Skype Journal

Home - Contact Us - Policies - Advertise - About News feed Independently covering the Talk Revolution since 2003

Tuesday, March 3, 2009

SILK performs better

Skype's been saying its SILK audio codec is better than others. They released some data today supporting their claim.

Key measure is Mean Opinion Score, which compares sound as perceived before and after processing. Higher is better, greater fidelity.

In this chart, the codecs are tested at low bitrates (hard, on the left) to high bitrates (easy, on the right). Lots of bandwidth makes it easy to replicate sounds. SILK does better even at dial-up speeds, and SILK climbs in quality with even a little extra freedom. 

People like SILK even at slow speeds by you.

That's with clean bandwidth.

SILK does well even with bad connectivity. This chart shows Skype degrades more gracefully than other codecs, twice as well as the popular free open source Speex codec and better than the Adaptive MultiRate WideBand (AMR-WB) speech codec. 

People like SILK even with data loss by you.

Three things contribute to SILK's attractiveness:

  • It's written in fixed point ANSI C, so it will run efficiently nearly anywhere.
  • It quickly adapts to changes in sample rate, network quantity/quality, and CPU resources. This minimizes audio artifacts and preserves quality.
  • Low delay frees up other parts of a system, cutting latency. SILK only needs 25 ms (20 ms frame size + 5 ms look-ahead). 

SILK does double duty with non-speech media. Skype's codec also works at music quality. Systems that stream music, television, movies, or ambient audio (games) will be able to use SILK.

Signal processing takes up huge overhead on mobile phones. As SILK moves from software to firmware, Skype suddenly takes up less memory, CPU, and power. Users get longer battery life, less heat, less latency. This would be a big win for Skype's mobile strategy. Skype would work on much dumber, cheaper, ubiquitous smartphones: a vastly larger market.

Notes from the data sheet:

MOS (Mean Opinion Score) listening test was performed for Wideband speech signals by Dynastat, an independent 3rd party laboratory. Confidence intervals (95%) are +/- 0.1 MOS. All bitrates are measured and averaged over frames containing active speech. SILK and Speex were run in the highest complexity mode. Packet Loss and Office Noise tests were done with all codecs running at 18.25 kbps.

See also:

Labels: , , , , ,

0 Comments:

Post a Comment

We've started to moderate comments to avoid spam. Please excuse the short delay. We'll get your post online a quickly as possible.

Links to this post:

Create a Link

<< Home