AET 2050 - DAW Production

2. Data Reduction for Audio

Objectives

• Identify the need for data reduction systems.
• Differentiate between lossless and lossy reductions.
• Define critical audio bands and their importance in perceptual coding.
• Compare and contrast the various coding systems currently in use.

Why Data Reduction?

As a data coding format, PCM is robust but inefficient. It requires data rate of 1.4 Mbit/s for 2 channels of 16 bits @ 44.1k.

Many contemporary delivery systems (such as those used on the Internet) require much lower data rates - as low as 128kbit/s per channel. This is more than 10 times data reduction needed.

Data Reduction Process

Direct data reduction

• Can be done one of two ways - reduce sampling or reduce quantization:
  • Reduction in sampling rate will reduce frequency response.
  • Reduction in quantization will reduce precision and introduce quantization noise.
• Usually, these compromises are unacceptable for high-quality audio.

Instead of reducing the actual data, a system is needed that can represent the data in an coded form that uses less actual data.

Data Reduction Coding

• Lossless coding - allows for data to be reduced and reconstructed into its original form.
  • Method used by Zipit or Stuffit
  • Method works well for most compression needs
  • Does not yield a good enough reduction for audio (about 2.5:1).
• Lossy coding - a system whereas data is not precisely reconstructed when decoded, but a (hopefully) accurate analogy is produced from the reduced data.
  • Lossy techniques reduce bit depth
  • Raises quantization noise, but do so in such a way that the noise is hidden.
  • Result is a change in dynamics and noise, but (optimally) small enough to be indistinguishable to the human ear.

Perceptual Coding

Perceptual coding is based on psychoacoustic principles which define how some audio is masked by other audio.

Critical Bands

• Masking occurs with tones inside of frequency bands
• A given tone will mask another tone within that band, but will not affect tones outside of that band
• Bandwidth of these bands approximated to be about 1/3 octave for frequencies between 300-20,000 Hz
• Bands are not fixed, but are continuously variable and and any audible tone will create a band centered around it.
• The masking tone raises the threshold of perceived hearing around that tone
  • Sound beneath that threshold is masked
  • Sound outside of the tone's critical band will not be affected.

Perceptual Coding Process

• Perceptual coding effectively reduces the bit rate of a signal by implementing the psychoacoustic principles based on critical bands and the masking phenomenon.
• The signal's sample rate is maintained, but the word length is selectively decreased dynamically based on signal conditions. Masking is considered so that the increase in quantization noise is rendered as inaudible as possible.
• Works by taking advantage of the masking characteristic of human hearing
  • Sounds that occur at the same time as louder sounds can be removed because we can't hear them
  • Frequencies above 15 kHz are sometimes not encoded at all (less common now)
  • Lower number of bits used to encode high frequencies
  • We can't hear the harmonics produced by a 20 kHz tone, so "squaring off' the waveform will be imperceptible

• The encoding process
  • Incoming audio is split into many narrow frequency bands
  • Audio is sent to an auditory model that decides what audio is "important' and what isn't
  • Each band can then be requantized using fewer bits
  • Only levels above the threshold of perception are quantized. The higher the level, the more bits that are used
  • Requantizing effects are constrained within the bands, and are are more effectively masked by the band's program material
  • Process repeats every few millisecond
• Although lossy, theoretically, the listener will not perceive the loss.
• Encoders generally allow the user to select a target bit rate, usually measured in kbps (kilobits per second)
• Results in audio files drastically reduced in size with little perceived difference in quality
• Useful for end-users only, due to the fact that audio quality decreases from generation to generation

Application

• Real-time streaming applications
  • Low bit rates allow for high-quality audio to be "streamed" from the web in real time
  • Speech audio transmitted through digital cell phone networks is encoded and decoded in real time
  • Delay can be noticeable
• Encoding software/algorithm plays an important role in the quality of audio encoded
  • Incoming audio is split into many narrow frequency bands
  • Audio is sent to an auditory model that decides what audio is "important' and what isn't
  • Process repeats every few milliseconds

     

  EBU subjective listening tests on low-bitrate codecs

Maximum Streaming Rates