@julian said:
However I'm still interested in the 10-1 VSL compression claim - no one else, to my knowledge, has got beyond 2-1 without affecting data integrity i.e. not lossless.
I suspect that the high compression ratio is due to the (highly repetitive) nature of the data being compressed.
Consider how (very) low quality sample libraries work (or the early ones) - a library might just have one sample for a pitch sampled at velocity 50. To play the same note at velocity 100 it would just double the amplitude of that sample, which although far from perfect is a reasonable approximation. So for compression purposes, if you do have a sample for velocity 100 what you could store rather than that sample itself is the difference between the velocity 100 sample and the velocity 50 sample with the amplitude doubled.
Now whilst that may not achieve a 10-1 compression, consider the difference between a velocity 51 sample and the velocity 50 sample (with an appropriate increase in amplitude) - the differences here would be pretty small (possibly larger than 10-1).
Similar compression can be achieved for pitch - instead of storing the entire c2 sample, store the difference between the c2 sample and the c1 sample with the speed doubled.
Now, I don't know if the VSL compression techniques are based on the above - but the above reasoning is enough to persuade me that it is plausible that there are characteristics in the sort of data needed in a sample library which can be taken advantage of to achieve higher compression ratios than are normally possible in generic more data sets.
Matthew
With many video compression techniques for each consecutive frame only the differences are sent as new data until the next keyframe (full data requirement) is sent. This allows for significant data reduction, though when the picture image changes rapidly like in a fast moving sequence there is less redundancy frame for frame and artefacts soon become apparent.
However when VSL sample a piano for example every sample is unique. It may have the same pitch and be slightly louder than the sample before but there is no redundancy between samples of the same pitch. Each and every sample is unique and cannot share data (however compressed between another sample). So VSL makes the original recording and then edits and produces the final uncompressed samples in 24bit/44.1k for current delivery.
The quoted data size for the complete set of samples is advertised at 500GB - a rough calculation using the bit rate of 2116 bps or 264.6 Bps for a 24 bit stereo 44.1k sample gives us a total recorded duration of, amazingly, 525 hours if all the samples were played back to back in their entirety.
So by reducing the data from 500GB to 50 GB VSL are squeezing 525 hours of recordings into the space normally occupied by 52.5 hours. Now we all know mp3 encoding achieves this all the time but I do not understand how the piano samples when played from the Vienna Imperial software engine can be the same quality as the original when it is re-created from only 10% of the original data.
I do hope someone from VSL will expand on how the seemingly impossible is being achieved!
Thanks
Julian
