Murf AI launches Falcon 2, a sub-100ms real-time voice model for AI agents

Murf AI, an India-based voice-AI company, unveiled Falcon 2, a real-time text-to-speech model built for voice agents that it says delivers its first audio in under 100 milliseconds, with public API access opening Aug. 20.

Latency is the pitch. Conversational voice agents feel broken when replies lag, and Murf says Falcon 2 has held a median time-to-first-audio of about 95 milliseconds over a live 30-day production window. The company says the model spans more than 150 voices across 35-plus languages, costs $0.01 per minute, and can handle up to 10,000 concurrent calls.

The launch pushes Murf into a crowded real-time voice market led by ElevenLabs and OpenAI’s Realtime API. Murf says the independent Artificial Analysis Speech Arena ranks Falcon 2 above ElevenLabs’ Flash v2.5 and Turbo v2.5, OpenAI’s Realtime API and xAI’s text-to-speech model on naturalness — a claim that rests on a third-party leaderboard Murf itself is citing, not a test the company commissioned.

Per its API documentation, Falcon 2 has processed more than 250 million characters since launch, with 99.9% uptime and data residency across 11 geographies — figures aimed at enterprise buyers weighing reliability and compliance alongside voice quality.

Benchmark rankings on naturalness are subjective and shift as rivals update their models, so the standing above ElevenLabs and OpenAI may not hold. The real test comes when the API opens Aug. 20 and developers run Falcon 2 against those incumbents at production scale.