Qwen3 TTS Model Card#

Model identity#

This package integrates the Qwen3-TTS 0.6B Base architecture through the NVIGI Qwen3 TTS plugin. The packaged model pair consists of:

  • qwen-talker-0.6b-base-Q4_K_M.gguf

  • qwen-talker-0.6b-base-Q8_0.gguf (optional)

  • qwen-tokenizer-12hz-F32.gguf

Q4_K_M is the default Talker for lower disk and VRAM use. Q8_0 is an optional, explicitly selected higher-precision Talker and may use more disk/VRAM and run more slowly. The tokenizer/codec remains F32 for both variants.

Intended use#

The model is intended for local text-to-speech, interactive narration, and voice-cloning workflows in Windows applications and games. The package provides CUDA, Vulkan, and D3D12 plugin variants.

The Qwen Base model supports the languages exposed by this package: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian.

Voice cloning#

The package does not include Qwen’s example reference WAV or a JSON derived from it. The opt-in provisioning script downloads the Qwen-hosted example at the user’s request and creates a local JSON containing the speaker and ICL-code latents. A user can also generate a JSON from other approved audio.

New references must be created only from approved and consented source audio. Use voice_cloning\get_voice_embeddings.py to generate a model-compatible JSON. Only Qwen3 voice-reference JSONs are compatible with this plugin.

Output and limitations#

  • Output is mono 24 kHz PCM audio.

  • Synthesized speech quality depends on the selected backend, text, language, reference quality, and available GPU memory.

  • A reference transcript should accurately match the audio to use ICL codes.

  • A precomputed JSON is tied to the Qwen model format used to create it.

  • Audio quality, speaker similarity, and suitability for a specific production use must be evaluated by the product team and QA.

License and attribution#

The packaged model assets are derived from Qwen/Qwen3-TTS-12Hz-0.6B-Base revision 5d83992436eae1d760afd27aff78a71d676296fc and Qwen/Qwen3-TTS-Tokenizer-12Hz revision 7dd38ad4e9bad454aae9cd937d0cd577604fe229. Both upstream repositories identify the assets as Apache-2.0. The package-root 3rd-party-licenses.html records attribution, conversion details, and model hashes; the complete license text is included locally at licenses\Qwen3-TTS-Apache-2.0.txt.