---
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`: 8 streams `[text_stream, audio_codebook_0, ..., audio_codebook_6]`
- `labels`: same shape as `inputs`, with `-100` for text/prompt positions and real targets only for audio tokens
- `attentionmask`: sequence mask, `1` for valid positions