Tontaube V1.1 adds open-weight voice cloning and lower-resource deployment
Tontaube developers released Tontaube V1.1, an open-weight, 2.9B-parameter autoregressive text-to-speech model for long-form audio. It supports zero-shot voice cloning from 5–60 seconds of reference audio and includes 46 voices. Developers say the 4-bit GPTQ version loads in about 4 GB of VRAM and runs at roughly real time on a 12-core desktop CPU. English and German are the main supported languages, while Spanish, French, Italian, Dutch and Portuguese have not been checked by native speakers. V1.1 adds five style controls, an audio-cleanup adapter and a server with GPU and CPU kernels.
The release combines long-form speech generation and voice cloning with a lower-resource 4-bit and CPU deployment path, making the model more accessible for local use.
We just released Tontaube V1.1, a text-to-speech model built for long-form audio (audiobooks, articles, voice agents) with zero-shot voice cloning. Our previous V1 release needed a lot of VRAM because of its vLLM setup, so this version is about making it more accessible: the 4-bit weights run in about 4 GB of VRAM, or at about real time on a desktop CPU. In a blind listening study with 400 pairwise comparisons, listeners preferred our V1 model's prosody over Fish Audio S2 Pro in about 70% of cases (details are in the paper). V1.1 is a post-trained version of this model.
What it is
- ~2.9B parameters, autoregressive, streams audio as it generates
- Clones a voice from 5–60 seconds of reference audio; 46 voices are bundled
- English and German are the main languages (works best in English). Spanish, French, Italian, Dutch and Portuguese work but have not been checked by native speakers
- Long texts are split at sentence boundaries and joined, so you can feed it whole chapters
Running it locally
- The default is a 4-bit (GPTQ) version: 1.8 GB download, loads in about 4 GB of VRAM, peaks at about 6.5 GB on long texts
- GPU speed: about 13x real time on an RTX 5090 for a single request (21–23x batched), ~190 ms to first audio when streaming. About 6–7x on an RTX 3090
- CPU: no GPU needed. About real time on a 12-core desktop CPU (Ryzen 9 5900X, dual-channel DDR4). It is limited by memory bandwidth and needs AVX2
- No vLLM or llama.cpp needed: it ships its own server with Triton kernels for the GPU and AVX2 kernels for the CPU. Full-precision weights and a vLLM backend are there too if you have 24 GB+ and want throughputgit clone --branch v1.1.0 --depth 1 https://github.com/craitech/tontaube.git cd tontaube uv run --python 3.12 --frozen tontaube serve