Craitech Unveils Tontaube V1.1: Local Voice Cloning TTS Running on CPU with 1.8GB Footprint
Craitech has released Tontaube V1.1, a local text-to-speech model featuring a lightweight 1.8GB 4-bit quantized edition, CPU inference support, and five speakin
On November 15, 2026, AI audio developer Craitech publicly released model weights and an inference server for 'Tontaube V1.1', a text-to-speech (TTS) synthesis system capable of running voice cloning and low-latency streaming generation directly on local personal computers.

Image source: Hane Lab
The central focus of this release is eliminating reliance on cloud APIs and processing audio locally without transmitting sensitive voice data across external servers. By introducing a 4-bit (int4) integer quantized build that compresses the model download footprint down to approximately 1.8GB, Craitech enables local speech synthesis not only on modest consumer graphics cards but also on standard central processing units (CPUs).
Core Architecture and Local Voice Cloning in Tontaube V1.1
Tontaube V1.1 is an open text-to-speech engine built to convert written text into natural human speech entirely within a user's local operating environment. While conventional cloud speech synthesis APIs charge recurring usage fees per audio minute and introduce latency alongside potential privacy liabilities, Tontaube operates in an offline, self-contained setup.
The model is engineered around three primary capabilities: voice cloning from a short reference audio sample, stable reading of extended long-form text, and low-latency sequential streaming generation where audio starts playing back from the very first spoken segment without awaiting complete sentence synthesis.
With the V1.1 update, Craitech introduces five granular speaking-style adjustment knobs alongside an integration pipeline for external audio models, specifically connecting with the VibeVoice ecosystem. This shifts the workflow from passive text rendering into an interactive audio workbench where creators can manipulate speech expression directly.
4-Bit Quantization and Low-Barrier CPU Execution
The most practical breakthrough in Tontaube V1.1 lies in the dramatic reduction of computational resource requirements.
Conventional high-fidelity speech synthesis models are typically distributed in 16-bit floating-point precision formats. Tontaube's standard full-precision weight distribution weighs roughly 6.4GB, and running it at high speeds via the open-source vLLM serving engine requires at least 24GB of dedicated GPU memory (VRAM). This requirement effectively locked out mainstream creators without enterprise or high-end workstation hardware.
To lower this deployment threshold, Craitech prepared an int4 quantized release:
- 1.8GB Download Footprint: Shrinking the model size by over 70% compared to the 6.4GB base weights, making local storage and bandwidth requirements manageable.
- Support for Modest GPUs and CPUs: Users without dedicated enterprise accelerators can load and execute the model directly across entry-level graphics cards or general-purpose desktop CPUs.
For creators operating under limited network connectivity or workflows requiring strict data residency guarantees, the 1.8GB footprint establishes a viable entry point for on-device voice generation.
Five Granular Speaking-Style Knobs and VibeVoice Integration
A recurring hurdle in local speech synthesis has been monotonic or emotionally mismatched voice outputs. Tontaube V1.1 tackles this by exposing five explicit parametric control knobs:
- Speaking Rate: Regulates the overall pace of recitation, accommodating both dense narrations and fast-paced conversational speech.
- Pauses: Controls the duration of breathing intervals between punctuation marks and paragraph breaks.
- Intonation: Modulates pitch contours and accents at clause boundaries to reinforce expressive clarity.
- Pitch: Finely shifts base vocal frequencies higher or lower to change vocal weight and warmth.
- Clarity of Pronunciation: Adjusts phonemic articulation intensity, shifting output between casual colloquial speech and sharp broadcast-style delivery.
Furthermore, built-in interoperability with VibeVoice allows creators to route synthetic voice outputs through secondary voice models for tonal re-shaping and acoustic texture enhancement.
Language Coverage Limitations and Dual Licensing Structure
Prospective implementers should evaluate several concrete constraints before integrating Tontaube V1.1 into commercial or personal pipelines.
First, language coverage remains strictly bounded. The current version officially supports seven European languages: English, German, Spanish, French, Italian, Dutch, and Portuguese. Asian languages, including Korean and Japanese, are not supported in this release, precluding native localized narration for those linguistic markets.
Second, Craitech maintains a bifurcated licensing approach:
- Inference Server: Distributed under the permissive Apache-2.0 license, permitting flexible pipeline customization and integration.
- Model Weights: Governed by a dedicated 'Community License' that outlines specific usage limitations, particularly around commercial deployment and redistribution.
Teams evaluating Tontaube V1.1 for client-facing software or commercial voice automation must review the community license terms hosted on the Hugging Face repository prior to production deployment.
Sources
- Hugging Face Official Model Card: TontaubeAI/TontaubeV1.1
- Hane Lab Tech News Flash: Generative AI News: Seedance 2.5 Flash, Tontaube V1.1 and Wonder 4