Solaria-3: Gladia Leading in Production Audio According to Its Own Metrics

Create a watch on this topic Receive upcoming developments if any, with an AI watch configured from this article.

Create my free watch

The API transcription market is shifting since 2024-2025 towards production audio - noisy meetings, accents, telephony - and Gladia's Solaria-3, published on June 10, 2026, ratifies this shift through a deliberate choice: the model progresses on real audio but retreats by 36% on Multilingual LibriSpeech compared to Solaria-1. The young Parisian startup claims the top position on Earnings22 Cleaned AA with 6.4% WER, according to its own measurements.

This trade-off is assumed: Gladia maintains Solaria-1 in parallel for extended multilingual use cases. The benefit of Solaria-3 is not uniform per language according to Gladia (English −26%, German −3% on its internal audio).

A table that can be read in both directions

The numbers published by Gladia draw a shift in specialization. The model progresses on the audio conditions encountered by a call center - 8 kHz telephony, multi-speaker meetings, non-native accents - and retreats on laboratory conditions where Solaria-1 maintains the advantage. The table below summarizes the measurements published by Gladia on June 10, 2026 (WER = word error rate, error rate on transcribed words).

Benchmark Audio Conditions Solaria-3 WER Reference Source Earnings22 Cleaned AA financial / professional speech 6,4 % AssemblyAI Universal-2: 6,9 % Gladia Switchboard degraded telephone 8 kHz 33,9 % ElevenLabs: 55,2 % Gladia Audio with noise background noise 1,4 % Mistral Voxtral: 1,0 % Gladia Multilingual LibriSpeech studio audio, multilingual 8,0 % Solaria-1: 5,9 % (+36 %) Gladia VoxPopuli Cleaned AA institutional / parliamentary audio 2,9 % Solaria-1: 2,2 % (+32 %) Gladia According to Gladia, Solaria-3 outperforms AssemblyAI Universal-2 on Earnings22 by 0.5 points (6.4 % vs. 6.9 %) - a difference that can be interpreted within the typical noise margin of WER measurements. On Switchboard, the publisher presents its model as the only one in the in-house comparison to score below 35 %. However, the scope is narrowing: Solaria-3 is optimized for five European languages (English, French, German, Spanish, Italian), whereas Solaria-1 is still reported by Gladia to support over 100 languages, including 42 exclusive ones. All these numbers are based on Gladia's internal dataset, which is proprietary and internally annotated—it is not public, making third-party replication impossible at this time.

What This Trade-off Says About the STT Market for a B2B Buyer

Earnings22, Switchboard, and VoxPopuli capture what a B2B buyer encounters daily: earnings calls, 8 kHz phone conversations, and accentuated parliamentary speeches. The sector is redefining itself around this second type since 2024-2025, and Solaria-3 confirms this shift for Gladia, the young Paris-based startup founded in 2022 and funded by a $16 million Series A round in October 2024.

The closest European competitor is Voxtral from Mistral AI, released in July 2024 and iterated until Voxtral Transcribe 2 in early 2026. Gladia chooses not to include Voxtral in its main comparison table, even though Voxtral outperforms it on noisy audio (1.0% vs. 1.4% WER) in the details of its own publication. Regarding compliance, Gladia announces SOC 2 Type II, HIPAA, RGPD, ISO 27001 certifications with EU and US clusters—this is a sovereign argument to consider: it applies to inference and customer data, not to training.

For a decision-maker selecting a transcription provider, the evaluation criterion thus shifts with the market. A use case focused on meetings and call centers (close to Earnings22, Switchboard) calls for a test on Solaria-3; a broad multilingual use case or clean audio (documentary transcription, institutional reading) refers to Solaria-1 or a competitor. The point that will decide Gladia's top claim boils down to one line: the publication, by a third-party evaluator, of WER measures under the same audio conditions - Earnings22, Switchboard, noisy audio - including Voxtral, Whisper, and APIs from major cloud providers absent from the house comparison.