Technology

How InterLive AI Interpretation Works

Whisper speech recognition → GPT translation → WebSocket streaming. Three steps completed in under one second.

What is real-time AI interpretation technology?

Real-time AI interpretation connects speech recognition (ASR), neural machine translation (NMT), and text-to-speech (TTS) in a pipeline that converts spoken words into another language instantly. InterLive combines OpenAI Whisper (ASR), GPT-4o (translation), and WebSocket streaming to deliver real-time interpretation with an average latency under one second.

Interpretation processing pipeline

Audio capture

The browser Web Audio API captures the microphone stream in chunks.

Speech recognition

The Whisper API receives audio chunks and converts them to text, including automatic language detection.

AI translation

GPT-4o translates into the target language while maintaining context from previous utterances.

Real-time delivery

WebSocket delivers translation results to the client instantly for simultaneous subtitle and TTS output.

Technology stack

Speech recognition (ASR)

OpenAI Whisper

Converts speech in 15 languages to text with high accuracy, robust even in noisy environments.

Translation engine

GPT-4o / GPT-4o-mini

Context-aware high-quality translation. Balance speed and quality with the inference level setting.

Real-time streaming

WebSocket

Results are sent instantly per utterance. Low-latency subtitle output without HTTP polling.

Speech synthesis (TTS)

OpenAI TTS

Translated text is output as natural AI speech. Choose from voices in 5 languages.

Meeting bot

Cloud Run + Playwright

A serverless bot that automatically joins Zoom, Teams, Meet and Webex. Works from just a URL.

File processing

PDF.js / mammoth / python-docx

Extracts text and structure from PDF, DOCX, and HWPX files for layout-preserving translation.

Technology FAQ

Which AI models does InterLive use?

OpenAI Whisper for speech recognition, and GPT-4o (precise) or GPT-4o-mini (fast/standard) for translation. The model is selected automatically based on the inference level setting.

How much latency does interpretation have?

Under one second on average. This varies based on network conditions, utterance length, and inference level.

How is voice data handled?

Audio streams are not stored on the server after processing. Only the translated text record is saved to the user account.

How does the Zoom/Teams bot work?

When you enter a meeting URL, a bot is automatically created in a Cloud Run serverless environment and joins the meeting. The bot receives audio and processes real-time interpretation.

Start now

Try free See features