AITRADER/dutch-tts-shards-v2

Dutch TTS Training Data (Multi-Speaker) v2

Pre-tokenized Dutch speech data for training Marvis-TTS models with correct tokenization and multi-speaker support.

⚠️ Important: v2 vs v1

This is v2 with critical fixes:

• ✅ Correct text tokenization: <|im_start|>[speaker]text<|im_end|>
• ✅ Multi-speaker support: Male (0), Female (1), Unknown (2)
• ✅ Speaker labels from source datasets

v1 (dutch-tts-shards) had broken tokenization that caused unintelligible output.

Speaker Mapping

Speaker ID	Gender	Source
0	Male	From gender labels or pseudo-random from speaker_id
1	Female	From gender labels or pseudo-random from speaker_id
2	Unknown	When gender info unavailable

Format

WebDataset shards (.tar) containing JSON with:

• text: Original text
• text_tokens: Tokenized as <|im_start|>[speaker]text<|im_end|>
• audio_tokens: 32 codebooks from Mimi codec (24kHz)
• speaker: 0 (male), 1 (female), 2 (unknown)
• dataset: Source dataset name

Source Datasets

• Mozilla Common Voice 17 (Dutch) - with gender labels ✓
• Google FLEURS (Dutch) - with gender labels ✓
• CML-TTS (Dutch) - speaker_id based
• MLS Dutch - speaker_id based
• Natural Accented Dutch
• CSS10 Dutch - single female speaker ✓

Usage

from datasets import load_dataset, Audio

ds = load_dataset("AITRADER/dutch-tts-shards-v2", streaming=True, split="train")
ds = ds.cast_column("audio.wav", Audio(decode=False))

for sample in ds:
    j = sample["json"]
    text_tokens = j["text_tokens"]    # Ready for model input
    audio_tokens = j["audio_tokens"]  # Ready for model output
    speaker = j["speaker"]            # 0=male, 1=female, 2=unknown

Training

python train.py config_nl_60m_cloud.json

Make sure your config uses "dataset_repo_id": "AITRADER/dutch-tts-shards-v2".

Speaker Distribution

See speaker_stats.json for detailed statistics per dataset.