Researchers present TontaubeV1, a text-to-speech model that preserves natural prosody while enabling streaming inference from a single consumer GPU. The system uses a hierarchical DualCodec representation at 12.5 Hz to separate semantic streams from acoustic refinements, allowing causal decoding despite the noncausal nature of the underlying codec.

  • The architecture employs four Qwen3-derived transformers totaling 2.9B parameters: one predicts the semantic stream and utterance duration, while three progressively smaller models add acoustic refinements.
  • Text is tokenized per character, with paired text and audio markers supporting long-form generation through bounded context.
  • On a single RTX 5090, the streaming path achieves approximately 200 ms to first audio, with an end-to-end real-time factor of 0.08 for one input.
  • The model matches ElevenLabs Flash v2.5 and outperforms Fish Audio S2 Pro, Gradium API, and Cartesia Sonic 3 on prosody in LLM-as-a-judge evaluations.

The release provides a solution that balances high perceptual quality with low latency, making it suitable for real-time applications on accessible hardware.