Whisper large-v3 + SPC (Soft Prefix Conditioning)

Fine-tuned from Shagufta/whisper-largev3-km-indomain20-withac with Soft Prefix Conditioning for trilingual Kreol Morisien / French / English code-switching ASR.

Architecture

Single utterance-level soft LID distribution, injected as a bias into the encoder output via a self-removing forward hook, once per utterance before decoding begins. Standard (non-language-specific) LoRA on the top 8 decoder cross-attention layers.

Parameters

Parameter Value
Base model Shagufta/whisper-largev3-km-indomain20-withac
LoRA rank 16
LoRA alpha 32
LoRA target cross-attention only, top 8 decoder layers
Learning rate 1e-05
Train utterances 3670
Val utterances 94
Best val WER 0.0114

Citation

PhD thesis: Soft Language Identity Modelling for Creole Code-Switching ASR

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support