mHashup has featured on BBC Click with Shazam, Midomi and Pandora, in a piece about music recognition by LJ RIch. You can watch the mHashup clip (1:35) or the full report (4:54) below.
The above demo uses Schubert's "Ständchen" recorded by German soprano Emmy Bettendorf (1895-1963) as starting point for finding similar melodies in the database. The closest match proves to be the same Emmy Bettendorf recording, though by examining the metadata, the title "Serenade" indicates a different publisher. The system has effectively found a duplicate of the same melody.
The next closest match, by Swedish tenor Jussi Björling (1911-1960), is the same melody, though an octave lower. This is followed by British tenors Webster Booth (1902-1984) and Frank Titterton (1893-1956). Note that the matching melody recorded by Titterton is located at 98 seconds in the track, as it is the second instance of that same melody in the track. The recording by John McCormack, 112 seconds into the track, is the same melody, but this time at a different pitch, a semitone higher. The system can accurately deliver similar melodic patterns, regardless of pitch or octave.
The first 6 tracks listed are versions of the same piece; 'Serenade' or 'Ständchen'. Having run out of the exact same melody, the system will match melodic sequences which are similar to the sampled melody. In recordings by Gerhard Hüsch and Lotte Lehman we can detect a similar style in the rendition of another Schubert piece.
The final match offers a surprise finding: a reading of Shakespeare's 'Henry V' by Welsh poet W J Gruffydd. On analysis, it becomes apparent that Gruffydd hits the very same notes that Schubert wrote.
Content courtesy of Dan Leech-Wilkinson and the King's Sound Archive, University of London.
Joyce Hatto was a concert pianist who stopped performing in public due to ill health. However, her latest recordings carried on being released on the Concert Artist label owned by her husband William Barrington-Coupe, to great reviews. Many questioned the artist's ability to record pieces of excellent quality, while being at the same time unable to perform in public. The story started unravelling when a musicologist pointed out that an allegedly Hatto-recorded misreading of a chord was identical to one on a Carlo Grande recording. Further studies, including a key series of tests by the CHARM research group at the Royal Holloway, University of London, proved that the celebrated late Hatto recordings had been slightly modified and engineered copies of recordings by a variety of other, sometimes lesser-known, artists.
When suspecting copyright infringement, the key is the ability to access the source of the infringement, with a precise point of origin within the copied track, for exact comparison. Where recordings have been copied but slightly altered by stretching or modifying the sound quality, the match may not be immediately obvious to the naked ear and one may have to trawl through hundreds of recordings looking for clues.
On testing copyright infringement with mHashup, a segment is chosen from a Chopin Mazurka credited to Joyce Hatto. One can expect to find virtually identical segments of the same Mazurka by different recording artists. However, the best match proves to be not only similar, but the exact same finger work, with the same accents and dynamics, audible even to a layman. The segment shows an exact match to an Indjic recording, with a precise location marker as to the starting point of the copied material.
This audio match can be tested further by choosing other segments of Indjic and Hatto. All the results point to the same: it quickly becomes apparent that the source of the Hatto Mazurkas is an original recording by Eugen Indjic.
Content courtesy of CHARM, Royal Holloway College, University of London.
As a possible entry point for mHashup, a multi-layered interface allows for query triggers from a geographical plane located at a level below the similarity interface. The inspiration for this interface were a set of ethnomusicological recordings of world music compiled by Alan Lomax as part of the Cantometrics project. The nature of the recordings suggested a search by geo-location. The layering of geo-positioning and similarity matching enabled the discovery of interesting anthropological and ethnomusicological relationships.
The Lomax recordings can be sampled at their geo-locations before a query is triggered. They are represented by simple, colour-coded symbols relative to the annotation: male or female singer, solo or group, vocal or instrumental, predominately based on the work of Polina Proutskova. Global positioning of the individual symbols was done manually, in absence of any geographical coordinates in Lomax's original annotation. This occasionally presented a problem: some of the countries and people's names have since changed, and some locations belong to a different country than they did in Lomax's time. We have tried to keep faithful to the original descriptions as much as possible, though cannot guarantee absolute accuracy.
On query launch, the recorded sample is matched to other Lomax recordings around the globe and brought to view in the mHashup similarity interface. The similarity interface and the geographical interface create layers which can be independently reduced in size or enlarged as required. The matched recordings can be sampled individually, with a highlighting device tracking the query trajectory from one part of the globe to another and creating a web of pathways indicating relationships between cultures in distant locations of the globe.
The geo interface was first presented at Unlocking Audio, British Library. London, March 2009.
Above is a version of the interface which focuses on the 100 closest relationships of shingles from files stored in a database to any starting point of a shingle within a chosen musical track. This provides a statistical overview of the tracks closely related to the query track, their most related sections (clusters of results form in a section of a track), and the number of times the same track is matched (suggesting a closer overall relationship).
The design provides a zoom function to reveal a landscape of 100 horizontally arranged tracks with their musical relationships revealing clusters of vertical links. Each link can be sampled by clicking on the matched shingle, which automatically highlights the shingle in the query track which it is related to. The audio feedback from the paired shingles gives evidence of the audio similarity relationship.
Content courtesy of AWAL - Artists Without a Label.
mHashup was a visual interface to large music collections for discovering musical relationships among tracks developed by Michela Magas in conjunction with the OMRAS2 group at the University of London (2007-2010). The academic research project was funded by the EPSRC. The project was followed by a commercial BETA proof of concept funded by Stromatolite with support from the UK Technology Strategy Board. The proof of concept developed new technologies to work with commercial music collections. The resulting Sonaris technology had its first public demo at the Cisco Big Awards in 2012.