Qwen3-TTS Pure C Implementation
A dependency-free C inference engine that runs Qwen3-TTS text-to-speech models on the CPU using only a BLAS library, with support for nine voices and ten languages.
This project is a lightweight, cross-platform C inference engine for the Qwen3-TTS text-to-speech models, in 0.6B and 1.7B sizes. It needs no Python, PyTorch or ONNX runtime, only a C compiler and a BLAS library, and it memory-maps model weights directly from safetensors files in BF16 so loading is close to instant and memory use stays low.
The engine runs the full text-to-speech pipeline itself: BPE tokenization, a 28-layer causal transformer it calls the Talker, a multi-pass code predictor, and a convolutional speech decoder. It includes a document mapping which layers of the model carry timbre, language, prosody and emotion, meant as background for voice cloning work.
On a 2020 Apple M1 CPU with no GPU, the 0.6B model generates speech faster than real time.