Similarly as on linux, there isn't enough control of the speech
queue to properly implement INTERRUPT_NO_REPEAT. So since this
commit we use our own queuing and use SAPI to speak each speech.
This is done outside the main thread.
Before I used SPD to queue messages and I had a copy of the queue,
so I could requeue everything when resume is called(). But more
control of the queue is needed, so I don't use the SPD's queue
and instead start speeches from my queue one by one from another
thread.
INTERRUPT_NO_REPEAT now behaves as described in the documentation
The NSSpeechSynthesizer is asynchronous and does not immediately start, pause,
or stop the speech. As a result querrying the state of the NSSpeechSynthesizer
does not alwats return the expected result (for example isSpeaking may not
yet been true just after we requested starting to speak). So instead the
TextToSpeechManager on macOS keeps track of the state itself.
It seems like, that at least some versions of speech-dispatcher
aren't able to successfuly pause and resume. For me, when trying
to pause, it still finishes the speech just being said instead
of pausing it and then it puts it at the end of the speech queue
with some speech-dispatcher internal commands added to it, which
are also hearable.
There is no way to find out where the speech ended when calling
pause, so it is just stopped and when resume is called it is
read from it's start again.
* Delete multiple empty rows
* Make getVolume non-virtual and leave just the implementation
in base class
* Resolve warning about signed / unsigned comparison in
gui-manager
* Clear availableVoices when updating voices on linux
* By default set language to transMan language on windows
(if the transMan is available)
* Remove freeVoices method from Windows ttsMan, it isn't needed
anymore
Voice is changed when changing language on windows, so when poping
state, the voice, that should get set has to be saved before
changing the language.
The speech shouldn't continue when changing state, so it is stopped
in pushState and popState.
Main changes are:
* Save age information about each voice
* Remove Sample TTS Voice from available voices, because it
basicaly cannot speak.
* Stop speech after silently playing the test speech (the last
voice could be heard speaking after the volume got
restored to its original value)
* Remove voice data freeing from freeVoices, because it gets
freed automaticaly
- Add comment to tts initialization on Windows
- Correctly free the voicesInfo in linux ttsMan
- Remove popState method from linux-text-to-speech.h and
windows-text-to-speech.h
- Add tts to help in configure
- Refactor language setting in gui-manager.cpp
It counted with english being the default language in
ttsMan constructors, which isn't true anymore.
The ScummVM was crashing because of an assert, when there was less
voices availaible, than what was set in the ConfMan.
Now the voice just falls back to 0th voice, if there are not
enough voices.
Unfortunatedly the encoding used by ScummVM breaks the
speech-dispatcher, so after trying to say non-ascii character
the connection has to be restarted. So for now I am restricting
the GUI TTS to english only.
RFC 2616 states that HTTP headers are not case-sensitive and also allows
arbitrary number of whitespace characters around header value. Previous
implementation was dependant on headers to be in "Title-Case" and to
have only one space before header value. That has lead to cloud sync
failure on Debian x64 (user's network environment was probably the
reason though).
This commit adds a new method, which parses headers name-value pairs
into HashMap. To ensure case-insensitivity, all headers names are
converted to lowercase, and thus code that uses this method should
specify headers in lowercase. All usages of raw headers contents were
updated to use this method.
With some older versions of SDL1, the SDL_iconv_string takes
char * instead of const char * as it's argument. This should
fix the build issue with gp2xwiz.
Because of how cyrilic transliteration and UTF-32 is handled on
Windows, it was unfortunately possible to get into an infinite
loop of conversions. The string would get converted to UTF-32
when transliterating, but because windows backend conversion
cannot convert from UTF-32, it would use Common::Ustr to convert
it to UTF-8, which would again get converted to UTF-32 when
transliterating and so on.
UTF-32 is used in transliteration in Common::Encoding, so it is
pretty important encoding and Windows should be the only thing,
that cannot convert it.
This removes the usage of memset to clear complex structures and replaces
them with constructor methods for the structures which will be executed
when these are instantiated.
If user doesn't have a "saves" folder, listing it as a first step of
syncing save files would result in 404 from OneDrive.
OneDriveTokenRefresher handles token-related errors (401), so when it
meets 404, it calls its finishError method. But because there was some
strange behaviour from OneDrive with sending invalid JSON, this method
tries fixing JSON and parsing it again. If it is valid, it calls
non-error method again, and in result we get stack overflow.
In order to fix that, I've added a non-JSON prefix "<irrecoverable>", so
finishError won't be able to parse JSON and thus won't call finishJson
again. Saves syncing callback does check string contents apart from
trying to parse JSON, so it still works and correctly handles the
situation when "saves" directory is missing. But, if needed, code can be
updated to search for the prefix I've added and remove it before parsing
original JSON.
The warning emitted here was due to debug() call with an empty format
string, so this could have been fixed by replacing this with "%s", "".
However, this change should be better since it avoids the duplication
of the number of file check over several lines and reduces the string
duplication i.e. DRY improvements.
In PR#1754 we've discussed and decided to ignore hidden (having a name
starting with '.') files while syncing saves or downloading game files.
This commit adds a CloudManager method to test whether file should be
ignored, and this method could be extended later if we need to ignore
some other specific file names.
- added missing 'd' in "%d" in SavesSyncRequest;
- removed trailing ',' in enum in gui/options.h;
- fixed #endif to have // before USE_LIBCURL in gui/options.h.
- StorageWizardDialog is removed, along with bmps it was using;
- EditTextWidget now accepts custom font in constructor;
- ScrollContainer scrollbar now jumps to top when content height changes
so it's "overscrolled";
- IndexPageHandler now does not awaits for `code` GET-parameter, as
local webserver is no longer used to connect Storages;
- CloudManager and all corresponding Storages are updated to support
disconnecting and to notify about successful connection.
This mostly affects OneDrive saves syncing, because it is the only cloud
provider to return 0 as timestamp of (non-)created file. 0 is treated as
EOF in /saves/timestamps file, thus all timestamps after such 0
timestamps were ignored and files were reuploaded as being "new". This
commit also adds more verbose debug information on SavesSyncRequest
decisions making for easier debugging.
ConnectionManager used to poll Cloud requests every second, while curl
requests were polled every 1/20th of a second. If curl request was over
at, say, frame #21, corresponding Cloud request would've only work with
that at frame #40 (950 ms later), which was making everything
cloud-related slower than it could be. This commit fixes it by making
Cloud polling period the same as curl polling period, and additionally
raises FPS to 25.
Something changed and old API endpoint "api.onedrive.com" now does not
work. The other one, "graph.microsoft.com", does, but there were some
other changes in JSON it returns. These changes are also in this commit.
While refreshing access_token, some cloud providers also pass a new
refresh_token. Google Drive does not, and accepts the same refresh_token
next time. These changes allow this to happen.
The GLSL version code has been taken from ResidualVM. The variable
'texture' is now a reserved keyword in GLSL 3.00, so it has been
renamed. This fixes compilation issues in AmigaOS4 (PR 1554).
This is necessary for properly identifying the Return key pressed from
the software or a hardware keyboard, and this was erronously removed
in commit e5709ed.
The cursor position was overflowing a signed 16-bits integer once
multiplied with MULTIPLIER when using a resolution such as 2560x1440.
It would be nice changing this code to make more sense, sadly it is
thightly coupled with platform specific subclasses.
Fixes#10996.
This implementation now matches previous ShellExecute() calls, so we are
now consistent in that regard. It also
silences a warning in Mingw complaining about a type mismatch.
