README.md
| 1 | --- |
| 2 | license: apache-2.0 |
| 3 | library_name: gguf |
| 4 | pipeline_tag: text-to-speech |
| 5 | tags: |
| 6 | - tts |
| 7 | - text-to-speech |
| 8 | - voice-cloning |
| 9 | - voice-design |
| 10 | - mandarin-dialects |
| 11 | - ggml |
| 12 | - gguf |
| 13 | - qwen |
| 14 | - qwen3 |
| 15 | - qwen3-tts |
| 16 | - cpp |
| 17 | language: |
| 18 | - zh |
| 19 | - en |
| 20 | - fr |
| 21 | - de |
| 22 | - es |
| 23 | - it |
| 24 | - pt |
| 25 | - ja |
| 26 | - ko |
| 27 | - ru |
| 28 | - ar |
| 29 | base_model: |
| 30 | - Qwen/Qwen3-TTS-Tokenizer-12Hz |
| 31 | - Qwen/Qwen3-TTS-12Hz-0.6B-Base |
| 32 | - Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice |
| 33 | - Qwen/Qwen3-TTS-12Hz-1.7B-Base |
| 34 | - Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice |
| 35 | - Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign |
| 36 | --- |
| 37 | |
| 38 | # Qwen3-TTS GGUF |
| 39 | |
| 40 | GGUF weights for [qwentts.cpp](https://github.com/ServeurpersoCom/qwentts.cpp), |
| 41 | a C++17/GGML port of Qwen3-TTS 12 Hz (Qwen team, Alibaba). Multilingual |
| 42 | zero shot TTS with named speakers and Mandarin dialects, 24 kHz mono. |
| 43 | Runs on CPU, CUDA, Metal, Vulkan. |
| 44 | |
| 45 | ## Files |
| 46 | |
| 47 | Two GGUFs load together : |
| 48 | |
| 49 | qwen-talker-{size}-{mode}-{variant}.gguf Qwen3 LM + code predictor MTP head + optional speaker encoder, text -> 12 Hz codes |
| 50 | qwen-tokenizer-12hz-{variant}.gguf SEANet + ConvNeXt + DAC v2 + RVQ, 12 Hz codes <-> 24 kHz audio |
| 51 | |
| 52 | Three modes are available across two talker sizes : |
| 53 | |
| 54 | | mode | 0.6B | 1.7B | use case | |
| 55 | |-------------|------|------|---------------------------------------------------| |
| 56 | | base | yes | yes | zero shot TTS with named speakers and dialects | |
| 57 | | customvoice | yes | yes | zero shot voice cloning from a reference clip | |
| 58 | | voicedesign | no | yes | voice synthesis from attribute description | |
| 59 | |
| 60 | The tokenizer is shared across every talker. |
| 61 | |
| 62 | | variant | talker 0.6B | talker 1.7B | tokenizer | use case | |
| 63 | |---------|-------------|-------------|-----------|--------------------------------| |
| 64 | | F32 | 3.7 GB | 7.7 GB | 647 MB | reference, debug, conversion | |
| 65 | | BF16 | 1.8 GB | 3.9 GB | 359 MB | source faithful, max precision | |
| 66 | | Q8_0 | 993 MB | 2.1 GB | 291 MB | recommended default | |
| 67 | | Q4_K_M | 629 MB | 1.2 GB | 255 MB | lowest VRAM | |
| 68 | |
| 69 | ## Quick start |
| 70 | |
| 71 | ``` |
| 72 | git clone --recurse-submodules https://github.com/ServeurpersoCom/qwentts.cpp.git |
| 73 | cd qwentts.cpp && ./buildcuda.sh |
| 74 | mkdir -p models |
| 75 | huggingface-cli download Serveurperso/Qwen3-TTS-GGUF \ |
| 76 | qwen-talker-1.7b-base-Q8_0.gguf qwen-tokenizer-12hz-Q8_0.gguf \ |
| 77 | --local-dir models |
| 78 | cd examples |
| 79 | ./base.sh # named speaker -> base.wav |
| 80 | ./clone.sh # voice cloning -> clone.wav |
| 81 | ./customvoice.sh # custom voice mode -> customvoice.wav |
| 82 | ./tts.sh # voice design -> tts.wav |
| 83 | ``` |
| 84 | |
| 85 | ## Backends |
| 86 | |
| 87 | Set `GGML_BACKEND` to force a device, otherwise the runtime picks the |
| 88 | best one available. |
| 89 | |
| 90 | | value | target | |
| 91 | |-----------|----------------------------------------------| |
| 92 | | `CUDA0` | NVIDIA GPU, fastest path on Ada / Blackwell | |
| 93 | | `Vulkan0` | Cross vendor GPU (AMD / Intel / NVIDIA) | |
| 94 | | `Metal` | Apple Silicon GPU | |
| 95 | | `CPU` | CPU fallback, x86 variant auto selected | |
| 96 | |
| 97 | ## Quantization policy |
| 98 | |
| 99 | Tokenizer GGUFs are not uniform quants. Three categories get a |
| 100 | dedicated treatment : |
| 101 | |
| 102 | | tensor | dtype across all variants | |
| 103 | |--------------------------------------------------------------|---------------------------| |
| 104 | | RVQ codebooks, input_proj / output_proj, speaker encoder fc | F32 | |
| 105 | | 1D tensors (gamma, biases, norms, snake alpha and beta) | F32 | |
| 106 | | Conv kernels with non alignable rows (K=7,3,1) | F16 in Q* variants | |
| 107 | |
| 108 | Conv kernel rows (K=7,3,1) never divide a K-quant block size, so the |
| 109 | quantizer skips the Q* intermediates and lands on F16 directly. This |
| 110 | is the last resort branch of llama.cpp's `tensor_type_fallback` |
| 111 | applied unconditionally for these kernels. F16 has no block size and |
| 112 | matches the runtime target dtype on every backend. The talker LM |
| 113 | (Qwen3 backbone, hidden divisible by 256) follows standard llama.cpp |
| 114 | K-quant across variants. The code predictor MTP head and the speaker |
| 115 | encoder live in the talker GGUF and share its quantization. |
| 116 | |
| 117 | ## License |
| 118 | |
| 119 | Upstream model : Qwen3-TTS by Alibaba / Qwen team, Apache 2.0 |
| 120 | Audio codec : Qwen3-TTS-Tokenizer-12Hz (Qwen team), Apache 2.0 |
| 121 | GGUF tooling : qwentts.cpp, MIT |