Solaria-3

Solaria-3 is the speech-to-text model of the Paris company Gladia, released on 10 June 2026 alongside the older Solaria-1, which remains on sale. It is deliberately narrow: instead of chasing language coverage, it is optimised for five European languages — English, French, German, Spanish and Italian — and for the kind of recordings contact centres and sales teams produce, where people interrupt each other and the line is poor. On the manufacturer's own dataset of annotated customer calls in English it records a 9.6 percent word error rate, ahead of the models it tested from ElevenLabs, AssemblyAI, Deepgram and Mistral, and 26 percent better than Solaria-1. On Switchboard, a public set of telephone conversations, it is the only model tested below 35 percent. On noisy audio it comes second to Mistral's Voxtral. The unusual part is what the company publishes next to those wins. Solaria-3 is worse than Solaria-1 on two benchmarks made of clean, formal speech: 32 percent worse on VoxPopuli, recordings of parliamentary debate, and 36 percent worse on Multilingual LibriSpeech, read audiobooks. Gladia prints both numbers on the model page and tells customers whose audio looks like that to stay on Solaria-1. Vendors rarely do this, and it is the clearest signal of what the model was traded for. Two limits matter for anyone choosing it. Solaria-3 transcribes recordings only — live streaming remains on Solaria-1, which returns partial results in under 103 milliseconds. And its code switching is described by the manufacturer as limited, where Solaria-1 supports it across more than 100 languages, 42 of which Gladia says no competitor covers. Both models can be deployed on the customer's own servers. All comparative figures above come from Gladia's own benchmark campaign, run on its own production audio alongside public test sets, and should be read as the manufacturer's measurement of its rivals. Soniox was excluded from the comparison.

▸News

The same hour of audio costs 10 cents or 61 — and no benchmark covers all the sellers

8/29/2026 · Speechmatics, Soniox, Gladia

▸Videos

No videos yet.

▸Reviews