So, can the TIA generate speech just using normal AUDC waveforms (versus AUDF based PCM)?

The initial results say no. Although you can hear some of the inflection.

In the attachment is are two WAV files (one with just a simulated TIA output and one with a the input and output in stereo) along the code & executable I used.

The code generates two frames worth of audio of each TIA AUDC+AUDF waveform (except AUDC 3 where the waveform is dependent on the initial state of the LFSRs). It then loads a frame worth of the input audio (8 bit signed PCM @ 31,500Hz) and does a brute force correlation to each of the waveforms at each sample offset (phase) to figure out which waveform best fits. It then writes out the best waveform at the average volume and loops back for the next frame of input.

One thing I will try is to see if adding a second waveform improves the output at all.

Update: wav2tia2.zip adds the second waveform and the result is more promising. A few of the letters are almost intelligible. However, I have a feeling getting better is going to be more experimentation and tweaking than following the correlation path.