ACI Yoruba Phase 9
Self-contained bundle: Phase 8's acoustic weights paired with a 5-gram KenLM decoder (alpha=0.5, beta=1.0) and a phonetic post-processing map for common ASR spelling errors.
Evaluation
Measured on a 55-sample held-out set (never used in any training for this model or its predecessors). WER reported with punctuation stripped symmetrically from both prediction and reference, since this model's vocabulary does not produce punctuation tokens.
| Model | WER (held-out, normalized) |
|---|---|
| Phase 9 | 12.63% |
| Phase 10 (fine-tuned acoustics) | 12.33% |
| Phase X (stronger decoder config) | 7.09% |
For production use, see Awesome-x/aci-yoruba-phaseX, which uses a stronger KenLM weighting and a larger unigram vocabulary for meaningfully better results on the same benchmark. Phase 9 is kept available for reproducibility and as the base this lineage was built from.
Quickstart
pip install torch transformers pyctcdecode librosa
import torch
import librosa
from transformers import Wav2Vec2ForCTC, Wav2Vec2ProcessorWithLM
REPO_ID = "Awesome-x/aci-yoruba-phase9"
model = Wav2Vec2ForCTC.from_pretrained(REPO_ID)
processor = Wav2Vec2ProcessorWithLM.from_pretrained(REPO_ID)
audio, sr = librosa.load("sample.wav", sr=16000)
input_values = processor(audio, sampling_rate=16000, return_tensors="pt").input_values
with torch.no_grad():
logits = model(input_values).logits
print(processor.decode(logits[0].numpy()).text)