Bland Speech v3 tops blind TTS benchmark with Elo 1237, second only to real humans
ycombinator · x · 2026-09-25
Voice-calling company Bland released Bland Speech v3, billed as the most human text-to-speech model. On intelligence.ai's Audio Realism Bench (blind Elo listening test), v3 scores 1237 — the best machine voice, behind only hidden real-human recordings (1384) — beating MAI-Voice-2, Grok TTS, GPT Realtime 2, Gemini TTS models, and Eleven v3. Customer Superunit reports v3 beat rivals by 27% on 19,000+ employee verification calls and made it the default voice.
More from Models
- Claude Opus 5.5 builds Minecraft from one prompt in ~1 hour for ~$20 — amasad · 2026-09-25
- Dev finds Opus 5.5 Medium reasoning so good that High feels unnecessary — rudrank · 2026-09-25
- PINNACLE: GPT-6 Sol cuts errors 2.5x at max effort, Claude Opus 5.5 doesn't benefit — ryanshrout · 2026-09-25
- 4B open model tops JevBench by being 5x faster and half the price — airesearch12 · 2026-09-25
- Terminal-Bench-Science leaderboard launches with GPT-6 Astra at 63% — scaling01 · 2026-09-25
- User says Opus 5.5 writes so well he deleted his 'don't write like a fuckhead' custom instruction — generativist · 2026-09-25