Gladia

If you’re assessing STT performance and your model looks accurate, but transcripts are off, the issue is often your ground truth. More specifically, how it’s defined and normalized. Without proper normalization, a model can score well on WER and still produce outputs that aren’t usable. At Gladia, normalization is built directly into the transcription pipeline, so outputs are aligned with real-world usage from the start. Learn more about this topic in the comments.

Exciting news on the latest performance benchmarks of our model Solaria 1! Across Switchboard, Gladia shows on average ~29% less WER than other providers. Why this matters: If performance only looks good on clean speech, it won’t hold up in conversation. Switchboard is among the most challenging datasets out there - noisy background, overlapping speech. True conversational audio. Which goes to show that our models are designed for real, messy, customer audio, not academic datasets. You can test it on your own audio in our playground. (Link to our open-source benchmarks in the comments)

👉 You may be familiar with common ASR datasets such as Common Voice or Earnings. But do you know the kind of audio they include? Datasets like Common Voice are a good starting point, but often short, scripted, and relatively controlled. Look beyond that, and we get very different datasets: Switchboard→ phone calls with interruptions and background noise VoxPopuli→ long and formal parliamentary speeches with varied pacing Earnings22 → multi-speaker, conference calls with financial jargon MLD (Multilingual LibriSpeech) → clean audiobooks, multilingual Knowing these differences can help you make an informed decision on which dataset best reflects the real-world performance you can expect from a model. And using your own audio, with clean ground truth and normalization, is ultimately the best WER test. Learn more about our results across datasets and top STT providers in the comments.

🔊Reminder: “Bring your own audio” is happening tomorrow. Over the past few days, developers have been sending us the recordings that usually break speech-to-text systems. Some of them are… brutal. Noisy calls. People talking over each other. Mics that sound like they came from 2007. Tomorrow we’ll run those files through multiple STT APIs live and see what actually happens. If you submitted audio → this is where you find out how your file performs. If not, it should still be a fun one to watch. (Link in the comments)

This is crazy. We just opened up the leaderboard of our blind STT comparison tool…and Gladia is now # 1. Almost 200 people submitted their audio and voted on the best of the two transcripts, based purely on quality. And the majority chose us - the smallest team in the batch, the underdog. I'll admit, I was scared to have this tool go live. Being truly transparent in the industry means holding ourselves to the same scrutiny as everyone else and I stand by that decision. But if living in the US has taught me anything, it's to celebrate our wins more. So here I am trying to lean into that (my French roots are fighting me on this, but I'm trying). If you want to keep challenging the leaderboard, I'll link the tool in comments.

The ASR leaderboard is live! 🇫🇷 ✊ Last week, we released a blind tool for you to compare STT models without bias. Now, it’s time we lay our cards on the table. See how every provider performs in the aggregate leaderboard—ELO scores, rankings, and all the metrics that matter. Is it what you expected? For a proper evaluation, we still recommend reproducing the results on your own audio using our open-source methodology. (link in the comments)

We’re heading to AWS Summit Paris 2026! 🚀 If you’re building anything with voice AI, come say hi! 📍 Palais des Congrès, 1st Floor - Startup Zone 📆 April 1, 2026 We’d love to hear about the challenges you’re facing and see where we can help. Lisa Surraco, Ludovic Teyssier, Constant Dacbert, Tarun Chellaboyina, Merwan Fagoul #AWSSummit #AWSSummitParis #VoiceAI

Our CTO, Maxime Gaudin, joined Davit Baghdasaryan, CEO of Krisp, for a conversation on what it actually takes to scale STT in production. They talked about the things you don’t always see — balancing latency, cost, and quality, what actually breaks at scale, and why running models in real world is very different from training them. Thanks again, Davit and Krips, for having us! Krisp is doing some seriously impressive work in voice AI 👏 Go check out their newsletter so you don’t miss it: https://lnkd.in/dVVx-vGQ

How do you compare speech models without brand bias? Proper testing is the way to go. But sometimes, it can also help to simply see the difference between two transcripts with your own eyes. So we built a small experiment 👀 One input audio clip, two output transcripts. You pick the winner. A blind test for speech-to-text, featuring Gladia, Deepgram, AssemblyAI, ElevenLabs, Mistral AI, and Speechmatics. Test it with any audio, speaking in different languages. See how far it can go. (Link in the comments)