Open-source TontaubeV1 TTS Model Uses Character-Level Tokenization
EAVDR · reddit · 2026-09-01
TontaubeV1 is a 2.9B-parameter open-weight TTS model built on Qwen3-1.7B, focusing on expressive speech, long-form narration, and low-latency local inference. It supports English and German with zero-shot voice cloning.
The release highlights two key design choices:
- Character-level Tokenization: Forces the model to process text as a sequence of individual characters instead of BPE tokens. This reduces out-of-distribution errors and rare token sequences while retaining language understanding capabilities.
- Chunking and Position Scheme: Processes text, semantic audio, and acoustic codebooks as parallel rows in a flat sequence. It assigns logical position IDs to align tokens representing the same audio frame across codebooks, using paired split markers and a sliding window to handle long-form context without leaking positions.
More from Models
- MOSS SFT vs Flux 3 Klein vs SD 3.0: Same-Prompt Comparison Exposes SD3's Age — ZootAllures9111 · 2026-10-18
- Quick take: GPT-6 matches Opus 5.5 at coding, far ahead at computer use — _vztu · 2026-10-18
- Open models now above Sol-level, outpacing expectations as community awaits V4.1 Pro — yacineMTB · 2026-10-18
- Codex took 3 days, Opus 5.5 nailed it in an afternoon — banned user's video protest — dotey · 2026-10-18
- Yacine MTB: Users will remember how Anthropic treated them once intelligence proliferates — yacineMTB · 2026-10-18
- Mistral's Timestral forecasting models start at 5M params and run on plain CPUs — lmoroney · 2026-10-18