Bland Speech v3 scores 1,365 on audio realism

Bland Speech v3 scored 1,365 on Design Arena’s Audio Realism Benchmark after training on more than 100 million human conversations, ranking behind only human recordings in Bland’s launch results. The API includes HTTP and WebSocket streaming, starts with 133,000 free characters, then costs $0.015 per 1,000 characters; for teams following the shift from typed agents to full-duplex voice, the number to file is not the vendor leaderboard win but the low-cost public endpoint that makes a blind listening test practical before choosing a speech stack.