Hack fix to #51 by monkey-patching the eSpeak speak() method.

I'm not totally sure why the fix works. The fix is to save the rate before the underlying speak() call and reset the rate after the call. However, the bug involves `<break>` and the speech isn't finished (probably not even started) by the time the reset happens. But it does work...

There is no other change to the eSpeak method, so hopefully there is no interference with other calls to it.

Note: because it was easiest, the monkey-patching is done on a espeak instance, not on the class itself. This is also a little safer because it won't happen unless someone is reading math.

Author: NSoiffer

NVDA-addon/addon/globalPlugins/MathCAT/MathCAT.py

@@ -83,32 +83,31 @@ def getLanguageToUse(mathMl:str) -> str:
 def  ConvertSSMLTextForNVDA(text:str, language:str="") -> list:
     # MathCAT's default rate is 180 wpm.
     # Assume that 0% is 80 wpm and 100% is 450 wpm and scale accordingly.
-    # log.info(f"Speech str: '{text}'")
+    # log.info(f"\nSpeech str: '{text}'")
     if language == "":    # shouldn't happen
         language = "en"   # fallback to what was being used
-
     mathCATLanguageSetting = "en"   # fallback in case GetPreference fails for unknown reasons
     try:
         mathCATLanguageSetting = libmathcat.GetPreference("Language")
     except Exception as e:
         log.error(e)    
 
     synth = getSynth()
+    _monkeyPatchESpeak()
     wpm = synth._percentToParam(synth.rate, 80, 450)
     breakMulti = 180.0 / wpm
-    synthConfig = config.conf["speech"][synth.name]
     supported_commands = synth.supportedCommands
     use_break = BreakCommand in supported_commands
     use_pitch = PitchCommand in supported_commands
-    use_rate = RateCommand in supported_commands
-    use_volume = VolumeCommand in supported_commands
+    # use_rate = RateCommand in supported_commands
+    # use_volume = VolumeCommand in supported_commands
     use_phoneme = PhonemeCommand in supported_commands
     # as of 7/23, oneCore voices do not implement the CharacterModeCommand despite it being in supported_commands
     use_character = CharacterModeCommand in supported_commands and synth.name != 'oneCore'
     out = []
     if mathCATLanguageSetting != language:
         try:
-            log.info(f"Setting language to {language}")
+            # log.info(f"Setting language to {language}")
             libmathcat.SetPreference("Language", language)
             out.append(LangChangeCommand(language))
         except Exception as e:
@@ -125,7 +124,7 @@ def  ConvertSSMLTextForNVDA(text:str, language:str="") -> list:
             if use_character:
                 out.extend((CharacterModeCommand(True), ch, CharacterModeCommand(False)))
             else:
-                out.extend((" ", "eigh" if ch=="a" else ch, " "))
+                out.extend((" ", "eigh" if ch=="a" and language=="en" else ch, " "))
         elif m.lastgroup == "beep":
             out.append(BeepCommand(2000, 50))
         elif m.lastgroup == "pitch":
@@ -156,7 +155,6 @@ def  ConvertSSMLTextForNVDA(text:str, language:str="") -> list:
             out.append(LangChangeCommand(None))
         except Exception as e:
             log.error(e) 
-        
     # log.info(f"Speech commands: '{out}'")
     return out
 
@@ -343,7 +341,6 @@ def __init__(self):
             speech.speakMessage(_("MathCAT initialization failed: see NVDA error log for details"))
         self._language = ""
 
-
     def getSpeechForMathMl(self, mathml: str):
         try:
             self._language = getLanguageToUse(mathml)
@@ -398,3 +395,91 @@ def getBrailleForMathMl(self, mathml: str):
     def interactWithMathMl(self, mathml: str):
         MathCATInteraction(provider=self, mathMl=mathml).setFocus()
         MathCATInteraction(provider=self, mathMl=mathml).script_navigate(None)
+
+CACHED_SYNTH = None
+def _monkeyPatchESpeak():
+    global CACHED_SYNTH
+    currentSynth = getSynth()
+    if currentSynth.name != "espeak" or CACHED_SYNTH == currentSynth:
+        return  # already patched
+
+    CACHED_SYNTH = currentSynth
+    currentSynth.speak = patched_speak.__get__(currentSynth, type(currentSynth))
+
+
+from speech.types import SpeechSequence
+from speech.commands import (
+	IndexCommand,
+	CharacterModeCommand,
+	LangChangeCommand,
+	BreakCommand,
+	PitchCommand,
+	RateCommand,
+	VolumeCommand,
+	PhonemeCommand,
+)
+def patched_speak(self, speechSequence: SpeechSequence):  # noqa: C901
+    from synthDrivers import _espeak
+    # log.info(f"patched_speak input: {speechSequence}")
+    textList: List[str] = []
+    langChanged = False
+    prosody: Dict[str, int] = {}
+    # We output malformed XML, as we might close an outer tag after opening an inner one; e.g.
+    # <voice><prosody></voice></prosody>.
+    # However, eSpeak doesn't seem to mind.
+    for item in speechSequence:
+        if isinstance(item,str):
+            textList.append(self._processText(item))
+        elif isinstance(item, IndexCommand):
+            textList.append("<mark name=\"%d\" />"%item.index)
+        elif isinstance(item, CharacterModeCommand):
+            textList.append("<say-as interpret-as=\"characters\">" if item.state else "</say-as>")
+        elif isinstance(item, LangChangeCommand):
+            langChangeXML = self._handleLangChangeCommand(item, langChanged)
+            textList.append(langChangeXML)
+            langChanged = True
+        elif isinstance(item, BreakCommand):
+            textList.append(f'<break time="{item.time}ms" />')
+        elif type(item) in self.PROSODY_ATTRS:
+            if prosody:
+                # Close previous prosody tag.
+                textList.append("</prosody>")
+            attr=self.PROSODY_ATTRS[type(item)]
+            if item.multiplier==1:
+                # Returning to normal.
+                try:
+                    del prosody[attr]
+                except KeyError:
+                    pass
+            else:
+                prosody[attr]=int(item.multiplier* 100)
+            if not prosody:
+                continue
+            textList.append("<prosody")
+            for attr,val in prosody.items():
+                textList.append(' %s="%d%%"'%(attr,val))
+            textList.append(">")
+        elif isinstance(item, PhonemeCommand):
+            # We can't use str.translate because we want to reject unknown characters.
+            try:
+                phonemes="".join([self.IPA_TO_ESPEAK[char] for char in item.ipa])
+                # There needs to be a space after the phoneme command.
+                # Otherwise, eSpeak will announce a subsequent SSML tag instead of processing it.
+                textList.append(u"[[%s]] "%phonemes)
+            except KeyError:
+                log.debugWarning("Unknown character in IPA string: %s"%item.ipa)
+                if item.text:
+                    textList.append(self._processText(item.text))
+        else:
+            log.error("Unknown speech: %s"%item)
+    # Close any open tags.
+    if langChanged:
+        textList.append("</voice>")
+    if prosody:
+        textList.append("</prosody>")
+    text=u"".join(textList)
+    # Added saving old rate and then resetting to that -- work around for https://github.com/nvaccess/nvda/issues/15221
+    # I'm not clear why this works since _set_rate() is called before the speech is finished speaking
+    oldRate = getSynth()._get_rate()
+    _espeak.speak(text)
+    getSynth()._set_rate(oldRate)