---
language:
- ar
task_categories:
- text-to-speech
pretty_name: Mimir Auto Tokenized SNAC
---

# Mimir Auto Tokenized SNAC

Tokenized TTS training dataset derived from the paired audio/text dataset.

Rows have exactly three columns:

- `inputs`: single token stream containing byte text, special markers, and flattened SNAC tokens
- `labels`: same length as `inputs`, with `-100` for text/prompt positions
- `attentionmask`: sequence mask, `1` for valid positions