Murf’s Falcon 2 Puts a One-Cent Bet on Real-Time Voice

Murf AI’s Falcon 2 arrives with a proposition that is easy to price but harder to validate as a whole: the company lists real-time voice generation at $0.01 per minute and says audio begins in less than 100 milliseconds. SMEStreet also reports a favorable naturalness result against prominent rivals, though that benchmark and Murf’s latency testing measure different parts of a live voice system.

Falcon 2 is Murf’s proprietary voice foundation model. SMEStreet lists its rate at one-fifth of the $0.05 per generated minute listed for ElevenLabs Turbo and Flash. At that published rate, one million minutes of Falcon 2 speech would cost $10,000, versus $50,000 for the two ElevenLabs offerings cited in the comparison.

That gap is the central commercial claim. Voice systems that handle customer support, booking flows or other conversations buy generated speech by usage, so a per-minute difference grows with call volume. Murf’s promise of sub-100-millisecond time-to-first-audio addresses a separate requirement: how long a user waits before hearing the system begin to answer.

SMEStreet says Falcon 2 ranked ahead of ElevenLabs Turbo v2.5, ElevenLabs Flash v2.5 and OpenAI Realtime for naturalness on Artificial Analysis’s Speech Arena. That result speaks to perceived speech quality, rather than the speed at which a session starts or the rate on a vendor’s price sheet.

Murf separately says latency tests using a third-party geo-distributed API relay placed Falcon 2 first across most regions it tested. The company describes a lightweight architecture designed for inference efficiency, or lower-compute model operation, alongside a semantic layer intended to adjust pacing, emphasis and pronunciation to a sentence’s meaning.

Murf says Falcon 2 supports more than 150 voices in more than 35 languages. Its training includes customer calls, support flows and booking conversations, while its text normalization is designed for numbers, currencies, dates, addresses, pincodes and vehicle numbers—inputs that commonly appear in customer-facing exchanges.

The operating claims buyers can evaluate

• Murf says Falcon 2 can handle up to 10,000 concurrent calls with stable latency.
• Murf says customers can use data residency across 11 geographies or deploy the model on premises.
• Murf says enterprises can create a custom voice from ten seconds of recording through instant voice cloning.

The launch gives voice buyers a compact set of comparison points: listed generation cost, initial-audio latency, naturalness, language range and deployment control. None substitutes for the others. The reported Speech Arena position is a naturalness comparison, while Murf’s regional latency result and its concurrency, residency and on-premise capabilities are company claims that matter most when tested against a buyer’s own traffic, geography and operating requirements.