HydrogenAudio

Lossy Audio Compression => AAC => AAC - Tech => Topic started by: yanchen on 2002-12-18 03:45:48

Title: psychoacoustic analysis about pre-procesing tools
Post by: yanchen on 2002-12-18 03:45:48
Roughly, we can conclude the aac feature tool -- M/S, prediction, TNS .... as a kind of pre-processing stage before quantization. However, we perform the psychoacoustic analysis directly from the input sequence or the output of MDCT. The psychoacoustic info used as masking curves has no business with the shaped MDCT coefficents. Ideally, we should also take some adaptations to the masking curves by the info feed back by pre-processing tool. As for M/S, it might be a easier task. But for TNS or another ones, it relates to a cross-domain management which needs complex consideration and seldom mention by publized literature.

Take the TNS as an example: by the analysis by synthesis manner, we can add a de-TNS machnism to complete the reconstruction that facilitate the coming quantization noise calculation. In this manner, the masking curve needn't be modified but a clever bit-allocation for compensation is required. Nevertheless,
analysis by synthesis  quantization is suffered from speed concern.  A distortion loop is abandoned always.

Therefore, a adaptive shaping of making curve is inevitable. Can anyone comment about this issue. thank you!
Title: psychoacoustic analysis about pre-procesing tools
Post by: Wen-Chieh Lee on 2002-12-23 03:40:11
Quote
However, we perform the psychoacoustic analysis directly from the input sequence or the output of MDCT. The psychoacoustic info used as masking curves has no business with the shaped MDCT coefficents. Ideally, we should also take some adaptations to the masking curves by the info feed back by pre-processing tool. As for M/S, it might be a easier task. But for TNS or another ones, it relates to a cross-domain management which needs complex consideration and seldom mention by publized literature.


Ya, I think your idea is right.  In fact, in M/S coding, there are several M/S transformed domain psychoacoustic models are proposed (you can find a good one in LAME code). In above methods, M/S bit allocation are calculated according to M/S psychoacoustic model instead of L/R one. But, as you think PNS vs. psychoacoustic vs. bit allocation is a more complex condition to consider.