Speech-to-text that holds up on real phone calls

For real-world contact center and business audio, our most accurate model is Solaria-3. On real customer recordings in English and core European languages, it ranks first against AssemblyAI, ElevenLabs Scribe v2, Deepgram Nova-3, Mistral Voxtral, and Speechmatics: 6.4% WER on Earnings22 financial calls, the only model under 7%, and 33.9% on the conversational Switchboard set, the only model under 35%. That's a 26% WER improvement over Solaria-1 on real English audio. Solaria-3 is async, which is the workflow most contact center QA and reporting runs on. See the full benchmark.

Yes, and it's where most generic STT APIs break down. We support 100+ languages with Solaria-1, including languages such as Tagalog, Bengali, Punjabi, Tamil, Urdu, Persian, and Marathi, which matters for BPO operations across Southeast Asia and South Asia. We also handle true mid-conversation code-switching, so when a caller and agent switch languages the transcript stays intact rather than degrading. For European business audio, Solaria-3 is our most accurate model.

Transcription sets the ceiling for everything downstream. Every QA score, coaching note, and CRM entry is only as reliable as the words captured in the first layer, so a wrong name or a missed disclosure silently corrupts the report a supervisor builds on top of it. Accurate async transcription lets QA teams score every interaction instead of sampling a small share by hand. Selectra's QA team now validates AI findings rather than manually reviewing calls, and Aircall cut transcription time by 95% while processing more than 1M calls per week through our API.

Yes. Gladia is EU-sovereign by design — a French company running EU and US workloads on separate infrastructure. We're GDPR and HIPAA compliant, SOC 2 Type II and ISO 27001 certified, and offer zero data retention for both async and real-time processing, with a published sub-processor list. For public sector, banking, or insurance deployments where EU hosting is a hard requirement, this is confirmed, not a roadmap item.

Not reliably. Overall WER can look strong while still missing what matters most for contact center workflows: phone numbers, account IDs, and names — the structured data downstream systems depend on. We benchmark on Earnings22, which uses dense financial and named-entity vocabulary as the closest public proxy for that kind of structured-entity accuracy, where Solaria-3 leads at 6.4% WER, the only model under 7%.