Speaker_partition_module
Analysis of voices based on the Mel-frequency band.
Goal: Identification of voices speaking (diarization) and calculation of speech partition (in %).
Methodology:
- Collect voice data
- Sample audio data of x speakers that talk y times to represent a round of people talking
- Annotate samples with labels and merge audio file
- Create train & test split of samples
- Train unsupervised clustering module to detect number of people
- Train supervised RNN classifier to determine who is speaking at time x
Preprocessing
- Convert files to .wav convertFlac2Wav.py
- Collect data via LibriSpeech voices library (audiofiles) audio_manipulation02.py
- Extract x random speakers with y audio samples per speaker Result: Generated audio samples of length 30-60 seconds
Feature extraction:
- Create mel-frequency spectrum for each audio file feature_extraction.py
- Define overlapping feature window for training
Training:
- Implementation of google-diarizer module
- Training accuracy is only at 40 %
Further activity
- Create own unsupervised clustering module
- Try out different libraries
