[PRESSWIRE] Paris, France – June 10, 2026 — Gladia, the AI audio infrastructure company, today announced the launch of Solaria-3 — its best model to date, ranking #1 among all leading speech-to-text providers on business audio and conversational speech, while delivering the strongest accuracy on real English customer calls of any model tested. Built for the audio that enterprises actually deal with in production — contact center calls, sales recordings, and meeting transcripts — Solaria-3 is the latest generation of Gladia’s proprietary Solaria model family and is available to try for free.

Built for real-world conditions

The majority of enterprise voice workflows operate in conditions that most speech models weren’t designed for. Contact center platforms deal with compressed phone audio, background noise, and overlapping speakers. Sales intelligence tools need to capture every word of a fast, jargon-heavy conversation. Meeting intelligence platforms handle multilingual teams switching between languages mid-call.

According to Gladia’s original research, 48% of enterprise users participate in multilingual meetings frequently or at every meeting — yet non-English transcription consistently scores lowest on user satisfaction. Solaria-3 was designed to close that gap, with consistent accuracy improvements across English, French, German, Spanish, and Italian on both internal production data and public benchmarks.

Better than the competition — #1 across key benchmarks

Solaria-3 was evaluated against all leading speech-to-text providers — including AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics — across a full suite of public benchmarks and Gladia’s own internal dataset of real customer calls, annotated by humans.

Solaria-3 ranks #1 in accuracy, measured by word error rate (WER), on:

• Earnings22 — the industry standard for financial and business speech, with 6.4% WER, the only model under 7%, ahead of AssemblyAI, ElevenLabs, and Deepgram.
• Switchboard — the most challenging conversational telephone benchmark, with 33.9% WER; every other model scores above 42%.
• Noisy audio — 1.4% WER on degraded real-world recordings, beating 4/5 providers tested.

Full benchmark results available at gladia.io/solaria-3.

Public leaderboard scores, however, only tell part of the story. Gladia continuously evaluates speech-to-text models on real customer audio — conversational, accented, and noisy — and the gap between benchmark and production performance can be concerningly wide. 

Models that claim sub-4% word error rate on standard benchmarks regularly score above 15% on a real sales call with background noise and a non-native speaker. Gladia’s internal dataset, drawn from actual customer recordings and annotated by humans, was built precisely to close that gap. On that dataset, Solaria-3 scores 9.6% WER — the strongest result of any model tested on actual customer audio.

“Solaria-3 is a genuine architectural leap, not an incremental update. The accuracy gains on real production audio are the biggest we’ve shipped, and the model is built for the volumes that enterprises actually operate at: high-throughput, mission-critical, and fully compliant with the security standards that regulated industries require,” said Maxime Gaudin, CTO of Gladia.

Solaria-1 and Solaria-3: complementary models

Solaria-1 remains available and recommended for use cases requiring broad multilingual coverage across 100+ languages, formal institutional speech, or clean read-speech transcription. The two models are designed to complement each other: Solaria-3 for European production audio, Solaria-1 for breadth.

Availability

Solaria-3 is available today via Gladia’s API. Documentation is available at docs.gladia.io. Gladia is GDPR compliant, SOC 2 Type 2 certified, HIPAA compliant, and ISO 27001 certified, with EU data residency and zero data retention for paid plans.

About Gladia

Gladia was founded in 2022 by Jean-Louis Queguiner with a mission to help companies turn audio data from calls and meetings into structured, actionable intelligence. Its API delivers production-ready speech-to-text and analytics across more than 100 languages, powering voice agents, meeting assistants, and customer support platforms. Headquartered in Paris (France) and New York (US), Gladia has grown to serve over 300,000 users and 2,000 enterprise customers, including Attention, Circleback, Method Financial, and Recall. More information can be found at www.gladia.io, or on Twitter or LinkedIn.

Media contact

Anna Jelezovskaia 

+33.766.868.657 

ajelezovskaia@gladia.io