An Open TTS Team Says the Standard "Treat It Like an LLM" Trick Is Wrong for Speech

Open text-to-speech has been copying the same moves from large language models: borrow the tokenizer, train a decoder on top, ship it. A new open-weight release this month breaks that template on purpose, swapping the inherited tokenizer for character-level input.

That release is TontaubeV1, a 2.9B-parameter open-weight model announced on r/MachineLearning. The team reads the input one character at a time rather than inheriting a language model's tokenizer, and argues the design choice is worth testing on its own.

Character-level tokenization means the model sees each letter of the script as its own token, rather than grouping common letter sequences into single tokens the way a language model would. The Tontaube team argues this avoids the rare-token combinations that speech training data under-covers, and gives the model a more stable mapping for long-form narration and one-minute reference voice cloning.

The architecture is a four-stage causal codebook stack built on DualCodec, a multi-codebook discrete audio codec. The first stage turns text into semantic audio and durations; three progressively smaller stages add acoustic detail. The semantic stage is initialized from Qwen3-1.7B, a 1.7-billion-parameter open language model from Alibaba's Qwen team, and forced to operate over characters rather than Qwen's native BPE tokens.

Beyond the tokenizer swap, the authors foreground a second design choice aimed at long-form generation. The model is trained with chunk boundaries encoded directly into the position layout of the token stream, rather than applied as a post-hoc split. The team's pitch is that this lets the model treat chunked narration as part of the training signal itself, so long passages can be generated without the boundary artifacts a sliding-window post-processing approach would produce.

Training data is described by the authors as roughly 200,000 hours of audio across seven languages. English and German are listed as the primary tested languages, with Spanish, French, Italian, Dutch, and Portuguese also supported. Voice cloning is zero-shot from up to one minute of reference audio, and the model ships with three style controls: audiobook, conversational, and agentic. Weights are distributed on Hugging Face under TontaubeAI/TontaubeV1, paired with a vLLM model adapter and a separate inference repository at craitech/tontaube.

All quality, latency, and comparison claims in this release are author-stated. The Reddit announcement, the Hugging Face model card, and the GitHub inference repository are all developer-controlled surfaces; no third-party benchmark, listening comparison, or independent reproduction is in the public record at the time of writing. The authors describe the model as running with sub-second local latency, but that figure is reported on the developers' own hardware and has not been verified outside the team.

Plenty of open TTS projects have inherited a backbone language model's tokenizer without revisiting the choice. TontaubeV1's pitch is that the tokenizer is a design decision in its own right, and the open-weight space will start to test that decision once listening comparisons can tell the choices apart. Even one contrarian release makes the prevailing option visible as a choice rather than a given.

The bet is testable. If character-level tokenization actually does make long-form TTS more stable than the inherited-BPE default, head-to-head listening tests on audiobook-length passages should show fewer mid-sentence artifacts, and replication of the chunking-into-position scheme should port cleanly to other backbones. If they don't, the inherited-tokenizer pattern will keep its hold on open TTS, and TontaubeV1 will read more like an interesting detour than the start of a new default. Either answer is a data point: the open releases worth watching in 2026 are the ones that test the move rather than copy it.