Machine learning model for signal isolation in drum recordings
Karaksela, Nikke (2024)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2024112831086
https://urn.fi/URN:NBN:fi:amk-2024112831086
Tiivistelmä
The aim of this thesis is to train an audio denoising machine learning model that can remove bleed from snare drum recordings while still attaining all the characteristics of the sound of the drum. Microphone bleed – a common prob-lem in music production – happens when a microphone unintentionally cap-tures unwanted secondary sound sources. This issue especially occurs when multiple sound sources and microphones are being used like when recording a drum set.
To develop a snare drum denoising model, a comprehensive dataset of snare drum recordings was gathered and used. The model was trained using ma-chine learning techniques. The results indicate that the proposed model suc-ceeds in reducing microphone bleed but not eliminating it entirely. The model faces challenges in accurately denoising recordings with ghost notes and dif-ferentiating between snare drums and other percussive elements such as tom and bass drums.
These findings suggest that further improvements could be achieved by incor-porating a higher sample rate, more diverse training data, and exploring differ-ent model architectures. Additionally, more computational resources could en-hance the training process and the overall performance of the model. This re-search contributes to the field by demonstrating the potential of machine learn-ing to solve practical audio engineering problems such as the issue of micro-phone bleed is.
To develop a snare drum denoising model, a comprehensive dataset of snare drum recordings was gathered and used. The model was trained using ma-chine learning techniques. The results indicate that the proposed model suc-ceeds in reducing microphone bleed but not eliminating it entirely. The model faces challenges in accurately denoising recordings with ghost notes and dif-ferentiating between snare drums and other percussive elements such as tom and bass drums.
These findings suggest that further improvements could be achieved by incor-porating a higher sample rate, more diverse training data, and exploring differ-ent model architectures. Additionally, more computational resources could en-hance the training process and the overall performance of the model. This re-search contributes to the field by demonstrating the potential of machine learn-ing to solve practical audio engineering problems such as the issue of micro-phone bleed is.