Case 1:14-cv-02396-PGG-SN Document 239-19 Filed 11/12/20 Page 1 of 2
`
`
`
`
`
`
`
`
`Exhibit 37
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 239-19 Filed 11/12/20 Page 2 of 2
`
`jim schrempp Uschrempgwenetnet]
`From:
`Monday, November 29, 1999 5:26 PM
`Sent:
`To: Wold Erling (E-mail)
`Vance ikezoye (E-mail)
`Cc:
`Wired Air Consulting
`Subject:
`Attachments: ToMFishl.ppt
`
`Hi,
`Attached is a file of two slides that will show you the kind of components I'm thinking of in
`the new application. In this application I envision a database of 100,000 to 5e0,008 song
`fingerprints. Each fingerprint will be the set of feature vectors that describe a 15 second
`piece of a song. My expectation is that the client software on the user's PC will start a
`stream analysis in MFish and take out 30 second chunks into a circular buffer.
`When the user clicks the "ID This" button the client will transmit the last
`180 seconds worth of analysis to a central Wired Air server. That server will invoke a lookup
`system that will tell us if any of the fingerprints from the database are present in the
`sample. The server will add the appropriate content (Title, Artist, Link, etc) and return
`that to the client for display.
`I'd like to have MFish build the lookup routine. As I envision it, our code will take out
`500,000 fingerprints from the database and load them into some structure. We will initialize
`the MFish routine by passing in this structure. Our code will accept 180 second samples from
`the client and pass it to the MFish lookup routine. MFish will return the AudioCBR name of
`any fingerprints found in the sample. I'd like the routines to return an incorrect
`identification less than 1% of the time and less than 4% of the time have the routines fail
`to identify a fingerprint that is in fact in the sample. I thinking that 99% of the lookups
`occur in less than 5 seconds and no lookup takes more than 15 seconds. We need to specify the
`system performance we need - how many concurrent lookups can occur on a certain processor and
`still maintain the performance.
`You asked for the first pass idea, this is it. I'd certainly consider you to do other parts
`of the system, but I think this one plays to your strengths. Brute force won't yield the
`performance or accuracy needed for this. Your expertise in both the analysis and
`identification make you the perfect place to implement some kind of clustering of the
`fingerprints prior to lookup - or some other strategy that optimizes the actual
`identification.
`We could also specify an interface that would allow you to cluster offline.
`We pass you the fingerprints, you pass back some information that we can store for you. When
`we next initialize your routine we also pass in this clustering information. This could even
`be some additional tag field in the AudioCBR object we store.
`
`Let me know your thoughts,
`Jim
`
`Jim Schrempp
`Wired Air, Inc.
`j_schrempOwiredair.com
`408 867-8645
`
`Date:
`Lana L
`
`Name:, ot-b
`CCP, CSR No. 966/
`
`Confidential - Outside Counsel Only
`
`AUDMAG01710721
`
`