shantipriya/odia-asr

Evaluation Results

• Latest checkpoint uploaded: best_model.pt (size: 368K)

• Training in progress: Currently at epoch 11 (see train.log for details)

• ASR WER (IWSLT 2026 test_set): 45.08%

• ASR CER (IWSLT 2026 test_set): 12.60%

Sample predictions and full results are in results_test_set.txt.

language: - or license: cc-by-4.0 tags: - automatic-speech-recognition - odia - conformer - ctc - ai4bharat datasets: - openslr - MUCS2021 metrics: - wer base_model: ai4bharat/indic-conformer-600m-multilingual

Odia ASR — Fine-tuned IndicConformer 600M

Fine-tuned version of ai4bharat/indic-conformer-600m-multilingual on Odia (ଓଡ଼ିଆ) speech recognition.

Model Details

Property	Value
Base Model	ai4bharat/indic-conformer-600m-multilingual
Language	Odia (or)
Task	Automatic Speech Recognition (ASR)
Architecture	Conformer + CTC head
Encoder Dim	1024
Vocabulary	91 Odia characters
Sample Rate	16 kHz

Training Data

• Train: OpenSLR-103 — MUCS 2021 Odia (~94.5h, 59,782 utterances)
• Test (OpenSLR): OpenSLR-103 test set (~5.5h, 3,471 utterances)
• Test (IWSLT 2026): OdiaGenAI/iwslt-odia-speech test set (~2.7h, 1,600 utterances)

Training Configuration

Hyperparameter	Value
Batch size	8 (effective 32 with grad_accum=4)
Learning rate	1e-4
Optimizer	AdamW (weight_decay=1e-4)
Scheduler	Cosine with 500 warmup steps
Max epochs	30
Max audio length	10s
GPU	NVIDIA A100 80GB

Architecture Notes

The base IndicConformer-600M model uses:

• Preprocessor: NeMo TorchScript mel-spectrogram extractor (80-dim, runs on CUDA)
• Encoder: ONNX InferenceSession (1024-dim output, CUDAExecutionProvider)
• CTC Head: nn.Linear(1024, 91) — the only fine-tuned parameters

The encoder is frozen (ONNX); only the Odia CTC projection head is trained.

Usage

from transformers import AutoModel
import torch

model = AutoModel.from_pretrained(
    "shantipriya/odia-asr",
    trust_remote_code=True,
)

Citation

If you use this model, please cite:

@misc{odia-asr-2026,
  title  = {Fine-tuned IndicConformer for Odia ASR},
  author = {Shantipriya Parida},
  year   = {2026},
  url    = {https://huggingface.co/shantipriya/odia-asr},
}