Qwen3-TTS

Disclaimer

This specific page uses temporary drafts written and maintained by Generative AI. While verified to be accurate, please cross-verify with source code.

Synthesize highly accurate multilingual character speech using Qwen3-TTS in both voice cloning and text-described VoiceDesign modes.

Qwen3-TTS 12Hz 1.7B provides high-accuracy local speech synthesis. Running servers/tts/qwen3tts/server.py in its default auto mode dynamically selects the Base voice-cloning model or the VoiceDesign model based on each incoming request.

Setup

Run these commands from the TomoriBot repo root, the folder where you cloned TomoriBot:

Windows PowerShell

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

Linux and macOS Bash

python3 -m venv servers/tts/qwen3tts/.venv
source servers/tts/qwen3tts/.venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r servers/tts/qwen3tts/requirements.txt
python servers/tts/qwen3tts/server.py

The default auto-mode endpoint URL is http://127.0.0.1:8012; set QWEN3TTS_PORT to use another port. 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.

Fast CUDA backend

By default the server uses faster-qwen3-tts, which the requirements install. On an NVIDIA GPU it synthesizes about 6x faster than the standard backend, so speech generates faster than real time instead of about twice as slow. It needs no extra setup, but it only runs on NVIDIA GPUs: on a CPU, a Mac, or an AMD GPU the server stops at startup and tells you to switch backends.

This backend reserves room for a fixed amount of speech per request, shared between the reference clip and the generated audio (about 170 seconds in total). With a 10-20 second clip that leaves well over two minutes of speech. If a request does not fit, the server returns an error instead of cutting the audio short; use a shorter clip or text, or the standard backend.

To use the standard backend instead, set QWEN3TTS_BACKEND=transformers before starting the server:

$env:QWEN3TTS_BACKEND = "transformers"
python servers\tts\qwen3tts\server.py

Register in TomoriBot

For most users, register the auto-mode server so one endpoint can support both voice-clone and VoiceDesign personas.

Run /providers, choose Add New Custom Endpoint, and use the speech API compatibility:

• API Compatibility: tts-clone
• endpoint_url: http://127.0.0.1:8012

After saving the connection, select it and use its model dropdown to add a Speech model. The model form asks for Voice Source Mode and Script Markup; choose Auto and Plain for the auto-mode server.

Use /providers for endpoint registration and model setup. Then open /config > Models > Switch Models to select and activate the registered endpoint.

Set up persona voices

Voice cloning

Use this for personas that should imitate a reference clip:

1. Prepare a clean 10-20 second voice clip with one speaker and no background music.
2. Open /config under Models > TTS Parameters & Voices and upload the clip.
3. Open /config under Persona > Voice, then choose the persona and the voice sample.

Qwen3-TTS supports reference clips as short as 3 seconds.

VoiceDesign

Use this for personas that should use a written voice description instead of a sample:

1. Open /config under Persona > Voice and choose VoiceDesign.
2. Choose the persona.
3. Enter a natural-language voice prompt, such as the speaker’s age, tone, accent, and delivery.

Optional: VoiceDesign-only server

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

Windows PowerShell:

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

Bash:

source servers/tts/qwen3tts/.venv/bin/activate
TOMORI_TTS_MODE=voice-design python servers/tts/qwen3tts/server.py

You can also pass --mode voice-design instead of setting TOMORI_TTS_MODE. The default VoiceDesign-only endpoint URL is http://127.0.0.1:8014.

Register it the same way as auto mode, but use endpoint URL http://127.0.0.1:8014 and choose VoiceDesign as the Voice Source Mode on the Speech model.