Whisper Largev3 Myst Gclora

--- license: apache-2.0 language: en library_name: transformers pipeline_tag: automatic-speech-recognition tags: - whisper - automatic-speech-recognition - peft - lora - gc-lora - myst --- # Whisper-large-v3 MyST (GC-LoRA) Whisper-large-v3 adapted to the **MyST** dataset with **GC-LoRA** (Gated Convolutional LoRA), from the paper *"GC-LoRA: Gated Convolutional LoRA for Parameter-Efficient Acoustic Adaptation"* (Interspeech 2026). - **Base model:** `openai/whisper-large-v3` - **Method:** GC-LoRA adapter on the encoder attention output projections (rank 8, kernel 31, scaling 16) - **MyST test WER:** **8.7%** - **Code:** https://github.com/balaji1312/gc_lora ## Usage The checkpoint bundles the frozen Whisper backbone together with the trained GC-LoRA adapter. Loading requires the custom modeling code in the [gc_lora repository](https://github.com/balaji1312/gc_lora); see `src/bin/decode_asr.py` there for loading and decoding. ## Citation ```bibtex @inproceedings{shankar2026gclora, author = {Shankar, Natarajan Balaji and Wang, Zilai and Zhang, Kaiyuan and Shi, Mohan and Alwan, Abeer}, title = {{GC-LoRA}: Gated Convolutional {LoRA} for Parameter-Efficient Acoustic Adaptation}, booktitle = {Interspeech 2026}, year = {2026}, } ```