Guardrails API

The Guardrails API adds safety, privacy, and compliance controls to AssemblyAI transcripts and audio. It covers PII redaction in both transcript text and the underlying audio, profanity filtering, and content moderation, applied as add-ons to a transcription request.

General
Developer	AssemblyAI
Type	Safety and compliance API
License	Commercial API
Documentation	assemblyai.com/products/guardrails

Core Features

PII text redaction: removes personally identifiable information from transcript text.
PII audio redaction: redacts personally identifiable information directly in the audio file.
Profanity filtering: filters profane language out of transcript output.
Content moderation: flags or filters unsafe or policy-violating content.

Pricing

Guardrails features are billed as add-ons on top of the base Speech-to-Text rate:

Add-on	Rate
Profanity Filtering	+$0.01/hour
PII Text Redaction	+$0.08/hour
Content Moderation	+$0.15/hour

Source: assemblyai.com/pricing

Use Cases

Regulated industries: healthcare and financial services use cases that require PII redaction before storing or sharing transcripts.
Customer support and contact centers: content moderation and profanity filtering on recorded or live calls.
Compliance workflows: redacting sensitive information from audio archives before retention or sharing.

Ecosystem and Integrations

Guardrails add-ons apply to output from both Speech-to-Text and Universal-Streaming, and can be combined with Speech Understanding add-ons on the same request.

Read the Guardrails docs to add safety and compliance controls to an existing transcription request.

NURDA6FOOT4

SDSADASDDASDSADASD ka dwkec keCJWeofweokMsdcosmks slnl;seNCVL;wenvipuwernfuiwnwoerf SDSADASDDASDSADASD ka dwkec keCJWeofweokMsdcosmks slnl;seNCVL;wenvipuwernfuiwnwoerf SDSADASDDASDSADASD ka dwkec keCJWeofweokMsdcosmks slnl;seNCVL;wenvipuwernfuiwnwoerf SDSADASDDASDSADASD ka dwkec keCJWeofweokMsdcosmks slnl;seNCVL;wenvipuwernfuiwnwoerf SDSADASDDASDSADASD ka dwkec keCJWeofweokMsdcosmks slnl;seNCVL;wenvipuwernfuiwnwoerf SDSADASDDASDSADASD ka dwkec keCJWeofweokMsdcosmks slnl;seNCVL;wenvipuwernfuiwnwoerf SDSADASDDASDSADASD ka dwkec keCJWeofweokMsdcosmks slnl;seNCVL;wenvipuwernfuiwnwoerf SDSADASDDASDSADASD ka dwkec keCJWeofweokMsdcosmks slnl;seNCVL;wenvipuwernfuiwnwoerf SDSADASDDASDSADASD ka dwkec keCJWeofweokMsdcosmks slnl;seNCVL;wenvipuwernfuiwnwoerf SDSADASDDASDSADASD ka dwkec keCJWeofweokMsdcosmks slnl;seNCVL;wenvipuwernfuiwnwoerf SDSADASDDASDSADASD ka dwkec keCJWeofweokMsdcosmks slnl;seNCVL;wenvipuwernfuiwnwoerf

WebInfo: Watch an AI Agent Find the Answer for You

WebInfo is an AI agent that does web research for you. Ask a question in plain English and it searches the web, opens the real websites in its own browser, reads them, pulls out the exact fact and highlights the sentence it came from on the real page. You watch it work live, with a breathing gaming-style cursor moving over the sites and a step-by-step agent log. Every answer comes with its source link and a word-for-word quote, and if the fact isn’t there it says so instead of guessing. How it works: FastAPI streams each step to the browser. Tavily finds the pages and their text in one search, headless Chromium (Playwright) opens them, Gemini Flash-Lite extracts the answer as JSON with verbatim quotes, and Postgres caches answers and tracks API usage. It runs in Docker on a free Google Cloud e2-micro VM, deployed from GitHub. What we optimized: answers dropped from 34 s to as fast as 5 s, because Gemini reads the search text while the browser opens the page in parallel. Results are ranked by relevance to the question, and model numbers and years must match. Each new question costs 1 search credit and 1–2 Gemini calls, and repeat questions cost zero. Hard caps keep it on free tiers, so it runs for $0. It blocks private addresses, respects robots.txt and treats page text as untrusted. Example: asked for the AMD Developer Hackathon ACT III prize pool on lablab.ai, it opened lablab.ai, answered $17,000+ and highlighted the proof on the page.

