Transcription of Piano Music
Recently, there has been a lot of research on automatic transcription of piano music. However, it seems that current methods aren’t as effective as they should be. The state of the art methods have been shown to be insufficient for transcription in certain contexts, like noisy or reverberant environments. Here, we will look at two methods that use a score-informed transcription to solve this problem.
The first approach, the spectral sparse coding method, aims at recognizing and estimating note waveforms in piano recordings. This approach combines the use of heuristic rules to find note candidates from piano signal segments. Furthermore, the spectrogram can be decomposed into polyphonic notes by using a non-negative matrix factorization.
The second approach uses signal models and sub-systems to recognize psycho-acoustic clues and combine them to create the desired transcription. A novel algorithm is used to distinguish between the pitches of underlying notes and track them to produce a final score. This algorithm is based on multiresolution techniques, such as the Fourier transform and maximum likelihood frequency estimator, and shows superior performance compared to existing commercial software.
Automatic Transcription of Piano Music
This research uses a hidden Markov model to achieve accurate transcription of piano music. Unlike traditional hidden Markov model techniques, this method does not require the use of a huge database of chord hypotheses. Instead, a trained likelihood model generates reasonable hypotheses for each frame. It then constructs a search graph using these hypotheses. We demonstrate our method on Mozart’s Sonata 18, K. 570.
The proposed method is sensitive to misalignment of onset labels. If onsets are misaligned by more than four frames, the output will be a completely different target. In addition, the method’s local maximum detection algorithm can’t detect consecutive onsets shorter than 40 ms. The proposed algorithm can’t detect note durations that are shorter than this, which is rare in a real piano performance.
The main tradeoff in automatic transcription is between Recall and Precision. High sensitivity leads to many notes yielded, but it also leads to more false positives. However, lowering this parameter results in higher Precision, and lower Recall. The F-measure, however, accounts for the harmonic mean between Recall and Precision.
Subscribe to Tartalover Youtube Channel for more Piano Transcriptions videos
AMT is generally performed on a digital computer. The input signal is generally represented as a sequence of discrete values. Hence, the onset and offset of a random frame will most likely be stationary. A common approach to this problem is to divide the input signal into multiple frames. Then, the results are combined to obtain a transcription.