Modelpublic

Iscte-Sintra/KeaASR-CTC-1B_v2

sourceHugging Facecc-by-nc-4.0updated 4mo agoView on Hugging Face

0likes

Model Card

KEA Omnilingual-ASR — Kabuverdianu Fine-Tune

Fine-tuned version of omniASR_CTC_300M for Kabuverdianu (kea) — Cape Verdean Creole, a low-resource language spoken by ~1 million people.

Model Details

Property	Value
Base model	omniASRCTC300M (Meta, ~300M parameters)
Language	Kabuverdianu (`kea_Latn`)
Task	Automatic Speech Recognition (CTC)
Framework	omnilingual-asr (fairseq2)
Dataset	shunyalabs/kabuverdianu-speech-dataset

Training Data

Split	Samples
Train	2715
Validation	366
Test	864

Training Hyperparameters

Parameter	Value
Training steps	N/A
Learning rate	N/A
Mixed precision	N/A
Gradient accumulation	N/A batches
Encoder frozen for	N/A steps
Min audio length	N/A samples (2 s)
Max audio length	N/A samples (60 s)
Max elements per batch	N/A

Evaluation Results

Metric	Value
WER (validation)	N/A
Train CTC Loss	N/A
Validation CTC Loss	N/A

Checkpoint Structure

checkpoint/sdp_00.pt   ← model weights (fairseq2 sharded format)
model.yaml             ← architecture definition
config.yaml            ← full training configuration
metrics/train.jsonl    ← per-step training metrics
metrics/valid.jsonl    ← per-step validation metrics (WER)