8VSB
OK, we've assembled the audio and video packet streams. And we've added the PSIP tables which tell the receiver how to assemble these streams into TV programs. What we have is a long stream of 188-byte data packets. They could be connected to the TV receiver with an Ethernet cable, but somehow I don't think that's what the NAB had in mind<grin>. The 8VSB modulator is responsible for taking these data packets and converting them to a radio signal that can be connected to an antenna and broadcast¹⁴.

Before anything RF happens, the data stream is passed to a “data randomizer”. A raw ATSC datastream has patterns. Certain sequences in the datastream are more common than others¹⁵. If used to directly modulate a RF signal, this would result in “clumps” of spectrum occupancy – and other “empty” areas. Instead, the ATSC datastream is mixed with a seemingly-random series of numbers. “Seemingly”-random, because the same series is generated at the receiver and used to retrieve the original datastream. The mixing process generates a new datastream in which all possible data sequences are roughly equally likely.

Next, the synchronizing byte is removed and 20 bytes of a “Reed-Solomon Forward Error Correcting Code” (FEC) is added. The FEC code is a series of numbers calculated based on the 188 bytes of “payload” data.

“Forward Error Correction” means the receiver can discover and repair some errors in transmission without asking the transmiter to retransmit the information. (which of course would be impossible in a broadcast environment!) As many as 10 bytes in the 188-byte data “payload” can be in error, and the FEC will still correct them. More than 10 bytes in error cannot be repaired, but the situation can be detected. The defective packet will be discarded.

Once the FEC is applied to the data packet, it is interleaved with other data packets. The contents of the packet are spread across a 4.5 millisecond period. The point is to reduce the exposure of any one packet to a burst of interference or noise. Again, the FEC can fix up to 10 bytes of errors. So it's better to allow a noise burst to wipe out 7 bytes of each of 10 packets, than to allow it to wipe out 70 bytes of one single packet. The former problem can be automatically fixed; the latter will result in the packet being discarded.

Then, “trellis coding” is applied. Every two bits are recoded to three. (for reasons for which I've not been able to find a readable explanation...) Three bits of information can represent up to 2^3, or 8, different numbers.

Next, the previously-stripped “sync byte” is replaced with a four-symbol segment sync. When the receiver detects this segment sync, “sync lock” is said to have occurred. Users of the WinTV-D card can tell when this happens by using the “Diagnostics” window. Sync lock is possible at 0dB signal-to-noise ratio – i.e., when the 8VSB signal and the noise are the same strength. A 15dB ratio is necessary to actually decode the signal¹⁶.

Finally, the stream of 3-bit trellis-coded numbers is placed on the RF carrier. In analog TV, the analog video is used to modulate the amplitude of a carrier 1.25MHz above the bottom of the TV channel. The instantaneous output power of the transmitter may be any figure between 5% and 75% of maximum rating¹⁷. The lower sideband of the analog TV signal is then “rolled off”, to prevent it from stretching into the adjacent lower channel.

In digital, the process is similar. The carrier is 0.31MHz above the bottom of the channel – 0.94MHz lower than in analog. The lower sideband is still rolled off – but so is the carrier. The carrier uses 7% of the average power of the 8VSB signal¹⁸. While an infinite number of signal amplitudes are possible for analog TV, only 8 amplitudes are allowed for 8VSB. (that's where the “8” comes from) The range of possible amplitudes is split into eight equal parts, tagged -7, -5, -3, -1, +1, +3, +5, and +7. Each possible 3-bit trellis-coded number maps to one of these eight amplitudes.