Character-Level TTS Model Released

TontaubeV1 is an 2.9B parameter open-weight text-to-speech (TTS) model released with optimizations for long-form generation and low-latency local inference. The model primarily supports English and German, offering zero-shot voice cloning via reference audio of up to 1 minute.

Character-Level Tokenization Applied

While most recent LLM-based TTS models use the BPE tokenizer of the backbone model, TontaubeV1 is based on the Qwen3-1.7B checkpoint but adopts a method that forces speech text to be tokenized as individual character sequences. The developers stated that this approach simplifies character-to-sound mapping while maintaining language understanding capabilities, thereby reducing out-of-distribution errors caused by rare token sequences.

Training Data and Architecture

The model is built on the DualCodec multi-codebook discrete audio codec and was trained on approximately 200,000 hours of audio data across 7 languages. Currently, testing and validation have been primarily conducted in English and German environments.