Tontaube V1.1 released: Open-weights TTS with 4-bit CPU support and 70% preference over Fish Audio S2 Pro

Details

Tontaube V1.1 is an open-weights, 2.9B parameter autoregressive TTS model designed for long-form audio generation with zero-shot voice cloning from 5–60 seconds of reference audio.
In a blind listening study with 400 pairwise comparisons, listeners preferred the V1 model's prosody over Fish Audio S2 Pro in approximately 70% of cases.
The 4-bit GPTQ version runs in about 4 GB of VRAM or at real-time speed on a 12-core desktop CPU (Ryzen 9 5900X) without GPU acceleration, using AVX2 kernels.
GPU performance reaches about 13x real time on an RTX 5090 for single requests and 21–23x when batched, with a time-to-first-audio of ~190 ms.
V1.1 introduces five style sliders for speaking rate, pauses, intonation, pitch, and articulation, along with an audio cleanup adapter.
Supported languages include English and German as primary, with unverified support for Spanish, French, Italian, Dutch, and Portuguese; the system is limited to Linux x86-64 and NVIDIA GPUs or CPU.
Code is licensed under Apache 2.0, while weights use a community license allowing free commercial use for entities with yearly revenue below EUR 200k.

Key point