Sequence Labeling

Description

use TIMIT dataset to predict phoneme sequences using provided mfcc or fbank features

Project Link

Requirements

• keras
• tensorflow
• python3
• h5py
• sklearn

Dataset

• TIMIT Dataset
• Features: mfcc and fbank
• Labels: 48 kinds of phones

Pre-Processing

Label Preprocessing

1. phone mapping 48 -> 39
2. converting sequences to one hot encodings
3. padding

Features Preprocessing

1. standardization
2. padding

Post-Processing

1. convert phoneme to alphabet
2. remove consecutive duplicates using a threshold
3. trim the 'sil' character

Results