Guardrails

Profanity filtering, content moderation, and PII redaction at the text and audio level. One API, compliant by default.

Get started See the models

Features

Output

Doctor: Hi, Mr. #####. How are you?

Patient: I’m good, Dr. #####. Nice to see you.

Doctor: Nice to see you again. What brings you back?

Patient: Well, my back’s been hurting again.

Doctor: Oh, I see. I’ve seen you a number of times for this, haven’t I?

Patient: Yeah, well, ever since I got hurt on the job 3 years ago, it’s something that just keeps coming back.

Doctor: It’ll be fine for a while, and then I’ll bend down or I’ll move in a weird way, and then Unfortunately, that, that can happen, and, and I do have quite a few patients who get recurring episodes of back pain. Have you been keeping up with the therapy that we had you on before?

Patient: Which, the pills?

Doctor: Actually, I was talking about the physical therapy that we had you doing. The pills are only meant for short-term because they, they don’t actually prevent the back pain from coming back.

Patient: Um, see, yeah, once my back started feeling better, I was happy not to go to the therapist anymore.

Doctor: Why was that?

Patient: Well, it was starting to become kind of a hassle, you know, with my work schedule, and the cost was an issue. But I was able to get back to work, so— and I could use the money.

Doctor: Do you think the physical therapy was helping?

Patient: Yeah, well, it was slow going at first.

Doctor: I see. Physical therapy is a bit slower than medications, but the point is to build up the core muscles in your back and your abdomen. Physical therapy is also less invasive than medications, so that’s why we had you doing the therapy. But you mentioned that cost was getting to be a real issue for you. Can you tell me more about that?

Code

import assemblyai as aai

aai.settings.api_key = "YOUR_API_KEY"

audio_file = "https://assembly.ai/wildfires.mp3"

config = aai.TranscriptionConfig(
    speech_models=["universal-3-5-pro", "universal-2"],
    language_detection=True,
    speaker_labels=True,
)

transcript = aai.Transcriber().transcribe(audio_file, config=config)

if transcript.status == aai.TranscriptStatus.error:
    raise RuntimeError(f"Transcription failed: {transcript.error}")

print(transcript.text)

for utterance in transcript.utterances:
    print(f"Speaker {utterance.speaker}: {utterance.text}")

Features

Every protection, built into your timeline

Ship Voice AI you can trust in production. Every guardrail you need, built into the API.

PII Text Redaction

Identify and remove personally identifiable information — such as social security numbers, dates of birth — from the transcript before it is returned to you.

PII Audio Redaction

Replace personally identifiable information — such as social security numbers, dates of birth — from the audio files with a "beep" for compliance.

Content Moderation

Detect sensitive content in your audio and video files — hate speech, violence, sensitive social issues, alcohol, drugs, and more.

Profanity Filtering

Identify and redact profane or sensitive terms, keeping your speech-to-text output clean and compliant across any application.

Speech Threshold

Transcribe only files with a minimum percentage of spoken audio.

Start & End of Transcript

Limit transcription to sections containing speech.

Comprehensive protection

Safety, compliance, and operational controls

Enterprise guardrails built into every request, so you can ship to production with confidence.

Safety Guardrails

Catch sensitive content, remove personal information from text and audio, and keep outputs production-clean.

Operational Controls

Speech Threshold and transcript start/end controls validate inputs before they hit the model. Transcribe only what matters, skip what doesn’t, and keep quality consistent at scale.

Compliance ready

Built-in PII Redaction and content controls help you meet GDPR, HIPAA, and PCI requirements. The guardrails regulated teams rely on to ship safely.

Guardrails built in, not bolted on

Model

AssemblyAI One API, every guardrail

Deepgram Limited safety surface

AWS Stitch your own pipeline

PII Redaction, text

Built-in

PII Redaction, audio

Built-in

—

Separate pipeline

Content Moderation

Multi-category

—

Toxicity only

Profanity Filtering

All models

Legacy models only

DIY filter

Speech Threshold

Built-in

—

Transcript start/end

Built-in

—

Playground

We’re not playing around, but you can

Put our Voice AI models to the test in our no-code playground.

Try it out

AI Speech-to-Text transcription in 99 languages

From Spanish to Korean, deliver accurate Voice AI in the languages your users speak.

🇪🇸 Spanish 🇵🇹 Portuguese 🇫🇷 French 🇩🇪 German 🇮🇳 Hindi 🇷🇺 Russian 🇳🇱 Dutch 🇯🇵 Japanese 🇮🇹 Italian 🇵🇱 Polish 🇺🇦 Ukrainian 🇮🇩 Indonesian 🇹🇷 Turkish 🇨🇳 Chinese 🇰🇷 Korean

Common questions

: Voice AI Guardrails are AssemblyAI's built-in safety features that keep voice pipelines safe and compliant. They include content moderation, PII redaction, profanity filtering, and operational controls like Speech Threshold — all available through a single API.
: Content Moderation for Voice AI detects sensitive or inappropriate topics in audio/video, including hate speech, violence, alcohol, drugs, and more. Results are returned with per-category confidence scores.
: Yes. AssemblyAI supports GDPR and PCI compliance, offers a DPA, and provides PII redaction for both text and audio output so regulated teams can ship safely.
: Enable Guardrails features via options in your transcription request. Each feature can be toggled independently in the API configuration — no separate pipeline required.
: Yes. Select Guardrails features are available with Realtime Speech-to-Text API for real-time safety filtering.
: PII Audio Redaction replaces sensitive spans of the original audio file with a "beep" so you can safely store or share recordings for compliance reviews.
: Guardrails coverage varies by feature. PII Redaction and Content Moderation support our core Universal model language set; see the docs for per-feature language availability.