Model
This model is Wav2Vec2-Large-XLSR-53 fine-tuned on the manually annotated subset of CMU's L2-Arctic dataset. It was fine-tuned to perform automatic phonetic transcriptions in IPA. It was tuned following a similar procedure as described by vitouphy with the TIMIT dataset.
Usage
To use the model, create a pipeline and invoke it with the path to your WAV, which must be sampled at 16KHz.
from transformers import pipeline
pipe = pipeline(model="mrrubino/wav2vec2-large-xlsr-53-l2-arctic-phoneme")
transcription = pipe("file.wav")["text"]
Results
The manually annotated subset of L2-Arctic was divided into training and testing datasets with a 90/10 split. The performance metrics for the testing dataset are included below.
WER - 0.425
CER - 0.128
Citation
If you find our model helpful, please feel free to cite us.
@article{Bo_Rubino_Xu_2024,
title={A Mispronunciation-Based Voice-Omics Representation Framework for Screening Specific Language Impairments in Children},
DOI={10.1109/ichi61247.2024.00045},
journal={2024 IEEE 12th International Conference on Healthcare Informatics (ICHI)},
author={Bo, Wei and Rubino, Matthew and Xu, Wenyao},
year={2024},
month={Jun},
pages={294–304}
}