NU2011-063

Inventors

Bryan Pardo*

Zafar Rafii

Short Description

REPET uses a simple but reliable technique for music/voice separation.

Abstract

Northwestern researchers have developed a novel technique for music/voice separation. Currently available separation methods are more demanding and complex, whereas this new invention called REpeating Pattern Extraction Technique or REPET uses a simple but reliable technique. It is also faster and completely automatable. In fact, evaluation of 1,000 song clips showed that REPET achieves better separation performance than existing automatic approaches. The advantage of REPET is its exploitation of a core principle in music – repetition. REPET separates music from voice, by extracting the repeating musical structure. After it identifies the period of the repeating structure and segments period boundaries, the segments are averaged to create a repeating segment model. Each time-frequency bin in a segment is then compared to the model, and the mixture is partitioned using binary time-frequency masking by labeling bins similar to the model as the repeating background. Because REPET utilizes “self-similarity” between repeating segments, it works on a variety of audio signals having one or more repeating patterns within a recording.

Applications

• Music search engines (simply by singing into a microphone)
• Music/Voice transcription
• Audio post-production
• Audio analysis for remixing
• Sound separation
• Sample-based musical composition
• Karaoke: Active noise removal

Advantages

• Entirely automated – no user input needed
• No system “training” needed
• Simple processing – no complex frameworks needed

IP Status

Issued US Patent No. 9,093,056