uctnlp/omniASR-CTC-300m-v2-Zulu

OmniASR CTC 300M v2 — isiZulu

This repository contains an OmniASR CTC 300M v2 model fine-tuned for isiZulu ASR.

Model details

• Repository: csikasote/omniASR-CTC-300m-v2-Zulu
• Framework: fairseq2
• Base model: facebook/omniASR-CTC-300M
• Language: isiZulu (zul_Latn)
• Best checkpoint: step 38,500
• Best validation WER: 27.2518
• Final training step: 40,000
• Final validation WER: 27.3033

Repository structure

.
├── README.md
├── omniASR-CTC-300M-v2-Zulu.pt
├── omniASR_tokenizer_written_v2.model
├── configs/
│   ├── ctc-finetune.yaml
│   └── model.yaml
└── training_checkpoint/
    └── step_38500/
        ├── model/
        ├── optimizer/
        ├── trainer/
        └── data_reader/

Native checkpoint format

This is a native OmniASR/fairseq2 checkpoint. It is not directly loadable with AutoModelForCTC.from_pretrained() without conversion.

Full training checkpoint

The complete checkpoint is included under:

training_checkpoint/step_38500/

It contains model, optimizer, trainer, and data-reader state.