
Tontaube
Open-weights, on-device TTS for voice agents and long-form narration, with one-file voice cloning.

About Tontaube
Tontaube: Text-to-Speech Models for Voice Agents and Long-Form Narration
Tontaube is a text-to-speech model built for voice AI agents and long-form narration, with open weights on Hugging Face and a pay-as-you-go API at $5 per million characters. It can run on-device, streams the first audio in about 200 ms on a single RTX 5090 and clones a voice from one audio file.
Key Features
Open weights: Tontaube V1.1 weights are on Hugging Face and the inference code is on GitHub under Apache 2.0.
Streaming output: about 200 ms to first audio and a real-time factor of 0.08 for one input on a single RTX 5090, measured on Tontaube V1. A 4-bit build also runs on small GPUs and on CPUs, at about real time on a 12-core desktop CPU.
Voice cloning: clone a voice from a single audio file; the model accepts up to one minute of reference audio.
Long-form narration: designed to keep prosody natural across long passages such as audiobook chapters, with bounded context for long inputs.
Style controls: V1.1 adds controls for speaking rate, pauses, intonation, pitch and articulation.
Languages: designed primarily for English and German, with additional support for Spanish, French, Italian, Dutch and Portuguese.
Python SDK: a client on PyPI lists speakers and generates speech in a few lines of code.
Published scores: Tontaube reports 1.66% Seed-TTS word error rate (English zero-shot, Whisper large-v3) for V1, with 2.9B parameters across four predictors.
Enterprise options: uptime SLA, private cloud or on-premise deployment, and custom DPA, with GDPR compliance and DPA and SCC on the usage plan.
Use Cases
Voice agents: low-latency speech output for conversational AI agents.
Audiobooks and long reads: narrating books, articles and documents; Tontaube also runs an iOS and Android app that turns PDFs and EPUBs into audiobooks.
Custom voices: cloning a speaker's voice from a short sample for consistent narration.
On-device and self-hosted TTS: running the open weights on your own GPU or CPU for research or product use.
Pricing
The API is usage based at $5 per million characters (about $0.0042 per minute of audio), with 200,000 free characters on sign-up, no subscription and a balance that never expires. Enterprise pricing is negotiated. The open weights ship under the Tontaube Community Model License 1.0: free for research, testing and commercial use below EUR 200,000 in group revenue (with "Powered by TontaubeV1" attribution); larger companies and anyone offering the model as a speech service need a commercial license.
Getting Started
Website: https://tontaube.ai
Documentation: https://tontaube.ai/api-reference
Model weights: https://huggingface.co/TontaubeAI/TontaubeV1.1
Tontaube gives developers a low-cost streaming text-to-speech model with voice cloning, available as open weights or as a pay-as-you-go API.

