Sometimes it is better not to reject the data on occurence of a short "glitch".

This kind of encoding-decoding (Manchester, biphase, "FM") is common in RFID. In one of our projects, with my colleague, we concurrently implemented two methods of decoding - I measured the edges and rejected data on out-of-bounds times. He started a one-and-half-bit-time timer on an edge and then sampled the current level (well that was not classical manchester, there was a guaranteed edge transition at each boundary of bit and an extra in the middle to distinguish 0 from 1); he rejected bad data based on the higher level protocol violation only. On the same hardware, "his" algorithm resulted in twice as big reading distance than mine...

But back to your raw problem.
- 2 halfbit = "long", 1 halfbit = "short"
- in case of length violation reset to state A
- start in
A - waiting for synchronising long pulse, if high (10), goto B, if low (01), goto D
B - halfway through 0, waiting for low pulse. If long (01), emit 0 and goto D. If short (00), goto C
C - in between 00. Wait for short high pulse, emit 0 and goto B.
D - halfway through 1, waiting for high pulse. If long (10), emit 1 and goto B. If short (11), goto E
E - in between 11. Wait for shirt low pulse, emit 1 and goto D.

Off the top of my head, please check for sanity.

Jan Waclawek