Digital Signal Processing Reference
In-Depth Information
of fixed size. In this case, packet sizes can be varied by changing the amount
of speech data (e.g., frames) placed into a packet: several speech frames could
be packetized together when the channel is good, just a frame or two when the
channel is noisy. Changing, however, the amount of speech data —be that PCM
samples or frames— encapsulated into packets causes causes potentially high
delays; moreover, such delay would be time-varying.
A different, constant-delay approach based on network-driven variable bit-
rate speech coders is possible. Variable bit-rate coders can compress speech seg-
ments in frames of different dimensions according to the selected codec mode.
For example, the GSM AMR can generate —as discussed in Section 15.2.2—
eight different frame sizes, all representing 20 ms of speech, ranging from
95 bits (lowest quality) up to 244 bits (highest quality). When a variable-rate
speech codec is available, higher rates (i.e., larger packets) can be used for good
channel conditions, lower rates for noisy conditions. The trade-off is, therefore,
between source coding quality —which is proportional to the speech coding
rate— and transmission performance —which depends on the packet size. For
at least some scenarios, the expectation is that the source coding degradation
experienced when the channel is noisy (when lower coding rates are employed)
is more than compensated by better transmission performance in terms of lower
packet losses, end-to-end delay and network congestion.
The proposed technique requires a channel estimation algorithm to select the
optimal output rate of the speech coder at any given time instant. The GSM
AMR standard leaves link quality estimation open. However, it provides an
example solution, which is based on burst-wise C/I estimates [11]. Channel
estimation schemes are covered in some detail in [4]. A carrier signal estimate is
computed using a training sequence known a priori. By comparing the received
training sequence and the known training sequence the receiver can estimate the
current C/I and communicate it to the sender. For codec mode adaptation, the
measure of the instantaneous channel quality has to be mapped to codec modes.
This is in principle done by quantizing the measurement where the levels of the
quantizer are mapped onto the different codec modes.
For the proposed technique, channel quality measurements are roughly quan-
tized in two states that represent good and bad channel conditions. In the bad
state, large packets have a higher probability to be in error and, therefore, to be
retransmitted, while small packets are more easily received without errors.