HandSpeak

SignBridge is a real-time AI-powered sign language interpretation system designed to make communication more accessible between sign language users and people who do not understand sign language. Using a camera, the system detects hand and body movements, analyzes gesture sequences, and converts recognized signs into natural-language text and speech. The project will leverage AMD GPU acceleration to improve real-time computer vision and AI inference performance, allowing the system to process video with low latency. Beyond individual signs, SignBridge aims to recognize sequences of gestures and reconstruct meaningful phrases, making communication more natural. The project combines accessibility, computer vision, AI, and high-performance AMD computing to create a practical assistive technology solution.

The Witness: payer call memory

A medical biller spends 25 minutes and $13.80 on one phone call asking a payer why a claim was denied (CAQH Index, 2024 edition, 2023 data year, the latest edition with per-transaction figures; the 2025 edition, 2024 data, reports $8.03 as the average across all manual medical transaction types, a different measure that is not a claim-status figure). The answer is spoken, unrecorded on the provider's side, and gone the moment they hang up. Weeks later a different representative gives a different answer, and nothing notices. About 19 percent of in-network claims are denied and under 1 percent of those denials are appealed (KFF, March 2026). The gap is not merit. It is evidence: appealing means reconstructing what the payer said, and nobody can. We don't predict what the payer will pay. We record what the payer said. The Witness rides along on the call. Every statement the payer representative makes becomes a timestamped, quotable row in an append-only, hash-chained ledger keyed to the claim. Five kinds of contradiction are detected against earlier calls on the same claim, in real time, and a refusal (the payer names a category but withholds the value) is recorded as a row of its own: the absence is the evidence. In Mode A the Witness conducts the call itself. It discloses that it is an AI assistant and that the call is recorded, challenges the contradiction on the recorded line ("on July eighth you told us no prior authorization; today you're saying timely filing, which one is on the claim?"), reads reference numbers back, and refuses to hang up until every required field is captured or refused. The appeal packet quotes the transcript with rep, badge, reference number, date and offset, and carries a hash anyone can recompute.

Reprise: A Podcast That Remembers You

Journals are lonely and easy to abandon. Podcasts need an audience and hours of audio editing. Reprise sits between the two: a private podcast you record by talking with an AI host that actually remembers you. Each take is a live, full-duplex conversation on AssemblyAI's Voice Agent API, with natural turn-taking and barge-in. The guest picks the host's voice. The host opens by recalling something you said in an earlier episode, quoted word for word from that episode's stored transcript. It never invents a memory. Names, places and callbacks come only from stored data. When you end the take, Reprise records both sides as separate stems and builds a draft. Gemini suggests cuts, a cold open, a title and show notes. You review the suggestions in the transcript and accept or revert each cut, so the model proposes and you decide. A pinned static FFmpeg renders the cuts. Speech understanding then runs on the finished file, not the raw take, so every timestamp matches what you hear. Universal-3.5 Pro with entity detection and key phrases produces chapters and the people and topics the host brings back next time. Episodes are private by default. Publishing gives one share link. Every paid call reserves budget first, and every session ends explicitly with a token cap. Reprise is live at reprise.nryn.dev, built in Go and Svelte 5.

Task Voice Agent

Voice Task Assistant is a practical productivity voice agent built entirely on AssemblyAI’s Voice Agent API. Users speak naturally to capture tasks, review their list, and mark items complete — without touching a keyboard. The entire voice stack (speech-to-text with Universal-3 Pro, LLM reasoning, and text-to-speech) runs through a single AssemblyAI connection with natural turn-taking and barge-in support. Replies are kept short and direct so the interaction feels fast and useful. This project demonstrates a clean, production-ready use of the Voice Agent API for everyday personal productivity. It requires no extra services for the core experience and can be extended later with HTTP tools for persistent storage.