MRM - Music Records Matching

  

The application demonstrates music identification technology. Given a short audio clip it finds the full record.

The application demonstrates use of two technologies developed by Sloud Inc.: Audio (Acoustic) Fingerprinting and Audio Identification.

Sloud MRM Acoustic Fingerprint (audio fingerprint) is a compact representation or index of an audio record. The audioprint can be used to find particular record in a large music database. This audioprint knows nothing of the title of the record, or performer, or any other type of textual description of the music. It's based exclusively on the sound data itself. Sloud MRM Audio Fingerprinting is constructed in two stages:

• conversion of audiodata to some standard format (PCM);
• indexing of the PCM data from the beginning to the end.

The current demo works with MP3 files because they seem to be the most commonly used. But any other type of music files can be used just as well and support for other formats is planned.

Sloud MRM Audio Identification is the technology which identifies the records by their Sloud MRM Acoustic Fingerprints. This component records a short 15-second fragment of the record, creates its audioprint, then searches a database for similar audioprints.

Both technologies are targeted at weakly distorted records. Such alterations appear, for instance, when the record is digitized, when it's converted from one audio format to another, or during compression/decompression. The modifications introduce artifacts that can be heard as a slight hissing or alterations in high or low pitch sounds of the music. Thus, only partial file content is used for indexing. For instance, two records digitized at two different sampling frequencies will produce either identical or very similar indexes.

In order to demonstrate technology clearly, the Sloud MRM Audio Identification records sound played by the user's media player (as opposed to directly searching for file fragments).

Technical Specifications

1. The probability of correct identification of the records in WMA or MP3 not worse than 98% (the query is in an MP3 or WMA format, the database indexes constructed from the original CD). If the formats are the same, the correct identification rate is very close to 100% (99.9%-99.99%).
2. Index size: 155 byte for 1 second of music. This is at least an order of magnitude smaller than in other products in this field.
3. Minimum query size (granularity): 15 seconds.
4. Projected scalability: to several million records (with clustering). The application could be useful to copyright holders in order to control distribution of music online or in broadcasts. It can also be used by search engines to identify music files found on the web. Or it can be used to identify video clips by the background sounds.
5. Search time (locally): about 2 seconds regardless of the database size.