Re: [codec] Thresholds and delay.

[Ben Schwartz <[e-mail]>]

On Tue, 2010-05-11 at 12:48 -0400, Marshall Eubanks wrote:
> As a point of order, I object to any graphs without an available paper  
> behind them.

I have located the first paper mentioned by Christian Hoene at
http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=81952
but of course it's paywalled.

One test in that paper told trained subjects to "Take turns reading
random numbers aloud as fast as possible", on a pair of handsets with
narrowband uncompressed audio and no echo.  Subjects were able to detect
round-trip delays down to 90 ms.  Conversational efficiency was impaired
even with round-trip delay of 100 ms.

Let me emphasize again that these delays are round-trip, not one-way,
there is no echo, and the task, while designed to expose latency, is
probably less demanding than musical performance.

In the presence of echo, round-trip delay must be kept below 30 ms to
ensure that the echo is perceived as sidetone, according to the Springer
handbook of speech processing:

(http://books.google.com/books?id=Slg10ekZBkAC&lpg=PA83&ots=wc9yM9WrCs&dq=sidetone%20delay%2030%20ms&lr&pg=PA84#v=onepage&q&f=false)

Such low delays are clearly impossible on many paths, but for Boston to
New York City (or London to Paris), ping times can be less than 18 ms,
making echo->sidetone conversion just barely possible for a codec with
5ms frames.

I accept Brian Rosen's claim that a slow conversation doesn't normally
suffer greatly from round-trip latencies up to 500 ms, but under some
circumstances much lower latencies are valuable.  Let's make sure
they're achievable for those who can use them.

--Ben