Supernova11c/Supernova-TeraiASR

Supernova TeraiASR v0.2

An experimental Nepali automatic speech recognition research model by Supernova11c.

Overview

Supernova TeraiASR is a character-level Nepali speech recognition experiment implemented in PyTorch. It uses 16 kHz audio features represented as 80-bin log-mel spectrograms and a CNN/BiLSTM/CTC-style recognition pipeline.

This repository contains the model state dictionary, character vocabulary, and configuration metadata. The model's original preprocessing, architecture and decoding implementation may require the matching training notebook.

Current evaluation

Reported metrics from the development process:

Metric	Result
Best reported validation loss	1.6994
Test character error rate (CER)	45.35%
Test word error rate (WER)	92.17%
Test samples	701

A later character-level diagnostic measured CER at 45.33% using the original reference strings. Small differences can arise from normalization and evaluation implementation.

These results indicate substantial recognition errors. This checkpoint is experimental and not production-ready.

Files

• model_state_dict.pt — PyTorch model parameters
• vocab.json — character-to-ID mapping
• config.json — model metadata and token IDs
• model_class_source.py — model class source, if extractable

Input and preprocessing

The experiment used 16 kHz audio and 80-bin log-mel features. Correct inference requires the exact feature extraction settings, tensor layout, model architecture, output-length calculation, and CTC decoding behavior used during training. The checkpoint alone is not a complete inference pipeline.

Limitations

• High character and word error rates
• Frequent missing characters and words
• Not validated for real-world or safety-critical transcription
• More training, evaluation, and inference-pipeline packaging are needed

Intended use

Research, educational experimentation, and further development of Nepali speech recognition.

License

supernova-horizon, subject to verification that the code, data-derived artifacts, and included materials are compatible with this license. Dataset terms may differ from this repository's license.