Skip to content

Fish Speech

aivrar edited this page Nov 18, 2026 · 3 revisions

Fish Speech

Fish Speech (OpenAudio S1-mini) is the engine to reach for when voice-cloning quality matters most. Give it a short reference clip and its transcript, and it produces very natural, expressive, multilingual speech.

Fish Speech panel set up with a reference voice

Use Fish Speech when you're making final output and want the closest match to the reference voice. XTTS and Kokoro are quicker for drafts and previews.

🔊 Listen: Fish Speech cloned-voice sample

Step by step

  1. Load the model. Pick a GPU and click Load. The first load downloads OpenAudio S1-mini.
  2. Pick a reference voice from voices/ (add clips in Upload & Transcribe). The preview player lets you check it.
  3. Paste the reference transcript into the reference-text box: exactly what is said in the clip. Fish uses it to align the voice, so an accurate transcript improves the clone a lot. Tip: Upload & Transcribe can produce it for you with Whisper.
  4. Choose the language and type your text.
  5. Set speed, Clean, De-ess, format and an optional save path, then click Generate.

Each chunk goes through generation → post-processing → (optional) Whisper verification:

Fish Speech generating

Fish Speech result

Settings that matter

Setting Default Guidance
Temperature 0.7 Lower is steadier; higher is more expressive.
Top P 0.7 Narrows word choices; 0.6–0.8 works well.
Verify with Whisper on Regenerates chunks that don't match your text.
Accuracy tolerance 80% Minimum Whisper similarity to accept a chunk.
Speed / Clean / De-ess 1.0 / 70% / 0% Same post-processing as the other speech engines.

Tips

  • The reference transcript must match the clip word for word. A wrong transcript is the most common cause of odd results.
  • Clean, dry recordings of 10–20 seconds work best.
  • Fish likes VRAM headroom (about 8 GB). For the best speed, give it a GPU that isn't shared with other loaded engines or apps.
  • For long scripts, set a save path so ##recover## can resume a job that stops part-way.

API

curl -X POST http://127.0.0.1:5006/fish_infer -H "Content-Type: application/json" -d "{\"text\": \"Hello from Fish Speech.\", \"voice\": \"MyVoice.wav\", \"ref_text\": \"Transcript of the reference clip.\"}"

Load first with POST /fish_load and {"device": "0"}. Use the panel's API Payload button for the full list of fields. More in API Reference and chapter 6 of the manual.

Clone this wiki locally