Bland AI Launches Speech v3, Topping Audio Realism Blind Benchmark
EXM7777 · x · 2026-08-05
Bland AI has introduced Bland Speech v3, a new text-to-speech model described by the company as the first "Human Speech Engine."
Trained on over 100 million real human conversations, the model ranks first on the third-party Audio Realism Bench with an Elo score of 1237. It outperforms major competitors including Grok TTS, GPT Realtime 2, Gemini 2.5 Pro TTS, and ElevenLabs. The company also showcased a use case where they used just 5 seconds of old audio footage to help a 49-year-old father who lost his voice to a stroke speak again.
Related event: Bland AI Launches Speech v3, Topping Audio Blind Tests(4 posts)→
More from Multimodal
- Qwen Image 2.1 vs Krea 2 Turbo: stress tests with kilogram-based physique prompts and dense crowd scenes — Crazy-Repeat-2006 · 2026-09-20
- Qwen-Image-2.1 INT8 ConvRot quantized version released on Hugging Face — Easy-Bike1524 · 2026-09-20
- Qwen-Image-2.1 hands-on: new open-source editing standard with 10 reference images and native RGBA output — listopalafoto · 2026-09-20
- Seedance 2.5 clip "Hey Janice, Please Don't Eat Me" wows with absurd drama — TomLikesRobots · 2026-09-20
- UK startup Unit1 raises $20m to stage gigs with hyper-realistic digital avatars — nordicinst · 2026-09-20
- Qwen-Image-2.1 natively generates and edits transparent RGBA images — Alibaba_Qwen · 2026-09-20