New tiny-voca text-to-speech model released on HuggingFace
On HuggingFace, on November 14, 2026, the model n4ze3m/tiny-voca was published — a compact text-to-speech for English with two voices: female F2 and male M2. The full model together with the signal generator contains 4.5 million parameters and was trained from scratch on 18 hours of synthetic speech. The author claims that on four CPU threads it runs about 70 times faster than real time. The weights are posted in ONNX, NPZ and PT formats, the license is Apache-2.0.
What it means
In our audio generators section, text-to-speech is represented by two tools: XTTS and Fish Speech. Both are noticeably heavier than tiny-voca: XTTS's largest weight file is 1.7 GB, Fish Speech's is 1.2 GB. For comparison, tiny-voca's PyTorch weight file weighs 18.2 MB, the ONNX export — 21.4 MB.
The licenses differ as well. XTTS is permitted for non-commercial use only, Fish Speech allows commercial use up to 500,000 dollars in annual revenue, and tiny-voca is distributed under Apache-2.0 without such restrictions. At the same time, tiny-voca is an English-only model with two fixed voices, whereas XTTS and Fish Speech support voice cloning and other languages.
The author of tiny-voca does not state VRAM requirements: the model is CPU-oriented and, according to him, runs even on a phone or in a browser. For comparison, YuE from our reference requires 24 GB of VRAM, but that is a tool for generating music, not speech.How the method works. The diagram was drawn for this news note.