- 4,537,291 parameters, 18 MB. This is the whole model, the waveform part is included. - Two voices, one female and one male. 24 kHz. - About 70x faster than real time on 4 CPU threads. - Runs with PyTorch, ONNX, and in the browser. - Trained from scratch in 7.7 hours on one RTX 4090.
What I think is different:
- Everything is open. Code, weights, the training recipe and the dataset. - Custom phonemizer. - No neural vocoder. The sound is made by a small source-filter synth, which is why it is fast.
What is not good yet:
- English only, and only these two voices. No voice cloning. - Foreign names and odd spellings are often read wrong. - It can sound a bit synthetic. - The training audio is synthetic too. I made it with Qwen3-TTS, so the model can't sound better than that.
This weekend, I did a fun experiment to create something similar to Jev. I LoRA fine-tuned Qwen3.5 4B using a mix of publicly available datasets and synthetic data.
For the synthetic data, I used DeepSeek V4.1 Flash, around 25M tokens. I trained the model for about 2 hours on a rented RTX 3090.
So far, the model performs pretty well. It's obviously nowhere near Jev yet, but compared to the base model, it works surprisingly well (0.596 -> 0.709 on typed-decisions, Jev is at 0.727).
I've open-sourced everything, including the synthetic dataset and a Jev-compatible API endpoint, so you can try it here: