Open-weights TTS model Tontaube V1.1 runs near real-time on a desktop CPU in 4-bit
EAVDR · reddit · 2026-11-14
Tontaube V1.1 is an open-weights, 2.9B-parameter TTS model built for long-form audio (audiobooks, articles, voice agents) with zero-shot voice cloning from 5–60 seconds of reference audio (46 bundled voices). In a 400-comparison blind study, its V1 prosody beat Fish Audio S2 Pro in 70% of cases.
- 4-bit GPTQ version: 1.8 GB download, 4 GB VRAM, or near real-time on a 12-core desktop CPU (memory-bandwidth bound, AVX2 required)
- 13x real time on an RTX 5090 (21–23x batched), 190 ms to first streamed audio; 6–7x on RTX 3090
- Ships its own server with Triton/AVX2 kernels plus HTTP/SSE/WebSocket API and browser UI; no vLLM or llama.cpp needed
- New in V1.1: five style sliders (rate, pauses, intonation, pitch, articulation), the 4-bit/CPU path, and an audio cleanup adapter
- Linux x86-64 only (NVIDIA or CPU); English/German primary; conversational style less reliable; code Apache 2.0, weights free for research/personal and commercial use under EUR 200k revenue with credit
More from Infra
- Developer now runs 100% of his agents locally on a Mac Mini, up from 25% earlier this year — josh_bickett · 2026-11-14
- Nvidia's $530B backstop for AI customers: the six assumptions that must hold — iamKierraD · 2026-11-14
- RTX 5070 Ti ComfyUI benchmarks: overclocking and launch flags yield surprising results — AnalElder · 2026-11-14
- Bernstein: Neoclouds earn ~6x more revenue per megawatt than Bitcoin miners — Beth_Kindig · 2026-11-14
- ZONOS3-Mini: MoE TTS model runs realtime on a Pixel 9 and MacBook Air — cocktailpeanut · 2026-11-14
- The Periodic Table of Inference: 12 LLM serving concepts as meme elements — sophiamyang · 2026-11-14