Fish Audio Realtime STT Understands Conversation Turns

Fish Audio is pushing STT beyond transcription toward conversation control, an important layer for making voice agents feel less brittle.

• –`transcribe-stream-1` provides turn detection, eager end-of-turn signals, backchannel awareness, and conversation context
• –Developers can use those signals to decide when an agent should answer, continue speaking, or stop
• –Fish Audio cites a 2.21% WER and first-place ranking on Cekura’s 20-system benchmark, though the current leaderboard shows a different snapshot
• –The WebSocket API is billed at $0.36 per session hour, making cost predictable for sustained voice-agent sessions
• –Preview status means latency, turn detection, and real-world interruption handling still need independent testing