File size: 3,256 Bytes
aa50616
 
 
 
 
 
 
 
 
 
 
 
 
 
 
03b27b4
 
a28c22a
40b9d12
 
 
03b27b4
 
aa50616
 
03b27b4
0d9e3a1
aa50616
 
 
 
 
 
 
 
 
 
 
 
 
0d9e3a1
aa50616
 
 
 
 
0d9e3a1
 
 
 
 
 
 
 
aa50616
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
---
license: openrail
language:
- en
metrics:
- f1
- recall
- accuracy
library_name: speechbrain
pipeline_tag: audio-classification
---
# Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->

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](https://boulderlearning.com/products/myst/) 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](https://phonbank.talkbank.org/access/Eng-NA/Providence.html) 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](https://openaccess.thecvf.com/content_cvpr_2013/html/Rehg_Decoding_Childrens_Social_2013_CVPR_paper.html) and [BabbleCor](https://osf.io/rz4tx/). 

## Model Sources
For more information regarding this model, please checkout our paper: (TO-DO)
- **Paper:** https://arxiv.org/pdf/2309.07287.pdf
  
## Model Description

<!-- Provide a longer summary of what this model is. -->
Folder contains the best checkpoint of the following setting
- **W2V2-MyST by fine-tuning on Librispeech 960h**: save_960h/wav2vec2.ckpt
- **W2V2-Pro trained on phone sequence**: save_MyST_Providence_ep45_filtered/wav2vec2.ckpt
- **W2V2-Pro trained on consonant/vowel sequence**: save_MyST_Providence_ep45_filtered_cv_only/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 there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
If you found this model helpful to you, please cite us as
<pre><code>
@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}
}
</code></pre>

# 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/