leaxer-qwen3-tts

Single binary, C++ implementation of Qwen3-TTS running on top of ONNX Runtime.

Usage

leaxer-qwen3-tts -m <model_dir> -p "Hello world" -o output.wav

# With language hint
leaxer-qwen3-tts -m onnx_kv_06b -p "你好世界" --lang zh -o chinese.wav

# Sampling controls
leaxer-qwen3-tts -m onnx_kv_06b -p "Hello" --temp 0.7 --top-k 30 --top-p 0.9

Options

Flag	Description	Default
-m, --model	ONNX model directory	required
-p, --prompt	Text to synthesize	required
-o, --output	Output WAV path	output.wav
--lang	Language: auto, en, zh, ja, ko	auto
--temp	Sampling temperature	0.8
--top-k	Top-k sampling	50
--top-p	Top-p (nucleus) sampling	0.95
--max-tokens	Max generation tokens	2048

Building

Requirements

• CMake 3.14+
• C++17 compiler
• ONNX Runtime (user-provided)

Build

# Download ONNX Runtime from https://github.com/microsoft/onnxruntime/releases
# Extract to a directory, then:

cmake -B build -DONNXRUNTIME_DIR=/path/to/onnxruntime
cmake --build build -j

./build/leaxer-qwen3-tts --help

macOS (Homebrew)

brew install onnxruntime cmake
cmake -B build -DONNXRUNTIME_DIR=$(brew --prefix onnxruntime)
cmake --build build -j

GPU Acceleration

# CoreML (macOS, requires ONNX Runtime with CoreML support)
cmake -B build -DONNXRUNTIME_DIR=/path/to/onnxruntime -DLEAXER_COREML=ON

# CUDA (NVIDIA)
cmake -B build -DONNXRUNTIME_DIR=/path/to/onnxruntime -DLEAXER_CUDA=ON

# ROCm (AMD)
cmake -B build -DONNXRUNTIME_DIR=/path/to/onnxruntime -DLEAXER_ROCM=ON

Note: GPU acceleration requires ONNX Runtime built with the corresponding execution provider. The default Microsoft releases include CoreML for macOS.

Models

Download ONNX models from zukky/Qwen3-TTS-ONNX-DLL:

# Clone model repo (or download manually)
git lfs install
git clone https://huggingface.co/zukky/Qwen3-TTS-ONNX-DLL models

# Use the 0.6B model
./build/leaxer-qwen3-tts -m onnx/onnx_kv_06b -p "Hello"

Required files in model directory:

• text_project.onnx — text token embeddings
• codec_embed.onnx — codec token embeddings
• code_predictor_embed.onnx — sub-codec embeddings
• talker_prefill.onnx — transformer prefill
• talker_decode.onnx — transformer decode (with KV cache)
• code_predictor.onnx — predict codebooks 1-15
• tokenizer12hz_decode.onnx — vocoder (codes → audio)

Also needs tokenizer files in ../models/Qwen3-TTS-12Hz-0.6B-Base/:

• vocab.json
• merges.txt

Architecture

Text → BPE Tokenizer → Talker (prefill/decode) → Code Predictor → Vocoder → WAV
                              ↓                        ↓
                           KV Cache              Codebooks 1-15

The model generates 16 audio codebooks per frame at 12Hz, then the vocoder upsamples to 24kHz audio.

Voice Clone

Clone any voice from a 3-second reference audio:

leaxer-qwen3-tts -m onnx/onnx_kv_06b -p "Hello world" --ref voice_sample.wav -o output.wav

The reference audio should be:

• WAV format (8/16/24/32-bit PCM or float)
• At least 3 seconds of clear speech
• Will be resampled to 24kHz internally

Roadmap

Feature	Requires	Status
Voice clone (--ref)	0.6B-Base	Done
GPU acceleration	ONNX Runtime + CoreML/CUDA	Done
Preset speakers (--speaker)	0.6B-CustomVoice	Planned
Voice instructions (--instruct)	1.7B-VoiceDesign	Planned
Single static binary	Static ONNX Runtime	Planned

Credits

• Qwen3-TTS by Alibaba
• zukky/Qwen3-TTS-ONNX-DLL for ONNX exports, big thanks to Mr. Daishi Suzuki (zukky)!

License

Apache 2.0 — see LICENSE