Praha-Labs/RIMA-TTS-v2-8Lang

Core 8-Language 45-Hour Run

This run trains one LoRA adapter for eight core Indic languages using one pass over substantially more unique audio than RIMA-TTS v1.

Languages:

hi bn mr gu ta te ml kn

Dataset plan:

• 45 hours per language
• 22.5 hours female and 22.5 hours male per language
• 360 unique hours total
• audio duration filter: 1-15 seconds
• one training epoch

Training configuration:

Setting	Value
Batch size	24
Gradient accumulation	1
Learning rate	4e-5
Epochs	1
LoRA rank / alpha	128 / 256
Checkpoint interval	1,000 steps
Checkpoints retained	6
Eval samples	4 per language every 1,000 steps

Run from the repository root:

export HF_TOKEN="hf_your_token_here"
bash IndicFinetuning/multilingual_indic/core_8lang_45h/prepare_dataset.sh
bash IndicFinetuning/multilingual_indic/core_8lang_45h/train.sh

The dataset is written under core_8lang_45h/dataset. Training outputs and checkpoint listening samples are written under IndicFinetuning/outputs/multilingual_indic_core_8lang_45h.

This run starts from the Chatterbox base model. It does not continue the RIMA-TTS v1 adapter, keeping the comparison with the earlier 12-language run clean.