Gladia

Our new model, Solaria-3, ranks #1 on real production audio — ahead of ElevenLabs, Deepgram, AssemblyAI, and Mistral. Not on curated benchmarks. On actual customer calls. We stopped trusting leaderboards and built our own dataset of real-life recordings, annotated by humans. On that dataset, Solaria-3 hits 9.6% WER. Compared to our own Solaria-1? 26% fewer mistakes. The standard benchmarks measure the best case. Your data is the real case: noisy, fast-paced, multilingual. We optimized for that, and we’re just getting started.

✨ Anthropic just dropped Claude Fable 5, and you can already use it inside Audio-to-LLM! Transcription, diarization, and LLM analysis all returned together in a single webhook. Our model field accepts any model available on OpenRouter. That's 400+ models, including every new frontier release. You pick the model. We handle the audio. Go play with it and tell us what you've built. #fable5 #claudefable5

We’re in the Sifted 100: France & Benelux. 🚀 That means we've been named one of the fastest growing startups across France, Belgium, the Netherlands, and Luxembourg — companies showing real traction and building businesses that matter. The ranking looks beyond funding rounds to spotlight the companies growing fastest in the region: revenue, headcount, user growth, and the fundamentals that actually count. Proud to be recognised alongside some of the most ambitious companies in the region. #Sifted100 #SiftedLeaderboards

🔥 Solaria-3 is out — the best STT model for business audio in core European languages! Most speech models are tuned for clean read-speech: none of that survives in real-life scenarios. Solaria-3 is built for noisy, fast-paced environments and beats every leading provider on business speech. • Best on real English audio and 26% better than Solaria-1 • #1 on Switchboard, the hardest conversational benchmark, staying ahead of ElevenLabs, Deepgram, and AssemblyAI • Superior performance across 5 European languages: EN, FR, DE, ES, IT Best part – Solaria-3 is free to use until June 21! Use the code TRY-SOLARIA-3 at checkout, redeemable once per account for async transcription. Now, go and play with it!

Gladia ranks #1 on a community-led ASR leaderboard! In case you missed it, we recently released a blind ASR test tool: submit your audio, get two transcripts, vote for the better one, then see which vendor produced each. The methodology behind it is open and fully reproducible.   Since then, 660 community votes, with Gladia leading the ELO of 1611 and a win rate of 50.5%. Our model stayed on top most of the time and never dropped out of the top 3, all from anonymous votes — a truly proud moment for us. And we won't stop here.   Share it with friends and keep the votes running ⚔️

A few weeks ago, we dropped the first market map of the meeting assistant space. People are still digging through it, and for good reason. 100+ tools, four moats, a year of research, and a clear picture of who's building what and where the gaps are. If you haven't found your spot on it yet, now's the time. Link in the comments. 👇

We're kicking off Gladia’s Unboxed series for voice apps, and Granola is the first one we put in the ring. Granola sits in an interesting position in the meeting assistant market – it strips away the meeting bot entirely and bets on a lightweight, local-first experience that stays out of your way during the call. In this review, we cover: → Setup and the first-run experience → How botless audio capture actually works → Real-time transcription, multilingual support, and code-switching → The notepad UX — jotting notes mid-call alongside the live transcript → Post-meeting AI: prompting, summaries, and action items Granola is only one of 65+ tools we mapped across 4 moats of defensibility for our Meeting Assistant Market Map. Full list in the comments.

Capture is where the AI meeting assistant market splits. Some products send a bot into the call. Some sit invisibly on the desktop. Some record from a device on your lapel. Amanda Zhu, Founder & CEO at Recall.ai, framed it sharply when we spoke to her: enterprises that care about user consent and adoption opt for the bot. PLG productivity apps tend to opt for invisible recording, where consent and compliance aren't the priority. And our new research backs that up — broader capture creates more value. Governance is what makes that capture enterprise-safe. Meanwhile, our 2,000+ user survey threw a curveball: only 1 in 6 users say botless capture is a top priority. An entire generation of invisible tools is being built on an assumption users haven't fully validated yet. The vendors who win will be the ones who understand exactly which buyer they're capturing for and what that buyer is willing to trade. Read a full interview with Amanda in the meeting assistants report, link in comments.

Today’s we’re open-sourcing our Lead ML Speech Engineer Bruno HAYS's original research aimed at achieving high-accuracy WER on lightweight real-time multilingual ASR, with code-switching. The system routes between small monolingual streaming models (Zipformer), using Silero VAD to detect speech boundaries and a SpeechBrain model for language identification, all while keeping language switches accurate, transparent, and CPU-friendly. The result — higher accuracy on WER performance when compared against Mistral, ElevenLabs, Deepgram, and AssemblyAI. Explore the code + benchmarks ⬇️

We surveyed 2,000+ active meeting assistant users across the 6 countries, the kind of research that usually sits behind a paywall. Here's what the data says: • Distribution is beating product: only 1 in 7 users have heard of Granola or Fathom, while @Microsoft Teams is used by 50%+ mostly because it came bundled. • This isn't a PLG category yet: 1 in 3 users chose their own assistant. The rest came via employer, IT, or a bundle. • Privacy outranks integrations: data privacy is the #3 user priority, ahead of speed and integrations. • Nobody asked for botless: Only 1 in 6 cite it as a top priority, despite the industry betting on it. Most telling of all, half the market is up for grabs, with 59% saying they'll switch tools within 12 months. We mapped 100+ tools across the four moats in the full report to help our community see where their next opportunity sits. Link in the comments.