I'm far away from being an expert, so don't take what I'm saying as universal truth, but as some kind of "insight help"...

The MDCT (if the window used satisfies some conditions) guarantees Perfect Reconstruction (i.e. identical input and output signals as long as there is no quantization noise introduced). First PR methods date from 1977, when PR was achieved with Quadrature Mirror Filters, combining anti-aliasing filters, down-sampling, up-sampling and interpolation. So I suppose that achieving PR using MDCT is more efficient than using QMF.

Maybe someone more experienced than me could correct me if I'm wrong.

Quote from: Ginswich on 2009-11-19 16:24:47

I'm far away from being an expert, so don't take what I'm saying as universal truth, but as some kind of "insight help"...

The MDCT (if the window used satisfies some conditions) guarantees Perfect Reconstruction (i.e. identical input and output signals as long as there is no quantization noise introduced). First PR methods date from 1977, when PR was achieved with Quadrature Mirror Filters, combining anti-aliasing filters, down-sampling, up-sampling and interpolation. So I suppose that achieving PR using MDCT is more efficient than using QMF.

Maybe someone more experienced than me could correct me if I'm wrong.

Well, the DCT is a transform.

The MDCT is an orthonormal (paraunitary, whatever) FILTERBANK.

Adjacent DCT's either have to overlap and be non-critically sampled, or alternatively have no control over frequency content across block edges.

QMF's were not originally PR, either. First FIR ones were constant delay but not amplitude flat. Then Mark Smith's IIR ones were amplitude flat but not constant delay. Then wavelets came along and the PR property was introduced.  All the forms of QMF cancel aliasing w/o any coding in between. Then there's PQMF's, which are a hybrid of a QMF and a polyphase filterbank, and can be a variety of things, including a PQMF representation of an MDCT, OBt, etc.

Malvar's book really would be a lot of help here.

Quote from: Woodinville on 2009-11-19 16:53:21

Quote from: Ginswich on 2009-11-19 16:24:47

I'm far away from being an expert, so don't take what I'm saying as universal truth, but as some kind of "insight help"...

The MDCT (if the window used satisfies some conditions) guarantees Perfect Reconstruction (i.e. identical input and output signals as long as there is no quantization noise introduced). First PR methods date from 1977, when PR was achieved with Quadrature Mirror Filters, combining anti-aliasing filters, down-sampling, up-sampling and interpolation. So I suppose that achieving PR using MDCT is more efficient than using QMF.

Maybe someone more experienced than me could correct me if I'm wrong.

Well, the DCT is a transform.

The MDCT is an orthonormal (paraunitary, whatever) FILTERBANK.

Adjacent DCT's either have to overlap and be non-critically sampled, or alternatively have no control over frequency content across block edges.

QMF's were not originally PR, either. First FIR ones were constant delay but not amplitude flat. Then Mark Smith's IIR ones were amplitude flat but not constant delay. Then wavelets came along and the PR property was introduced.  All the forms of QMF cancel aliasing w/o any coding in between. Then there's PQMF's, which are a hybrid of a QMF and a polyphase filterbank, and can be a variety of things, including a PQMF representation of an MDCT, OBt, etc.

Malvar's book really would be a lot of help here.

Can you share the book name ?