Training

Training Vivi

At the beginning, Vivi knows nothing about violin sounds – she needs a human to teach her. The human helps Vivi to learn by pointing out good and bad sounds. How important is training? Well, take a listen:

Example of training scales

Number of audio snippets trained

Approximate time spent

Audio examples

32

5 minutes (basic training only)

separate scale mp3 slurred scale mp3

94

15 minutes

separate scale mp3 slurred scale mp3

203

40 minutes

separate scale mp3 slurred scale mp3

All sound analysis and machine learning (training and classifying) is done with Marsyas.

User: basic training

Vivi asks you to classify short (approximately 0.5 seconds) bits of audio. These are created without any attacks, and are intended to test her "steady" sounds.

The sounds are classified as follows:

1. Much more force

mp3

almost silent!

2. More force

mp3

3. Ok sound

mp3

"ok" in comparison to the other sounds

4. Less force

mp3

5. Much less force

mp3

not a violin?

Automatic: calculating parameters

Once Vivi knows what a "good steady sound" should be, she tries to figure out how to generate those sounds.

  • if she hears a bad sound, how much should she modify the bow force?
  • how much bow force should she use before the bow is moving?

User: practice scales

Once Vivi has some basic knowledge and some bow forces, it’s time to help her practice scales!

Select a scale (and dynamic) to practice, then listen to Vivi’s performance. If any note sound questionable, you can examine that note and identify any areas in which Vivi’s judgement was incorrect.

Screenshot of interactive
training