AI voice
We follow voice models, the people building them, and what happens when you put them to work.
Start here
Speech Recognition Models Repeated Numbers That Researchers Had Removed From the Audio
Researchers silenced numbers in public benchmark recordings, but several speech recognition models still returned the exact values. The audit asks whether some models learned the answer keys.
People Preferred Gemini's Voice Agent Even Though Other Models Completed More Tasks
In Artificial Analysis's 2026 Speech Agent Arena, Gemini 3.1 Flash Live led user preference with 74.6% task success, while less-preferred rivals reached 89.4% and 94.7%.
Latest articles
Where the Delay Comes From After You Stop Talking to a Voice Agent
A pause after your last word can come from turn detection, transcription or response generation. Five timestamps show where a voice agent spent the time.
An AI Voice Agent Said It Changed a Customer's Address Even Though the Database Never Updated
In a 278-task test, voice agents completed 26% to 51% of customer-service work, and one agent confirmed an address change without making the required tool call.
A Voice-Cloning Sample Contains Your Words, Delivery, Identity and Room Noise
A voice-cloning reference mixes the speaker's words, identity, delivery and room. Different model designs preserve different parts in the generated voice.
Get Rohkun by email
New tests and explanations when we have something worth sending.