Indian startup Murf AI aims to compete with OpenAI in crowded voice arena

A Bangalore-based startup has entered the crowded race to build artificial voices that sound human, with an AI model that garners high scores versus systems from OpenAI and ElevenLabs Inc. 

Murf AI's text-to-speech Falcon 2, publicly available from Aug. 20, ranked higher than some platforms from better-funded players such as OpenAI's Realtime API in benchmarks on Artificial Analysis, the independent platform that tracks AI performance.

  Like other voice foundation models, it aims to provide big savings for users like call centers, banks and airlines, which hope to employ the technology to handle large volumes of customer communications. Falcon 2 is priced at $0.01 per generated minute - a fifth of the price of larger competitors such as ElevenLabs - making it attractive for conversational AI and real-time voice agents.  It's also faster than many competitors, responding in under 100 milliseconds.

 

  Murf AI was founded in 2020 by the engineer trio of Sneha Roy, Ankur Edkie and Divyanshu Pandey, alumni of the premier Indian Institute of Technology Kharagpur. The Matrix Partners and Elevation Capital-backed startup serves over 10 million users in 300 companies including Honeywell International Inc., Cisco Systems Inc., Pfizer Inc., VMware, Nestle SA and Air France-KLM. Falcon 2 was in limited release to select clients ahead of the broader rollout.

  The startup's India-based engineering team developed the platform to take on a central challenge in real-time voice AI: balancing natural-sounding speech with faster speeds and lower prices. It supports over 150 voices in more than 35 languages, the company said.

  "For years, the most important AI models have largely emerged from the US," Roy, Murf AI's chief operating officer, said in a statement. Falcon 2 is "competing directly with established global models on independent benchmarks, while being engineered around the economics and deployment requirements of enterprise production."

  The price difference becomes noticeable at scale. For every million minutes of generated speech, Falcon 2 costs $10,000, compared with $50,000 for ElevenLabs Turbo and Flash.

  Falcon 2's training included real-world interactions such as live customer calls and booking flows rather than just scripted narration. It's built to interpret the meaning of a sentence while adjusting pace, emphasis and pronunciation around the context.

"We work with real voice actors who consent to having their voices used, with transparent contracts, ongoing royalties and data protection," Roy said in an interview via email.

  The startup says its model improves voice delivery on details that confound many AI systems, such as currencies, dates, addresses and license plate numbers. It can handle as many as 10,000 calls concurrently, Murf AI said. It will provide cloud support for data to reside in 11 geographies, and also offers deployment within in-house clouds for greater security.