TontaubeV1 Verbalizer

The TontaubeV1 Verbalizer is an optional English text-normalization model for the TontaubeV1 text-to-speech system. It converts written forms such as numbers, dates, times, currencies, symbols, abbreviations, and initialisms into text better suited for speech synthesis while aiming to preserve the original wording.

It is a separately trained Qwen3-1.7B-derived causal model and is not part of the four-codebook TTS graph. Numeric input uses digit-split tokenization in the release runtime. Inference is deterministic and uses vLLM n-gram prompt-lookup speculative decoding.

Technical reportTry TontaubeV1 onlineTontaubeV1 modelInference repository

Use with TontaubeV1

Enable the verbalizer when starting the inference server:

ENABLE_VERBALIZATION=1 uv run tontaube serve

When verbalization is enabled, the pinned snapshot is downloaded into the standard Hugging Face cache on the first server startup. The verbalizer is English-only and disabled by default. Requests opt in with "use_verbalization": true; using this option for another language returns an error. Requests that do not opt in bypass the verbalizer.

Limitations

Text normalization is generative and can produce incorrect expansions or alter wording. Applications requiring exact wording, legal or medical accuracy, or reliable handling of identifiers should validate its output or disable it. It is not intended as a general conversational language model.

Licence and attribution

This checkpoint is released under the Apache License 2.0. Third-party material remains subject to its respective terms. See Third-Party Notices.

The model is derived from Qwen3-1.7B, released under Apache License 2.0.

Downloads last month
-
Safetensors
Model size
2B params
Tensor type
BF16
Inference Providers NEW
This model isn't deployed by any Inference Provider. 馃檵 Ask for provider support

Model tree for TontaubeAI/TontaubeV1-Verbalizer

Finetuned
Qwen/Qwen3-1.7B
Finetuned
(1053)
this model