Adaptive Transform Acoustic Coding - ATRAC3 (LP2 and LP4 Modes)

ATRAC3 (LP2 and LP4 Modes)

Like ATRAC1 and MP3, ATRAC3 is also a hybrid subband-MDCT encoder, but with several differences.

In ATRAC3, Three stacked QMF split the signal into 4 parts:

• 0 to 2.75625 kHz (DC to f/16)
• 2.75625 to 5.5125 kHz (f/16 to f/8)
• 5.5125 to 11.025 kHz (f/8 to f/4)
• 11.025 to 22.05 kHz (f/4 to f/2)

The four subbands are then MDCT encoded using a fixed length transform. Unlike nearly all modern formats, the transform length cannot be varied to optimize coding transients. Instead, a simpler transient encoding technique called gain control is used, in which the gain of different subbands is varied during a transient prior to MDCT and then restored during decoding after the inverse MDCT to try to smooth over transients. Additionally, prior to quantization, tonal components are subtracted from the signal and independently quantized. During decoding, they are separately reconstructed and added back to reform the original MDCT coefficients.

Sony claims the major advantage of ATRAC3 is its coding efficiency, which was tuned for portable DSP which provides less computing power and battery life. However, as ATRAC is a hybrid subband-MDCT codec that is algorithmically very similar to MP3, any advantage is probably exaggerated. Furthermore, compared to newer formats such as Windows Media Audio which use a simple MDCT rather than a hybrid, ATRAC3 must perform an additional and computationally expensive inverse-QMF, although the hybrid system does significantly reduce memory usage, which was likely a factor given the limited memory available when ATRAC was originally developed.

LP2 Mode

This uses a 132 kbit/s data rate, the quality of which is advertised to be similar to that of MP3 encoded at a similar bit rate. However, in an independent double-blind test (2004/05) without format encoding parameters reference against Ogg Vorbis, AAC, and LAME VBR MP3, ATRAC3 came last. Unfortunately, due to lack of transparency in ATRAC encoder versioning, it is not known if the ATRAC3 encoder tested was optimal, and subsequent investigation was inconclusive. It is possible that newer ATRAC3 encoders offer better performance.

LP4 Mode

This reduces the data rate to 66 kbit/s (half that of LP2), partly by using joint stereo coding and a lowpass filter around 13.5 kHz. It allows 324 minutes to be recorded on an 80 minute MiniDisc, with the same padding required as LP2.

FFmpeg has an implementation of an ATRAC3 decoder, which was converted to fixed precision and implemented in the Rockbox series of firmwares for ARM, Coldfire and MIPS processors. RealAudio8 is an implementation of ATRAC3.

The PlayStation 3 game Race Driver: Grid uses 224 simultaneous streams of ATRAC3 compressed audio, with between one and eight channels per stream at sample rates between 24 and 48 kHz, each filtered using 512 frequency bands of adaptive equalisation, routed via six reverb units running on the same SPU co-processor (one of eight on the PS3's Cell chip), alongside 7.1 channel hybrid third-order Ambisonic mixing.