Updating Voxin TTS Server To Avoid A Possible ALSA Bug

1 Summary

I recently updated to a new Linux laptop running the latest Debian
(Rodete). The upgrade went smoothly, but when I started using the
machine, I found that the Emacspeak TTS server for Voxin (Outloud)
crashed consistently; here, consistently equated to crashing on short
utterances which made typing or navigating by character an extremely
frustrating experience.

I fixed the issue by creating a work-around in the TTS server
atcleci.cpp::xrun
— if you run into this issue, make sure to update and rebuild
atcleci.so from GitHub; alternatively, you'll find an updated
atcleci.so in the servers/linux-outloud/lib/ directory after a
git update that you can copy over to your servers/linux-outloud
directory.

2 What Was Crashing

I use a DMIX plugin as the default device — and have many ALSA
virtual devices that are defined in terms of this device — see my
asoundrc. With this configuration, writing to the ALSA device was
raising an EPIPE error — normally this error indicates a buffer
underrun — that's when ALSA is starved of audio data. But in many
of these cases, the ALSA device was still in a RUNNING rather than
an XRUN state — this caused the Emacspeak server to
abort. Curiously, this happened only sporadically — and from my
experimentation only happened when there were multiple streams of
audio active on the machine.
A few Google searches showed threads on the alsa/kernel devel lists
that indicated that this bug was present in the case of DMIX devices
— it was hard to tell if the patch that was submitted on the
alsa-devel list had made it into my installation of Debian.

3 Fixing The Problem

My original implementation of function xrun had been cloned from
aplay.c about 15+ years ago — looking at the newest aplay
implementation, little to nothing had changed there. I finally worked
around the issue by adding a call to

snd_pcm_prepare(AHandle) 

whenever ALSA raised an EPIPE error during write — with the ALSA
device state in a RUNNING rather than an XRUN state. This
appears to fix the issue.