kapilkarda/indic-multilingual-tts-v2

Indic Multilingual TTS v2

Combined dataset for training multilingual Indian-language TTS models, specifically prepared for Spark-TTS BiCodec and LLM fine-tuning.

Dataset Summary

Metric	Value
Total samples	556,524 (550K train + 6.5K val)
Total duration	~1,251 hours (1,237h train + 14h val)
Languages	13 (12 Indic + Indian English)
Sources	IndicVoices-R, Rasa, IndicTTS
Audio format	WAV, 16kHz mono
Emotion labels	6 emotions + neutral + domain tags

Languages

Language	Code	Train Samples	Source(s)
Assamese	as	65,733	IVR + Rasa
Tamil	ta	61,370	IVR + Rasa
Telugu	te	56,081	IVR + Rasa
Bengali	bn	53,936	IVR + Rasa
Malayalam	ml	51,338	IVR + Rasa
Nepali	ne	49,990	IVR + Rasa
Punjabi	pa	40,123	IVR + Rasa
Odia	or	36,759	IVR + Rasa
Kannada	kn	36,363	IVR + Rasa
Marathi	mr	35,521	IVR + Rasa
Hindi	hi	30,338	IVR + Rasa
Gujarati	gu	19,857	IVR + Rasa
English (Indian)	en	12,615	IndicTTS

Data Sources

• IndicVoices-R (281K samples): Large-scale crowdsourced Indian language speech corpus
• Rasa (256K samples): Emotion-labeled conversational speech in Indian languages
• IndicTTS (12.6K samples): High-quality Indian English TTS data from NPTEL

Emotion Labels

From Rasa corpus, 6 real emotions are labeled:

Emotion	Samples
anger	8,914
happy	8,779
fear	8,086
surprise	8,045
sad	7,956
disgust	7,644
neutral	293,857

Domain tags (conv, wiki, book, news, etc.) are also preserved in the emotion field for non-emotion Rasa samples.

File Format

Metadata CSVs use pipe (|) delimiter with columns:

path|text|lang|emotion|duration|utt_id|speaker_id|source

• train.csv — 550,024 training samples
• val.csv — 6,500 validation samples (500 per language, stratified)

Audio files are organized under audio/ by source:

audio/
  hi/          # Rasa Hindi
  ta/          # Rasa Tamil
  ...
  ivr_hi_v2/   # IndicVoices-R Hindi
  ivr_ta_v2/   # IndicVoices-R Tamil
  ...
  rasa_hi_v2/  # Rasa Hindi (v2 pipeline)
  ...
  indictts_en/ # IndicTTS English

Usage

For Spark-TTS BiCodec fine-tuning

python scripts/training/train_bicodec.py \
    --config scripts/training/configs/bicodec-multilingual.yaml

For Spark-TTS LLM emotion fine-tuning

First extract emotion training data in CoT format:

python scripts/data/extract-emotion-data.py \
    --data_dir data/multilingual_v2 \
    --model_dir pretrained_models/Spark-TTS-0.5B \
    --output_dir data/emotion_training

Then fine-tune:

source venv/bin/activate
python scripts/training/train_emotion_finetune.py \
    --config scripts/training/configs/emotion-finetune.yaml

Preparation Pipeline

Data was prepared using scripts/data/prepare-bicodec-data.py which:

1. Downloads IndicVoices-R from HuggingFace
2. Downloads Rasa dataset
3. Downloads IndicTTS (Indian English from NPTEL)
4. Resamples all audio to 16kHz mono WAV
5. Filters by duration (0.5s-30s)
6. Deduplicates by utterance ID
7. Creates stratified train/val split (500 per language for val)

License

This dataset combines data from multiple sources. Please refer to the original dataset licenses:

• IndicVoices-R: CC-BY-4.0
• Rasa: CC-BY-4.0
• IndicTTS: Check original source terms