In order to create short utterances from audio recordings, I found two solutions:
However, I am struggling to find efficient solution because recordings may contain noise or music. Moreover, I have neither a trained model nor an annotated corpus to build a new model.
Any advice would be greatly appreciated.
Thanks in advance.
Dear Selma Kali Ali
You can Simply search for Audio Segmentation which there are plenty of resources available.
If I understand correctly you want to extract segments with speech and disregard areas without speech - silence, noise, music, etc.
You are right that one way to go is to properties (i.e., features) that identify speech, but energy is not certainly not such a property (unless you are looking for energy contours).
You could work with any of the well-known speech-specfic feature sets (e.g., LPC cepstrum), but it will require a lot of tweaking (how do you differentiate between speaking and singing?)
Another possibility is to forget about features and work directly on samples. You really don't need a prepared dataset - just take a few recordings of speech and a few of music and mix segments together. You know where the speech is, so you can train. Interesting how much data you will need to make that work...
Dear Yaakov,
maybe You could try Audacity od Praat - these are programs You can easily download on your computer.
Hope it was helpful.
- Badges
- Science topic
- Similar topics
- Education