Python 3.5
- Separate the each audio file into segments (one per conversation turn) using transcription
- Use force alignment tool to create further segments (one per word)
- Using the most frequent words, choose the most important keywords to detect
- Train a classifier to detect keyword
- this repository is for 1~2.
From CHILDES Dataset
- Gleason (2~5 years old) - father/mother children conversation
- Weist (2~5 years old) - 6 kids. Little bit noisy but okay
- Van Houten (2-3 years old) - bit short, but okay
Which one to use? Documentation (v), Python (v), Easy to use (v), Need to test Different algorithms
- AENAS (DTW algorithm) : See
aeneas_test.ipynb
FAVE (HMM; HTK based)Montreal (HMM; Kaldi based)SPPAS (HMM; Julius based)