4 articles
NVIDIA’s Nemotron 3.5 Transcribe and Granary v2 point to a speech stack where transcription, adaptation, and multilingual data creation are joined.
Hugging Face’s Open TTS Leaderboard adds scalable evaluation for multilingual speech and voice cloning, exposing the tradeoff between naturalness, identity, and consent.
YODAS v3’s million-hour speech dataset points to the scale open voice models need and the provenance questions that scale makes impossible to ignore.
Hugging Face’s FFASR leaderboard pushes speech recognition toward the conditions that actually matter: noise, distance, latency, and real deployment hardware.