One week after OpenAI launched GPT-Realtime-Translate, Silicon Gardens portfolio company VoiceFrom ran it through a public head-to-head benchmark. VoiceFrom Pro won 6 of 8 language pairs.
On May 8, OpenAI added GPT-Realtime-Translate to its Realtime API. The first model built for live speech translation. Within a week, Silicon Gardens portfolio company
VoiceFrom ran it through a public head-to-head benchmark across eight language pairs. VoiceFrom Pro won six of them.
.webp){kind=spacer}
Tested against OpenAI, Google Meet, LiveVoice, and Palabra using GEMBA-MQM v2 , VoiceFrom delivered lower error magnitude across every major language pair. OpenAI's model is faster at the median. VoiceFrom is more accurate, including on real-world edge cases.
OpenAI's launch tells the industry that live translation is a real category. The benchmark tells the industry which system you can put on a stage today — VoiceFrom team
→ Full benchmark & audio: VoiceFrom Blog post
→ Open-source eval harness: GitHub
Big congrats to the VoiceFrom Team!
Connect with us