Voice-cloning models are evaluated and marketed with their substantially high speaker similarity scores, but would this impression of fidelity be challenged if we listened to the generated audios and saw for ourselves?

Open-weight TTS models can look remarkably close to the best closed-weight models on benchmarks and demos. But are they reliable enough to actually replace them? We stress-test Qwen3-TTS and VoxCPM2 in real-world scenarios, uncover reliability failures that standard evaluations miss, and develop an inference harness that eliminates the reliability gap. 🧵

This is what context does to your speech-to-text system! Our new paper studies the impact of contextual information on the accuracy of leading open-source and proprietary systems.

Voice is becoming the most powerful interface for AI — but real-world issues can still break today's agents. Join founders from @ai_coustics, @covaldev, @argmax, and @livekit for The Audio Layer, a deep dive into making voice systems reliable in the real world. Dec 10 • 6–9 PM • Frontier Tower SF RSVP: luma.com/the-audio-layer

Introducing Custom Vocabulary in Argmax Pro SDK Customize speech-to-text with contextual keywords at runtime to surpass the accuracy of generic frontier speech models! You can try it on @superwhisper and @argmax Playground today. More demos and details in comments.

Going to @icmlconf and want to learn about the frontiers of on-device AI? Catch my talk on July 18 after @DAlistarh and before @songhan_mit! I will discuss the secret sauce and benchmarks behind how @argmax built WhisperKit to outmatch several top cloud providers for real-time transcription without Nvidia GPUs.

Been experimenting with gpt-oss:20b locally on my M1 Max and pit it against Gemini 2.5 Pro on the same type of deep research question I often explore: 'How does @macwhisperapp’s speaker detection actually work?' Not surprisingly, the tiny gpt-oss punches way above its weight. ↓

the tide is shifting to local models.

It's been fun collaborating with @superwhisper, a blazing fast AI speech-to-text app, with support for local and cloud models. By putting @vercel CDN in front of their model API, they're seeing up to 350ms gains in some geos 🤯 To be clear, this is 350ms+ faster by *just* putting a proxy in front backed by our global network. They haven't even replicated their db / compute yet! It's free real estate. When we talk about 'fast data/origin transfer', this is what we mean. We partnered with @awscloud to back vercel with a fully redundant 400 GbE fiber network backbone. No noisy neighbors, top notch security, no last-mile unexpected rerouting, no free-tier throttling, no enterprise-only POPs or TCP optimizations, no Telekom peering issues…

Introducing the world's best (and open) speech recognition models!