Musicians and artists have been using software to aid in music production for decades. As software tools have grown in sophistication, an emerging industry of artificial intelligence (AI) in music has emerged with applications in the area of music composition, performance, theory, and digital sound processing. Though music purists may take issue with AI being used to create or generate music, these technologies are also being used to refine human-generated music in the editing and mastering process.
Mastering is the process of preparing and transferring recorded audio from a source containing the final mix of instruments and/or voices to a data storage device known as the master recording. This is the source from which all copies will be produced. During mastering, equalization, compression, limiting, and other levels are set, and the audio files are cleaned to remove excess hisses, noise, pops, or other artifacts. Other processes, including specifying the gaps between tracks, adjusting level, and fading tracks in and out, can also be handled so that the final master is ready for duplication.
Tonic to the Rescue
Traditionally, the mastering process required critical listening and sound engineering skills, particularly with respect to setting levels, applying equalization curves, and other processes. If applied incorrectly, these processes could make the mix sound muddy, overly bass-heavy, or trebly. It is also possible the mix would not be able to be accurately and evenly reproduced on consumer listening devices. LANDR’s latest update to Tonic, a cloud-based music mastering and distribution platform, helps solve these issues. Tonic is designed to provide its users with machine learning-driven mastering styles that can be applied directly to a song or album without requiring music creators to touch even a single fader.
According to the Leigh Smith, LANDR’s senior research engineer, LANDR’s mastering engine is currently proprietary and in the process of patent application. The company has indicated that it approaches mastering as a multivariate regression problem that considers how to control a large number of parameters via its audio signal processing algorithms to arrive at a desired musical outcome. The goal is to allow users to quickly apply one of three distinct mastering styles that embody specific equalization and compression levels, creating specific musical flavors:
- Warm: Vintage warmth with softer compression for thick, smooth sound
- Balanced: Controlled, with focus on balance, clarity, and depth
- Open: Modern, open sound with emphasis on punch and presence
Supervised and Unsupervised Learning Create Different Flavors
Smith notes that Tonic uses a variety of machine learning and AI processes that interoperate to achieve these mastering styles. Tonic draws from datasets such as users’ audio tracks so that its systems learn exactly what to expect from user input; it also learns from commercial music datasets. The algorithms incorporate specially curated datasets created by LANDR’s in-house team of mastering engineers, and both supervised and unsupervised learning methods are deployed in different parts of its system. They also consider how the track will sound in many different listening environments to ensure the final recording is equally suitable for headphones, car speakers, or the PA system at a large sports stadium.
Smith adds that the use of supervised learning techniques allows the algorithms to learn specific behaviors that are desirable for a specific mastering style, such as identifying the relative equalization and compression levels for a specific style. However, these mastering styles are indeed styles; different mixes of a single track will produce somewhat different sounding masters due to the individual variances of frequencies contained in each mix.
Because the master styles are created via an algorithm, there are no user-controlled fine-tuning options. LANDR’s goal was to simplify and speed up the mastering process for creators. This update of Tonic is designed to provide different flavors of mastering that are commonly requested of traditional engineers without adding too much complexity to the process for music creators, Smith says.
The company was launched in 2014, and since then, LANDR’s mastering tool has been used by more than 2 million musicians from 160 countries around the world.
Through Tonic, the mastering process, which once required a significant amount of recording engineering knowledge and experience to achieve professional results, is now able to be handled with a simple click. It leverages machine learning to assist humans, rather than replace them. By using a trove of sound recordings and trained algorithms to identify the ideal settings that will yield a specific sound style, even relative amateurs are able to approximate the sound and feel of a professional recording.