Structured Training for Large-Vocabulary Chord Recognition — Proceedings of the 18th International Society for Music Information Retrieval Conference (ISMIR 2017)

Claims

1. data-scarcity
   Large-vocabulary chord recognition must overcome data scarcity, because some chord classes occur much less often than others.

   "Automatic chord recognition systems operating in the large-vocabulary regime must overcome data scarcity: certain classes occur much less frequently than others, and this presents a significant challenge when estimating model parameters."

   Locator: page: 188 · Quote language: en
2. no-inversions
   McFee and Bello's model does not predict inversions.

   "the model does not predict inversions."

   Locator: page: 193 · Quote language: en
3. rare-classes
   The model was biased toward major and minor chords, which made up most of the training data, and performed poorly on the rarest class, minmaj7 (0.03% of the data).

   "The CR2+S+A still performs poorly on the rarest class minmaj7 (0.03% of the data)"

   Locator: page: 193 · Quote language: en
4. structured
   McFee and Bello's convolutional-recurrent model is trained to output both chord labels and binary encodings of chord roots and qualities.

   "we develop a deep convolutional-recurrent model for automatic chord recognition over a vocabulary of 170 classes. To exploit structural relationships between chord classes, the model is trained to produce both the time-varying chord label sequence as well as binary encodings of chord roots and qualities."

   Locator: page: 188 · Quote language: en
5. vocabulary
   The vocabulary has 170 classes: two special symbols and 168 combinations of 12 roots and 14 qualities.

   "The final vocabulary contains 170 classes: 2 special symbols (N, X), and 12 × 14 = 168 combinations of root and quality."

   Locator: page: 190 · Quote language: en