Qwen3-TTS

Use servers/tts/qwen3tts/server.py for both Qwen3-TTS 12Hz 1.7B modes. By default it starts in auto mode, which chooses the Base voice-clone model or VoiceDesign model from each request shape.

python -m venv servers\tts\qwen3tts\.venv
servers\tts\qwen3tts\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install -r servers\tts\qwen3tts\requirements.txt
python servers\tts\qwen3tts\server.py

You can also specify auto mode explicitly:

python servers\tts\qwen3tts\server.py --mode auto

Auto mode inspects each /synthesize request: requests with ref_audio use the clone model, while requests with instruct use the VoiceDesign model. It keeps only one model loaded at a time and swaps models when the request type changes, so the first request after a swap may be slower.

In TomoriBot, select Auto as the speech endpoint’s voice source mode when one endpoint URL should support mixed personas. Personas configured with /speech voice-assign use clone synthesis; personas configured with /speech voice-design set use VoiceDesign synthesis. Voice samples and VoiceDesign prompts are kept as reusable persona data; the most recently selected voice mode becomes active without deleting the other saved setup.

Register with /provider custom-endpoint add:

• capability: speech
• api_style: tts-clone
• endpoint_url: your wrapper URL
• script_markup: plain
• supports_instruct: false

VoiceDesign

Start the same server in VoiceDesign mode when serving Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign.

servers\tts\qwen3tts\.venv\Scripts\Activate.ps1
$env:TOMORI_TTS_MODE = "voice-design"
python servers\tts\qwen3tts\server.py

You can also pass --mode voice-design instead of setting TOMORI_TTS_MODE.

For mixed clone and VoiceDesign usage on one endpoint URL, prefer --mode auto and register a TomoriBot custom endpoint with the Auto voice source mode.

Register it the same way as the clone wrapper, but select the VoiceDesign voice source mode. TomoriBot treats VoiceDesign mode as instruct-capable automatically; the Supports Instruct checkbox does not need to be selected separately.

Set each persona’s natural-language voice description with /speech voice-design set. Remove it with /speech voice-design remove. During generation, TomoriBot sends that prompt in the /synthesize JSON body as instruct; when the tool includes voice_instructions, those one-off delivery notes are appended to the instruct text for that message only. The wrapper still accepts ref_text or TOMORI_TTS_DEFAULT_INSTRUCT as fallbacks for manual testing or single-voice deployments.

Debugging VoiceDesign Requests

TomoriBot logs VoiceDesign /synthesize requests with endpoint metadata, script length, instruct length, and short script/instruct previews. Set TTS_VOICE_DESIGN_LOG_PAYLOADS=true in the bot .env to print the full script and full instruct payload before the request is sent.

The Qwen3-TTS wrapper also logs each request’s inferred mode, text length, instruct length, active model, and generation timing. Set TOMORI_TTS_LOG_PAYLOADS=1 before starting server.py to print the full request text and instruct prompt in the wrapper logs; otherwise it prints previews controlled by TOMORI_TTS_LOG_PREVIEW_CHARS.

$env:TOMORI_TTS_LOG_PAYLOADS = "1"
python servers\tts\qwen3tts\server.py --mode auto

When vLLM-Omni or another stable hosted serving path supports this model, prefer documenting that as the production path and keep this wrapper as a reference implementation.