All these recommendations are very relative, because recognition accuracy strongly depends on initial musical performance. Some instruments, especially with smooth attack are badly detected by TS-AudioToMIDI recognition engine. Drums can't be recognized precisely; instead ordinary instruments emulate them.
Besides, it is difficult to describe sound with words, for this may lead to misunderstanding of some expressions, such as smooth attack, for example. It is intuitively evident, that some instruments sound smoother than others, but it is tricky to introduce any quantitative measure for this "smoothness".
The following recommendations on improving recognition quality are particularly effective in case you experiment with parameters as well:
1. Always use appropriate note detection algorithm. Polyphonic for recognizing performances with several voices, monophonic for transcribing monophonic performances or for detecting one voice from a polyphonic performance.
2. It is recommended to avoid recognizing performances with drums, or at list cut drums off with the help of Equalizer. Drums usually give strongest surges on low and high frequencies, so you may need to reduce amplification of edges and increase it in the middle frequencies.
3. Set up Harmonic Model for instrument used in performance you are recognizing. See Setting Harmonic Model for method of determining appropriate harmonic ratio.
4. Use Auto Tune if you are not absolutely sure that performance was recorded with well-tuned instruments.
5. Use Selectivity, Sharpness and Threshold parameters to optimize TS-AudioToMIDI between producing many confusedly placed notes on the one hand (when too much notes are passed through) and silence, sometimes interrupted by notes on the other hand (too much notes filtered out). Optimal recognition settings lie just in the middle between these two extremes.
6. Do not forget about Minimal Note duration parameter. It is very powerful tool for reducing amount of "mesh" (short confusedly placed notes). Remember that Minimal Note duration parameter is taken into account only in non real-time recognition mode.
7. Do not be afraid to experiment with parameters. If you find some parameter configuration producing almost clear transcription, save it to Settings file and try to improve it. Usually there is no optimal settings configuration for the performance, for example, reducing Threshold can be compensated by reducing Sharpness or increasing Minimal Note duration.
Take a look at sample files we offer you to download from our site. Each sample is provided with Settings file for you to see how the settings were set. This may help you to get empirical understanding of the way settings can be adjusted.
Related topics:
Choosing Recognition Algorithm
Setting Harmonic Model
|
