Model Card for Model ID

We build a CTC-based phoneme recognition model using wav2vec 2.0 (W2V2) for children under 4-year-old. We use three-level fine-tuning to gradually reduce age mismatch between adult phonetics to child phonetics.

W2V2-Libri100h: We first fine-tune W2V2-Base using 100 hours of LibriSpeech pretrained on unlabeled 960 hours LibriSpeech adult speech corpus with IPA phone sequences.
W2V2-MyST: We then fine-tune W2V2-Libri100h using My Science Tutor corpus (consists of conversational speech of students between the third and fifth grades with a virtual tutor).
W2V2-Libri100h-Pro (two-level fine-tuning): We fine-tune W2V2-Libri100h using Providence corpus (consists of longititude audio of 6 English-speaking children aged from 1-4 years interacting with their mothers at home) on phoneme sequences.
W2V2-MyST-Pro (three-level fine-tuning): Similar as W2V2-Libri100h-Pro, we fine-tune W2V2-MyST using Providence on phoneme sequences.

We show W2V2-MyST-Pro is helpful for improving children's vocalization classification task on two corpus, including Rapid-ABC and BabbleCor.

Model Sources

For more information regarding this model, please checkout our paper:

Paper: https://arxiv.org/pdf/2309.07287.pdf

Model Description

Folder contains the best checkpoint of the following setting

W2V2-Libri100h: save_100h/wav2vec2.ckpt
W2V2-MyST: save_100h_MyST/wav2vec2.ckpt
W2V2-Libri100h-Pro: save_100h_Providence/wav2vec2.ckpt
W2V2-MyST-Pro: save_100h_MyST_Providence/wav2vec2.ckpt

Uses

We develop our complete fine-tuning recipe using SpeechBrain toolkit available at

https://github.com/jialuli3/speechbrain/tree/infant-voc-classification/recipes/RABC (used for Rapid-ABC corpus)
https://github.com/jialuli3/speechbrain/tree/infant-voc-classification/recipes/Babblecor (used for BabbleCor corpus)

Paper/BibTex Citation

If you found this model helpful to you, please cite us as


@article{li2023enhancing,
  title={Enhancing Child Vocalization Classification in Multi-Channel Child-Adult Conversations Through Wav2vec2 Children ASR Features},
  author={Li, Jialu and Hasegawa-Johnson, Mark and Karahalios, Karrie},
  journal={arXiv preprint arXiv:2309.07287},
  year={2023}
}

Model Card Contact

Jialu Li (she, her, hers)

Ph.D candidate @ Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign

E-mail: jialuli3@illinois.edu

Homepage: https://sites.google.com/view/jialuli/

lijialudew
/

wav2vec_children_ASR