The First Promptable Speech-to-Text Model Tested Live

👉 Sign up for LiveKit Cloud: https://cloud.livekit.io/signup?utm_source=youtube&utm_medium=video&utm_campaign=devrel&utm_content=x2L8FXvJAS4 AssemblyAI's Universal-3 Pro is the first speech-to-text model you can prompt like an LLM. That means you can control how it handles domain vocabulary, multilingual audio, verbatim capture, and punctuation rules — all without retraining or fine-tuning. In this video, we run four real-world streaming STT tests using a LiveKit Agent with LiveKit Inference. Every session is logged in LiveKit Cloud's observability dashboard so you can see exactly what each model transcribed and where accuracy breaks down. What we test: - Entity accuracy (credit card numbers, emails, phone numbers) with no configuration - Multilingual code-switching between English and Spanish using `language_detection` - Domain-specific vocabulary using `keyterms_prompt` (names, product terms, technical jargon) - Verbatim disfluency capture for legal, medical, and coaching use cases Swapping between Universal Streaming and Universal-3 Pro is a single line of code through LiveKit Inference, so the only variable is the model itself. AssemblyAI reports up to 45% accuracy improvement with targeted keyterms prompting. We show that directly in the transcript output. Note: `prompt` and `keyterms_prompt` cannot be used at the same time. It's one or the other depending on your use case. Universal-3 Pro is available now through LiveKit Inference with just your LiveKit API key. 📚 Resources 📚 AssemblyAI docs: https://docs.livekit.io/agents/models/stt/assemblyai/?utm_source=youtube&utm_medium=video&utm_campaign=devrel&utm_content=x2L8FXvJAS4 Agent docs: https://docs.livekit.io/agents/?utm_source=youtube&utm_medium=video&utm_campaign=devrel&utm_content=x2L8FXvJAS4 🤝 Join the Community Slack: https://livekit.io/join-slack?utm_source=youtube&utm_medium=video&utm_campaign=devrel&utm_content=x2L8FXvJAS4 #livekit #ai #voiceai