melvinemmanuel001/whisper-naija-pidgin-lora

Whisper Nigerian Pidgin ASR — LoRA Adapter

A LoRA adapter that fine-tunes `NCAIR1/NigerianAccentedEnglish` for Nigerian Pidgin automatic speech recognition. The base model handles Nigerian-accented English; this adapter teaches it the structural particles of Pidgin that the base model mistranscribes as English equivalents.

Results

Evaluated on the 892-example test split of `asr-nigerian-pidgin/nigerian-pidgin-1.0`:

Model : WER

NCAIR1/NigerianAccentedEnglish (baseline) | 66.22% NCAIR1 + this LoRA (checkpoint-500) | 28.83% Absolute improvement | -37.39 pp | Relative error reduction | ~56% |

This adapter is checkpoint-500, selected by a checkpoint sweep — it outperformed both checkpoint-400 (32.19%) and checkpoint-510 (31.72%). The regression at step 510 indicates the final training steps overfit; the best saved checkpoint is step 500.

Usage

from transformers import WhisperForConditionalGeneration, WhisperProcessor
from peft import PeftModel

BASE = "NCAIR1/NigerianAccentedEnglish"
processor = WhisperProcessor.from_pretrained(BASE)
base = WhisperForConditionalGeneration.from_pretrained(BASE, use_safetensors=False)
model = PeftModel.from_pretrained(base, "%s").merge_and_unload()

Limitations

• —Trained and evaluated on a single Pidgin corpus; performance on noisier real-world audio (market noise, phone-mic recordings, heavy code-switching) is untested.
• —The checkpoint sweep spanned only steps 400-510, so step 500 is the best saved checkpoint, not necessarily the global optimum of the training run.

Base model

Built on NCAIR1/NigerianAccentedEnglish by Awarri Technologies, in partnership with Nigeria's minister of information.