Murf AI Says Falcon 2 Voice Model Beats OpenAI, ElevenLabs on Naturalness

Murf AI has launched Falcon 2, a proprietary voice foundation model that the company says ranks ahead of OpenAI Realtime API and ElevenLabs’ Flash and Turbo models for naturalness on the Artificial Analysis Speech Arena, while delivering sub-100 millisecond time-to-first-audio at a lower cost.

Falcon 2 is priced at $0.01 per minute of generated audio, compared to $0.05 for both ElevenLabs Turbo and Flash. Murf AI said the model recorded higher naturalness scores than ElevenLabs Turbo v2.5, Flash v2.5 and OpenAI Realtime on the independent benchmark.

The model aims to address a key trade-off in real-time voice AI: balancing naturalness, latency, and cost. Murf said Falcon 2 uses a lightweight proprietary architecture focused on inference efficiency rather than primarily increasing model size.

The Bengaluru-based company is also positioning Falcon 2 around the requirements of the Indian market, including multilingual and code-mixed conversations, diverse accents, and the infrastructure costs associated with running voice AI at scale. The model supports 150+ voices across 35+ languages, including extensive coverage of Indian languages.

“Falcon 2 represents a different proposition: a voice foundation model developed by our India-based teams and competing directly with established global models on independent benchmarks,” Sneha Roy, Co-founder and COO, Murf AI, said in a statement. She added that the model was engineered around the economics and deployment requirements of enterprise production.

Beyond speech generation, Falcon 2 includes a deeper semantic layer designed to interpret sentence meaning and adjust pacing, emphasis and pronunciation based on context. Its conversational training includes patterns from customer calls, support interactions, and booking conversations.

The model also improves text normalisation for enterprise-specific information such as numbers, currencies, dates, addresses, pincodes, and vehicle numbers.

Murf AI said Falcon 2 can support up to 10,000 concurrent calls with stable latency, offers data residency across 11 geographies and supports on-premise deployment. Enterprises can also use instant voice cloning to create a custom brand voice using 10 seconds of recorded audio.

At the announced pricing, one million minutes of generated speech would cost about $10,000 on Falcon 2, compared with $50,000 on ElevenLabs Turbo and Flash. Murf said the economics are intended to make high-volume, real-time voice deployments more viable for enterprises.

Latency is another area Murf is highlighting. In tests conducted using a third-party geo-distributed API relay, Falcon 2 ranked first across most regions tested, with time-to-first-audio below 100 milliseconds.

The launch also comes as enterprises increasingly experiment with voice agents for customer support, commerce, bookings, and other phone-based interactions. In India, Murf has identified multilingual speech, code-switching, latency, cost per minute and local infrastructure as particularly important constraints for scaling such systems.

“The challenge was never simply reducing latency or improving speech quality independently. It was engineering a system where both improve together without increasing the computational cost of deployment,” said Ankur Edkie, Co-founder and CEO, Murf AI.