CITATION — REFERENCE ENTRY
Chord Segmentation and Recognition using EM-Trained Hidden Markov Models — Proceedings of the 4th International Conference on Music Information Retrieval (ISMIR 2003)
- Key
- sheh2003-hmm-chords
- Authors
- Sheh, Alexander; Ellis, Daniel P. W.
- Issued
- 2003
- Type
- paper-conference
- Container
- Proceedings of the 4th International Conference on Music Information Retrieval (ISMIR 2003)
Raw CSL JSON
{
"URL": "http://ismir2003.ismir.net/papers/Sheh.pdf",
"type": "paper-conference",
"title": "Chord Segmentation and Recognition using EM-Trained Hidden Markov Models",
"author": [
{
"given": "Alexander",
"family": "Sheh"
},
{
"given": "Daniel P. W.",
"family": "Ellis"
}
],
"issued": {
"date-parts": [
[
2003
]
]
},
"event-place": "Baltimore, MD, USA",
"container-title": "Proceedings of the 4th International Conference on Music Information Retrieval (ISMIR 2003)"
}
Claims
-
Tested on a small set of 20 early Beatles songs, the system reached about 75% frame-level accuracy when aligning known chord sequences to the audio.
"show frame-level accuracy of around 75% on a forced-alignment task"
-
In Sheh and Ellis's experiments, pitch class profile features performed up to four times better than MFCC features.
"The strength of the PCP representation, and the model averaging approach, is clearly demonstrated, with the PCP ROT models performing as much a four times better than the best MFCC counterparts."
-
Sheh and Ellis judged their recognition accuracy not yet sufficient for usable transcriptions of unknown audio, though alignment of known progressions could be useful in itself.
"Although recognition accuracy is not yet sufficient to provide usable chord transcriptions of unknown audio, the ability to find time alignments for known chord progressions may be useful in itself."
-
Sheh and Ellis trained hidden Markov models using only the chord sequences of training songs, without the timings of chord changes, which were determined automatically during training.
"Crucially, this allows us to use as input only the chord sequences for our training examples, without requiring the precise timings of the chord changes"
Available in