A quasi-stationary source is one which is stationary over short periods, but changes occasionally to a new mode of stationarity; a typical example is the speech source. Convolutional codes are designed for the speech source for use with search algorithms such as the Viterbi, stack, andM-algorithms; the design is first by heuristic means and then by an empirical optimization. These codes have the same generating structure as the usual binary convolutional codes, but employ ordinary arithmetic. It is found by experiment that, at rate 2 bits/sample (which allows telephone quality speech), such codes need not have constraint length longer than five or six, which implies a generating circuit of about 1000 states. Encoding noise becomes white and uncorrelated with simple code searching; this shows that a short fixed code can successfully decorrelate waveforms from different source modes.