facebook/mms-300m fine-tuned (character CTC) on YECS Yoruba-English code-switching, plain transcription. Control for the language-tag-injection A/B (tagged twin: LyngualLabs/yecs-asr-mms-lid, which adds 99.41% per-word LID).

Results (YECS full held-out test, 9,949 utts)

Metric	This model	mms-300m-yoruba-english (reported)
WER tone-aware	19.91%	29.11%
WER tone-insensitive	17.00%	—
CER tone-aware	6.23%	—

A 5-epoch fp16 fine-tune brings MMS-300m well below the previously reported baseline.

Usage

python

from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import torch, librosa
repo = "LyngualLabs/yecs-asr-mms-plain"
proc = Wav2Vec2Processor.from_pretrained(repo); proc.tokenizer.set_target_lang("yor")
model = Wav2Vec2ForCTC.from_pretrained(repo).to("cuda").eval()
a,_ = librosa.load("utt.wav", sr=16000)
iv = proc(a, sampling_rate=16000, return_tensors="pt").input_values.to("cuda")
print(proc.batch_decode(torch.argmax(model(iv).logits, dim=-1))[0])

Training

facebook/mms-300m SSL -> fresh CTC head; 5 epochs, lr 1e-4, fp16, dropout 0.05, warmup 500, batch 32, 1xH200. W&B run mms-plain. Code: https://github.com/osinkolu/yecs-asr-benchmark (tag-injection/).

Evaluation note

Scored on the full held-out YECS test split (9,949 utts); test audio is the canonical Mozilla/MDC release. Test reference labels are held privately by LyngualLabs (public test_metadata.csv has targets stripped), so metrics are internal and not independently reproducible.