluated in its local context by a Boolean isInsult question plus a four-level severity Score, then intervals are padded, merged, and overlaid as a beep through ffmpeg. The demo ships a word-timestamp sidecar so the decision and DSP path is reproducible without an ASR vendor.
Only whole files today. The checked-in demo reads a WAV end to end. The README describes a live version as three swaps: an ASR adapter emitting stable partial words every 100-250 ms, per-window bounded queues with one request in flight, and a 300-600 ms PCM ring buffer. That streaming path is a design sketch, not implemented code.
A false positive destroys audio, which the README calls out as the expensive failure. The default threshold of 0.72 is described as a POC starting point rather than production calibration, and it recommends hysteresis so a lower bar extends an adjacent span without starting a new one. Calibration by language, dialect, and quoted speech is left to you.
Both are sent to whichever providers you configure, and the README warns against logging raw audio or keys. Nothing about retention, encryption, or deletion is implemented. It only advises setting a retention policy, encrypting temporary media, deleting it promptly, and verifying the provider's data-handling terms for your account first.
Node.js 22 or newer, ffmpeg on PATH, and a TypeSafe AI key with Jev access. `npm install`, export TYPESAFE_AI_API_KEY, then `npm run check` and `npm run demo`. It writes output/prueba-censurada.wav plus a .wav.json audit report holding each word, its Jev probability, score and confidence, and the final beep intervals. For another recording, npx tsx src/cli.ts takes a WAV, a --transcript sidecar, --out, and --threshold.
The official @ai-sdk/typesafe-ai provider with the AI SDK evaluation API, against the System One endpoint and the jev-latest alias, which JEV_MODEL overrides. Every word is evaluated against its local context using two typed questions: isInsult, a Boolean returning P(true), and severity, a Score over four ordered levels with TypeSafe confidence metadata. ASR sits behind its own seam, with AsrProvider as the interface a streaming provider would implement.
There is no license file in the repository, so the terms are unclear rather than permissive. The project describes itself as private and a proof of concept, with an explicit 0.72 threshold it calls a starting point and not production calibration, so treat the pipeline as a reference implementation to read rather than a dependency to depend on.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
YouTube: Building a contextual brand-safety scorer on Jev from scratch.
X: A website slop detector that checks 35 tells in 243 ms
X: Breaking, golden nugget, or slop
X: A filter for reply-guy comments
More Jev resources like this
A self-hosted Discord bot that classifies messages with Jev in real time, deletes spam and scam links, escalates repeat offenses through warnings to timeouts, and profiles members from message history.
Jev Tools & CLIs#community#python#moderation
A research-stage proposal-review contract in TypeScript: an LLM proposes one action, Jev answers four pinned questions, and code produces evidence the host can consider.
Jev Tools & CLIs#community#typescript#agents
Community
TypeSafeAI/jev-harnessTypeScript
TypeSafeAI/jev-harnessTypeScriptAn open-source Discord moderation bot powered by Jev. Owners sign in to a hosted dashboard, add the bot, supply their own TypeSafe key, test in Monitor only mode, then let Jev rules log, delete or time out spam, phishing, harassment, hate and threats.
Jev Tools & CLIs#community#github#typescript
Community
i098/jev-modTypeScript
i098/jev-modTypeScriptA guard that stands outside an LLM agent and uses TypeSafe Jev to score both incoming prompts and proposed actions, turning the score into a block, a question, or a pass; in the README's Gandalf test it rejected 17 of 21 published injection prompts before send.
Jev Tools & CLIs#community#github#security
Community
JamesANZ/JevPromptShieldTypeScript
JamesANZ/JevPromptShieldTypeScriptA Pi coding-agent extension with Jev as a decision layer: a gate judges bash, write, and edit calls before they run, and an output judge flags leaked secrets and classifies failures.
Jev Tools & CLIs#community#typescript#coding-agent
Community
y0usaf/pi-jevTypeScript
y0usaf/pi-jevTypeScriptA Hermes skill that calls Jev as a typed tiebreaker: the agent submits a state and a question, and Choice, Score, or Noul answers with probabilities let it act or escalate.
Jev Tools & CLIs#community#python#skills
Community
shantanugoel/ask-jev-skillPython
shantanugoel/ask-jev-skillPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.