ATC TTS Voxtream Dataset

This dataset is prepared for training the Voxtream TTS model and follows the format of herimor/voxtream-train-9k.

Source: Based on the Singaporean SG Eleven dataset: aether-raid/sg-aviation-el-combined

Description:

All features are provided as single batched .npy files:
- mimi_codes_16cb.npy: Mimi codec tokens (16 codebooks)
- phone_emb_indices.npy: Alignment of phoneme tokens to Mimi frames
- phone_tokens.npy: Phoneme tokens
- sem_label_shifts.npy: Monotonic phoneme alignment labels
- spk_templates.npy: Speaker templates (first 3 seconds, if available)
uids.npy contains the list of utterance IDs, matching the batch dimension of all features.

Usage: See herimor/voxtream-train-9k for loading and training examples.

License: See original dataset for licensing and usage restrictions.