October 19, 2026
By Steven Van ·
Existing falcon-2, falcon2 and FALCON identifiers now run Falcon 2.1 automatically, and the model adds IPA and dictionary pronunciation control.
Murf's Falcon 2.1 speech model is now live on Stream Speech and WebSockets in the Murf AI API, as listed in the API documentation. Murf describes it as a model for real-time conversational AI.
What changed in the model
- More natural voice quality across languages and voices.
- Lower time to first spoken word, as benchmarked independently by Coval. The model page cites about 100ms time-to-first-audio.
- Fewer artifacts and hallucinations, with more consistent loudness between responses.
- More languages and voices in the Voice Library.
Pronunciation control
On both Stream Speech and WebSockets, a word can be given an exact pronunciation in two ways. IPA can be written inline in the text, for example Welcome to "/mɜːrf/". Alternatively, a pronunciationDictionary sets how a word is said every time it appears. Murf suggests this for brand names, customer names and medicines.
How to switch
Send model: "falcon-2.1" to use the new model. The earlier identifiers falcon-2, falcon2 and FALCON keep working and now use Falcon 2.1, so existing integrations upgrade without code changes.
Other details from the model page
Falcon 2.1 voices can switch between languages within a single sentence. A volumeOffset setting raises or lowers loudness relative to the default, from -50 to 50, with values outside that range clamped. Pricing is 1 cent per 1000 characters. The concurrency limit is 5 for the US-East region and 2 for all other regions.