Tontaube V1.1 compresses a 2.9-billion-parameter voice model into roughly 4GB of VRAM and can generate near real time on a 12-core desktop CPU. That is a clear sign that the barrier to local speech synthesis has fallen, and it is worth our attention.
What this is
Tontaube is a text-to-speech model designed primarily for long-form audio such as audiobooks, article narration, and voice agents. It uses open weights and can clone a voice from 5–60 seconds of reference audio, while offering controls for speaking rate, pauses, intonation, pitch, and pronunciation.
The focus of V1.1 is not parameter count but deployment efficiency: the quantized version is about 1.8GB to download and occupies roughly 4GB of VRAM once loaded, with peak usage of about 6.5GB for long text. It can run at approximately 13 times real-time speed on an RTX 5090 and near real time on a Ryzen 9 5900X desktop processor.
Industry view
We believe this reflects a shift in speech-model competition from whether a model can speak at all to whether it can speak continuously and at low cost. Cloud speech services remain the mainstream choice for commercial deployments, but they come with usage-based fees, data uploads, and external dependencies. Local inference is better suited to organizations that require privacy, low latency, or bulk generation.
The risks are equally clear. Blind-listening results published by the developers showed that its prosody won about 70% of 400 pairwise comparisons against Fish Audio S2 Pro, but these were the team’s own tests and primarily covered English and German. Its conversational and agent-oriented speech performance has also been acknowledged as less stable than its audiobook mode. In addition, the model currently supports only Linux x86-64; Windows users must run it through WSL2. Its weights use a custom community license and cannot be treated as equivalent to a standard open-source license. Teams should verify the authorization boundaries before commercial use.
Impact on regular people
For enterprise IT: Businesses can use ordinary desktop hardware to generate speech in bulk for courses, customer service, or internal training, reducing some cloud API costs while keeping sensitive text on local infrastructure. Even so, deployments still require GPUs, storage, and engineering expertise.
For individual professionals: Knowledge workers can quickly turn articles, reports, and presentations into audio. However, voice cloning raises consent and impersonation risks, so anyone publishing content in a specific person’s voice should obtain explicit permission first.
For consumer markets: As local generation costs fall, podcasts, audio content, dubbing tools, and voice assistants may become cheaper. Whether the experience reaches mainstream users will still depend on Chinese-language quality, platform compatibility, and licensing rules.