wfloat OpenAI-Compatible Server

Small local server for using wfloat with SillyTavern's OpenAI Compatible TTS provider.

It uses the wfloat-python package.

Install

git clone https://github.com/wfloat/wfloat-openai-server.git
cd wfloat-openai-server

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run

uvicorn wfloat_openai_server.app:app --host 127.0.0.1 --port 8123

SillyTavern

In SillyTavern:

1. Open Extensions -> TTS
2. Set Enabled to on
3. Set Provider to OpenAI Compatible
4. Set Provider Endpoint to http://127.0.0.1:8123/v1/audio/speech
5. Set Model to wfloat-tts
6. Leave the API key empty
7. In Available Voices, paste something like:

clever_villain_woman,strong_hero_man:anger:0.8,skilled_hero_woman,mad_scientist_woman:surprise:0.7

8. Click Reload
9. Click Available voices to test playback

Voice format

The voice field can be:

• voice_name
• voice_name:emotion
• voice_name:emotion:intensity

Examples:

• outgoing_anime_woman
• outgoing_anime_woman:joy
• mad_scientist_woman:surprise:0.9

Voice options:

• skilled_hero_man
• skilled_hero_woman
• fun_hero_man
• fun_hero_woman
• strong_hero_man
• strong_hero_woman
• mad_scientist_man
• mad_scientist_woman
• clever_villain_man
• clever_villain_woman
• narrator_man
• narrator_woman
• wise_elder_man
• wise_elder_woman
• outgoing_anime_man
• outgoing_anime_woman
• scary_villain_man
• scary_villain_woman
• news_reporter_man
• news_reporter_woman

Valid emotions:

• neutral
• joy
• sadness
• anger
• fear
• surprise
• dismissive
• confusion

Intensity range:

• 0.0 to 1.0

Notes

• The server uses the local wfloat package.
• speed is passed through to wfloat.
• The server always returns WAV audio.
• SillyTavern sends response_format: "mp3" for this provider, but it still plays WAV responses correctly.

Environment

• WFLOAT_MODEL_NAME Default: wfloat/wfloat-tts
• WFLOAT_DEFAULT_VOICE Default: wise_elder_man
• WFLOAT_DEFAULT_EMOTION Default: neutral
• WFLOAT_DEFAULT_INTENSITY Default: 0.5