Speaking certain characters with sd_voxin makes speech-dispatcher unresponsive

joanmarie commented 1 year ago

Step to reproduce

In a terminal, enter the following:

$ spd-say -c -o espeak-ng "<"
$ spd-say -c -o espeak-ng "<"
$ spd-say -c -o voxin "<"
$ spd-say -c -o voxin "<"

Obtained behavior

With espeak-ng, the character is spoken each time and spd-say exits. With voxin, nothing is heard after the first attempt, but spd-say exits. With the second attempt, nothing is heard and spd-say does not exit. The sd_voxin process's CPU usage spikes to 100%. (Are there debugging symbols we can get our hands on? Output from gdb was useless.)

Expected behavior

The character would always be spoken, but failing that, spd-say wouldn't hang.

Environment, logs, etc.

The contents of my voxin.log is just a single line: Error opening configuration file '/etc/speech-dispatcher/modules/voxin.conf'. But voxin works fine (modulo this particular bug). For instance if I replace "<" with "h", the bug does not occur.
speech-dispatcher.log
Fedora 39, speech-dispatcher-0.11.5-1.fc39.x86_64
Voxin 3.3rc5 (Guess I should check if there's an update...)

sthibaul commented 1 year ago

Since sd_voxin comes from the Voxin package, that'd rather have to be discussed with the Voxin authors?

sthibaul commented 1 year ago

Note: to get more logs, you can put

Debug 1

in /etc/speech-dispatcher/modules/voxin.conf

sthibaul commented 1 year ago

What I can see in the logs on my system is that in both < and > cases, < and > are sent to the synth. Only the former case seems to be posing problem, speech-dispatcher's ibmtts.c (used for the sd_voxin module shipped by speechd) is stuck in the eciAddText call, so it cannot do better here...

joanmarie commented 1 year ago

Makes sense, but any guesses as to why it's only a problem now? It was reported on the Orca list, and I'm rather surprised that we're only just now hitting it....

joanmarie commented 1 year ago

For completeness, here's the voxin.log from 3.3rc6.

joanmarie commented 1 year ago

Also for completeness, I just emailed the contact address for Oralux to report this issue. (Couldn't find an issue tracker.) I CCed @sthibaul. :crossed_fingers:

sthibaul commented 1 year ago

Makes sense, but any guesses as to why it's only a problem now?

Some improper parsing of the < entity, but I don't have any more precise idea.

It was reported on the Orca list, and I'm rather surprised that we're only just now hitting it....

It depends on the actual configuration. Actually in my setup I had to disable the symbols.dic processing, which was replacing it with an actual word.

sthibaul commented 1 year ago

Ah, in your setup it's actually a mere < that is sent to the synth. But that's the same behavior as I observe.

joanmarie commented 1 year ago

It just occurred to me that maybe the reason we're seeing it now (at least as far as Orca is concerned) is that Orca is no longer converting the symbol before it reaches speech-dispatcher. With all the other changes (new keyhandling, de-pyatspification, etc.), I had completely spaced that out.

jeremyp3 commented 1 year ago

for your information, I had already reported this problem to @gcasse

I received a patch from him to correct the voxin crash. https://oralux.org/soft/tmp/3.3rc6/fix_spelling_char_lower_than.tgz

An update is now available pending a future future delivery from voxin. This update assumes that version 3.3rc6 has already been installed under /opt . This quick fix will prevent voxin from looping endlessly when reading of this character. A real fix will require more time and will be included in a future delivery of voxin.

joanmarie commented 1 year ago

for your information, I had already reported this problem to @gcasse

I received a patch from him to correct the voxin crash.

Good to know there's already a fix. Now we need it in a release.

jeremyp3 commented 1 year ago

@joanmarie I edited my comment to include the patch link, if that helps.

gcasse commented 1 year ago

Hello,

The 'lower than' character can crash voxin (+ IBM TTS or Vocalizer Embedded).

This issue will be fixed in the next release (expected before mid october); today a partial fix for version 3.3rc6 and x86_64 is available at: https://oralux.org/soft/tmp/3.3rc6/fix_spelling_char_lower_than.tgz

Its sha256sum is: c8f9c95d959efa3c71ed7998f5ddcce2fbf453b1d0f07dc9f79b5b1e34c2190f

Best wishes, Gilles

Le lun. 2 oct. 2023 à 16:11, jeremyp3 @.***> a écrit :

@joanmarie https://github.com/joanmarie I edited my comment to include the patch link, if that helps.

— Reply to this email directly, view it on GitHub https://github.com/brailcom/speechd/issues/857#issuecomment-1743093298, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB7ZH2MXY6WVQYBYJDQQMKTX5LDSZAVCNFSM6AAAAAA5PH7RSGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONBTGA4TGMRZHA . You are receiving this because you were mentioned.Message ID: @.***>

brailcom / speechd