Yemberzal — Kashmiri SraVaani ASR (Epoch 28)

Yemberzal is a Kashmiri automatic speech-recognition checkpoint, full fine-tuned from SraVaani. This model is the epoch-28 checkpoint used as the primary Kashmiri ASR model in the associated data-collection service.

Model lineage and details

• Base model: ARTPARK-IISc/SraVaani-1.0
• Adaptation: full fine-tune for Kashmiri ASR
• File: best_epoch28.nemo
• Framework: NVIDIA NeMo EncDecHybridRNNTCTCBPEModel
• Primary decoding path: TDT / RNN-T
• Recommended runtime: Python 3.11 and nemo_toolkit[asr]==2.7.3

Training and validation

This is the epoch-28 checkpoint from a full fine-tune on 45.0 hours of Kashmiri IndicVoices-R data. The experiment recorded a validation WER of 37.83% at epoch 28. The available report does not state the validation split's number of utterances or duration, so those values are intentionally not inferred here.

Held-out test set

Evaluation uses the official IndicVoices-R Kashmiri test split, after the configured 0.5–30 second clip-duration filter and transcript deduplication:

Test-set property	Value
Utterances / clips	403
Duration	0.9043 hours (about 54.3 minutes)
Speakers	16
Reference words	9,357
Reference characters	49,326

Test results

Metrics are corpus-level micro averages. Use the TDT / RNN-T decoder for this checkpoint.

Decoder	WER	95% bootstrap WER CI	CER	Exact utterances
TDT / RNN-T (primary)	38.13%	36.57–39.69%	12.74%	10 / 403
CTC (diagnostic)	42.60%	40.93–44.25%	14.95%	8 / 403

Loading locally

from nemo.collections.asr.models import EncDecHybridRNNTCTCBPEModel

model = EncDecHybridRNNTCTCBPEModel.restore_from("best_epoch28.nemo")
model.change_decoding_strategy(decoder_type="rnnt")
model.eval()
text = model.transcribe(["sample.wav"], batch_size=1, verbose=False)[0]

Use and limitations

This research/deployment checkpoint is for the associated data-collection workflow. The held-out test set is relatively small and does not substitute for native-speaker acceptance review. Validate transcripts—particularly proper nouns, code-switching, short utterances, and noisy speech—with a fluent Kashmiri reviewer before treating them as ground truth. Before redistribution or commercial use, review the base model and training-data terms that apply to your use case.