This adapter targets Korean speech recognition across dysarthric and typical speech. It was trained on the PL-HFT combined corpus. The release contains the LoRA weights and the native v2 tokenizer.
omniASR_CTC_1B_v2, the fairseq2 encoder with a CTC output head. Use the official v2 checkpoint via the native model card.q_proj, k_proj, v_proj, output_proj and inner_proj.FEATURE_EXTRACTION is the wrapper option; transcription uses the native CTC decoder.Use Python 3.12 and install the dependencies below in an isolated environment. The included requirements target the CPU runtime. For CUDA, use the matching fairseq2 and PyTorch wheels. The first load needs access to the official base checkpoint.
python -m pip install "huggingface-hub>=0.34,<1"
hf download PL-HFT/omniasr-ctc-1b-v2-ko-dysarthria-combine-lora --local-dir ./omniasr-combine-lora
cd ./omniasr-combine-lora
python -m pip install -r requirements.txt
Run the following from the adapter directory. The packaged loader calls PeftModel.from_pretrained on the official fairseq2 model. It uses the included tokenizer and the native audio processing and CTC decoding pipeline.
from pathlib import Path
import torch
from load_omniasr import load_pipeline
pipeline, adapter = load_pipeline(Path("."), device="cpu", dtype=torch.float32)
with torch.inference_mode():
print(pipeline.transcribe(["sample.wav"], batch_size=1)[0])
Use short audio clips. Training used 16 kHz mono segments up to 29.5 seconds. The bundled tokenizer is a native SentencePiece model and the processor is provided by omnilingual-asr.
combine Dataset)The combined Korean profile uses dysarthric speech from AI-Hub 013 and typical speech from AI-Hub 109. These are counts from the run records and the held-out evaluation. The repository contains adapter artifacts and aggregate metrics.
| Split | Utterances | Audio hours |
|---|---|---|
| Train | 321,135 | 514.37 |
| Validation | 38,459 | 64.77 |
| Internal test | 38,599 | N/A |
The trainer retained all prepared train and validation segments. It added minimal right zero padding to 40 train and 3 validation segments to meet the CTC frontend minimum. The test split was held out during training and selection.
20261020T163744095984Z on NVIDIA A100-SXM4-80GB.Scores are from archived Korean evaluations of the native checkpoint identified by the source adapter SHA-256. The scorer uses normalized Korean text. Each error rate divides the sum of edit distances by the sum of the corresponding reference lengths. The combined internal row pools both internal subsets by their reference lengths.
N = 38,599)| Condition | CER | WER |
|---|---|---|
| Base | 11.76% | 37.02% |
| Combine LoRA (This Model) | 4.30% | 15.69% |
| Evaluation set | Utterances | Base CER / WER | Combine LoRA CER / WER |
|---|---|---|---|
| Dysarthric speech (AI-Hub 013) | 15,495 | 11.78% / 37.92% | 2.11% / 7.24% |
| Typical speech (AI-Hub 109) | 23,104 | 11.75% / 36.46% | 5.70% / 20.96% |
| External Korean lectures (AI-Hub 71627) | 10,000 | 16.86% / 43.64% | 16.49% / 42.53% |
These benchmark results correspond to one training run. The packaged PEFT loader was checked against the native checkpoint using a local CPU smoke test. See the smoke-test receipt for the scope of that check.