A Santali speech-synthesis model for text written in the Ol Chiki script.
The model was trained using the Piper VITS training pipeline and exported to
ONNX for offline inference and deployment on supported platforms.
Model details
| Property |
Value |
| Language |
Santali |
| Language code |
sat |
| Script |
Ol Chiki |
| Architecture |
Piper VITS |
| Piper version |
1.8.0 |
| Sample rate |
16 kHz |
| Speakers |
1 |
| Training utterances |
1,200 |
| Training epochs |
50 |
| Batch size |
16 |
| Training hardware |
2 × NVIDIA Tesla T4 |
| Trainable parameters |
~70.4M |
| Final training epoch |
49 |
| Reported validation MOS |
2.1369 |
Files
santali_piper_vits.onnx
The exported ONNX inference model.
This is the primary deployment artifact.
santali_piper_vits.onnx.json
The corresponding Piper model configuration.
config.json
Training configuration containing the model and text-symbol configuration.
phonemes.json
The custom text-symbol inventory used for Santali written in Ol Chiki.
metadata.csv
Training metadata for the speech samples used in this training run.
MODEL_MANIFEST.json
Machine-readable metadata describing the model, training run, dataset, and validation results.
Training checkpoint
The PyTorch Lightning checkpoint is retained separately for reproducibility and
possible future model work.
Training data
The model was trained using a selected Santali speaker subset from the
AI4Bharat Rasa dataset.
Dataset:
https://huggingface.co/datasets/ai4bharat/Rasa
The upstream dataset should be obtained directly from its original repository.
Please refer to the dataset repository for its current license, attribution,
and usage requirements.
Model architecture
The model uses the Piper VITS architecture with the following configuration:
- ResBlock:
2
- ResBlock kernel sizes:
[3, 5, 7]
- ResBlock dilation sizes:
[[1, 2], [2, 6], [3, 12]]
- Upsample rates:
[8, 8, 4]
- Upsample initial channel:
256
- Upsample kernel sizes:
[16, 16, 8]
- Filter length:
1024
- Hop length:
256
- Window length:
1024
- Segment size:
8192
A custom text-symbol representation was used for Santali/Ol Chiki rather than
an English phoneme inventory.
Inference
The exported model is intended to be used with Piper-compatible inference
software.
The ONNX model accepts:
- tokenized text input
- input sequence lengths
- synthesis scales
and produces waveform audio.
Validation
The exported model was checked using ONNX's structural validator and loaded
successfully using ONNX Runtime.
- ONNX checker: passed
- ONNX Runtime: 1.30.0
- Runtime provider tested: CPU
- Piper end-to-end synthesis: passed
Intended use
This model is intended for:
- Santali text-to-speech
- Offline speech synthesis
- Low-resource language research
- Educational applications
- Accessibility applications
- Edge and on-device inference
- Research and experimentation
Limitations
This is a low-resource, single-speaker speech-synthesis model trained on
approximately 1,200 utterances.
Possible limitations include:
- pronunciation errors for unseen words
- reduced quality on text outside the training distribution
- sensitivity to text normalization
- limited speaker diversity
- reduced naturalness compared with larger speech datasets
- occasional pronunciation or prosody errors
The reported validation MOS is a training-time metric and should not be
interpreted as an independent standardized human evaluation.
Reproducibility
The repository includes the model configuration, custom text-symbol mapping,
training metadata, and trained model artifacts needed to document this
training run.
The original speech dataset is not redistributed here.
Attribution
This model uses the AI4Bharat Rasa dataset.
Please retain the attribution and comply with the applicable license terms of
the upstream dataset when using or redistributing derivatives.
Dataset:
https://huggingface.co/datasets/ai4bharat/Rasa
Piper:
https://github.com/OHF-Voice/piper1-gpl
License
The repository metadata uses Apache-2.0 for the model artifact unless otherwise
specified by the upstream dataset or dependency licenses.
Users are responsible for complying with the licenses and attribution
requirements of the underlying training data and software dependencies.
