I noticed a blurb in another thread about a non-linear psymodel. Is there more information regarding what state of quality it delivers? Any opinions as to it's possible benefits and drawbacks? I'm surprised there hasn't been more of a discussion on this topic.

Sorry I haven't been around lately but the LAME project seemed to have largely fallen off in the last year and there wasn't enough to read about it until recently.

I would imagine a non-linear psymodel is capable of looking at a wav file entirely and being able to apply blocks, bits, and other tricks in a much more productive way to improve pre-echo and other artifacts caused by masking and bit-deficiency. Is this indeed the case? And if so, is anyone in the LAME project working to develop this psymodel for a future LAME release?

tlc

"Non-linear" psymodel is a vague term.. it could mean just anything where some intermediate result is calculated by non-linear mean (exponential, power-law, etc...)

I guess you are referring to:

http://www.mp3-tech.org/programmer/docs/AES99-FB.zip

This model uses assumptions found by Lutfi in JAES article "Additivity of simultaneous masking” - that additions of individual masking components are not linear but somewhat exponential - i.e. the total masking capability of one frequency band is greater than linar sum of individual maskers.

This might, or might not be better than linear summation in psymodel II used in many codecs, but the problem is that we could only talk about the efficiency of the complete coding system - not just psymodel itself.

This method is used in ITU-R BS.1387 PEAQ tool.

Ideal way to judge this would be to implement this in "threshold simulator" or encoder which directly injects noise - like MPC, or MP3/AAC with scale factor estimation - so you could directly tune it up and compare to Psymodel II.



Actually, LAME already has this code (not enabled by default) - but they are using it in a wrong way IIRC (maybe I am wrong) - they estimate tonality and then perform non-linear masking and this results in overmasking - non-linear spreading actually replaces tonality estimation - these two tools are not really compatible with each other.

Apart from this, there are some better ways ot estimate masking power without this operation (as reported by research data), which is very CPU intensive because instead of simple addition you must perform one pow() (or, exp() at least) per sample - trully a killer for CPU



There is no difference how the model "looks" between Psymodel II and this model - only difference is the way how the masking power is determined for each band - and this nonlinear model use signal-adaptive spreading function, while LAME uses fixed Psymodel II one.

MPC also uses signal-adaptive spreading function.

Thanks! You answered all my questions and even added some cool info on top of it. Sounds like it could potentially be a good psymodel but from what you're saying it will take quite a bit of new re-tooling to get LAME to work with it properly.

mp3

You can download a lame_enc.dll using this from: http://dspguru.doom9.net/. I haven't tried it so I can't comment on how it performs.