Fish Audio Drama 3 for voice agents
October 17, 2026 · release
Fish Audio is now a cloud speech vendor on VoiceThere. Drama 3 is the default TTS model and transcribe-1 is the STT model, and you can use Fish on one leg or both. Your Fish key goes in as a project secret, so Fish bills you directly and VoiceThere speech credits are not involved.
Drama 3 takes direction tags in the text, such as [warm, friendly] or [small laugh], and wraps like <whisper>…</whisper> and <emphasis>…</emphasis>. VoiceThere passes them through unchanged, so your agent can write them into its replies.
We measured the speed from our CI against Fish's API on 2026-10-16, using one sentence of about 80 characters and three runs per path. Audio started after 291 to 658 ms over HTTP and after 963 to 1035 ms over WebSocket, so HTTP streaming is what VoiceThere uses by default. The STT final arrived about 500 ms after the end of the turn (494 to 505 ms).
Fish STT does not send partial transcripts, only finals. Because of that, barge-in for a project on Fish STT follows voice activity instead of waiting for a partial, and the transcript arrives once the caller pauses rather than word by word. Fish bills STT for the audio it receives, and VoiceThere only streams the speech it detects, not the silence between turns.
Try it
- Docs: Fish Audio speech on VoiceThere
- Overview of all vendors: STT & TTS vendors
- Dashboard: VoiceThere dashboard → your project → Voice (STT / TTS) → pick Fish Audio, add
AGENT_FISHAUDIO_API_KEYas a secret, then Deploy to cloud - CLI:
voicethere projects voice set --project YOUR_PROJECT_ID \
--stt-provider fishaudio --tts-provider fishaudio \
--stt-model transcribe-1 --tts-model drama-3
Set AGENT_FISHAUDIO_VOICE_ID if you want a specific Fish voice. Without it, Fish uses the default voice of the model.