This is part of the (complicated) question of how to deal with ambiguity in pronunciation and part-of-speech tags. If there is a "real" pos-tagger at work, or if pos-tags are set by the user, these should be used to select pronunciation. Otherwise the lexicon might just as well set it. Changed (test) to set part-of-speech in the same way as pronunciation has been set before: Use the first entry in the lexicon, unless another one is explicitly set as preferred.

Words with > (glottal stop) in transcription can now be played.

I think Hanna has to upload a new version to Docker hub.
But it should also work to run the wikispeech server directly from the source cloned from github (I'm doing that).

Tokenisation problem in Arabic marytts. Is it even wrong? The input is very unlikely for an Arabic synthesis. However there could definitely be a better error message!
Maybe the problem is in handling the various types of quotes used in Arabic? Or just that quotes work with Arabic script input but not with Latin script.
This works:
وكتابه "نهاية الإيجار في دراية الإعجاز" يعتبر من المراجع البلاغية المهمة.
But the example in this ticket doesn't and neither does this:
"Bob Dylan"

Yes, English marytts does this. English Flite says "hash" followed by the number. Swedish marytts actually just skips the whole "#N" expression.
So for now what you can do is to change the pronunciation using ssml.
But it also reraises a question we discussed before: Should there be a preprocessing step with e.g. regular expressions to modify input? Maybe a good idea.

Corrected in git source.
Needs more testing with various kinds of input.
The input text above now gives:

Changed in git source

changed to WARNING

No, not too big, the version on github now has a fix for it.

I'm using the current from github. But not too sure about when these parts were last changed, I think it was long ago, so maybe the problem isn't there at all. Could be an encoding issue?
Anyway the right thing to do isn't to replace tokens back to original form as an afterthought, as it is now, but to either send the tokens from the client or to make sure that the original input is kept and returned from the server.

Hm.. The example works for me (see json output below).
But that doesn't mean that it's actually solved in a sensible way. It's just a guess at what the client wants. Bound to go wrong in complicated cases like this, with text going in both directions and unusual punctuation characters.
The specific problem in your case is in the word ومرن ، which has an upside down comma after it, you can see the comma as a separate token in the output token list but not in the input.
So even if you can update so that we use the same version, and solve this problem, we should still think of a better way of solving the issue.

What call are you using for this?
It seems to work for me, so maybe we are doing it differently somehow. Or else my tests just happen to work. Can you give an example of a call that fails?

New version of flite_adapter.

In what way does it sound weird? I can't reproduce, what I've tested sounds fine, or at least the same no matter which textprocessor is used..
If it doesn't, yes that should be fixed!

Did you update lexdata, copy the new symboltable, restart pronlex server, update wikispeech, restart wikispeech server?
In that case I must have forgotten to commit something.

Cronjob or just test script that calls the server every couple of seconds or so?

Symboltable added to git/lexdata. Mapper added to voice_config for Norwegian.
Works for me now.

The lexicon wasn't setup right in wikispeech_server/voice_config
But now that it is, it's worse, all transcription fields are now wrong and there seems to be no mapper available for norwegian marytts sampa .
Nothing should have worked before so in that sense it's better now. Right place to look, anyway, at the mapper.

The error message is very confusing - but what actually happens is that marytts returns an error, because there is an unknown symbol in the transcription. "." instead of "-" as syllable separator.
The mapper isn't called in that case. Why not?
The transcription in the table on the left is mapped, and in the "ssml transcription field". But not in the "lexicon entries" field.

Added info message "The wikispeech server can be accessed at http://localhost:10000/wikispeech"

The reason for this is described at http://flask.pocoo.org/docs/0.12/quickstart/#a-minimal-application under "Externally visible server".
It's no longer needed and can anyway be done in different ways, so I removed it.
However, when I ran the server with "host=0.0.0.0", it did serve at that URL, as well as 127.0.0.1, and "localhost". So I'm not sure what the problem was?

Edit "audio_url_prefix" in the config file to do this.
We could change so that the audio link is always at the request host, but sometimes this is not right - if the request is being proxied through another webserver, for instance.

Yes I could replicate on morf.se. Updated the offending line. Works now on morf.se, commited bugfix to github.

I can't replicate this error.
http://localhost:10000/wikispeech/?lang=sv&input=tilltugg,%20i%20bakverk
and
http://localhost:10000/wikispeech/?lang=sv&input=tillverkningen
both work fine for me.

Fixed in a very simple way by inserting a space before the hyphen, in marytts_adapter.py
Not sure that's actually good enough..

The marytts textprocessing uses icu locale files for number expansion. In words like "sexhundra", it returns a soft hyphen "sex\xadhundra". This causes lexicon lookup to fail. Simple fix is just to remove "\xad" if it is found before doing lexicon lookup.

Updated handling of test errors.
TODO: set up a minimal flite voice.

There is now a TEST version of an Arabic lexicon, to be used for foreign names and the like.
It only contains four entries at the moment, so not much real use, but it means that it is now possible to add words to the lexicon to improve pronunciation of a wikipedia article.

There is now a branch "packaging" in https://github.com/stts-se/wikispeech_mockup, that I think contains everything needed for fairly simple setup of the ws server. Lexicon and marytts servers are separate, as previously.

Closing this. But that does not mean that there are no possible future issues.

Closing this. But that does not mean that there are no possible future issues.

This is not a subtask of T122157

It is possible to choose between voices.