Reload --vosk-grammar-file JSON on SIGHUP

[KJ7LNW]

This commit adds support for reloading the grammar configuration on SIGHUP. This is accomplished by moving the recognizer initialization into its own function so that the signal handler can call it on reload.

A reload of the recognizer is fast because it does not reload the model and only applies the grammar.

Signed-off-by: Eric Wheeler vosk-git@z.ewheeler.org

[KJ7LNW]

fixed merge conflict

[KJ7LNW]

Fixed mypy --strict with return type -> None.

BTW, the diff shown by github isn't pretty because of the indented code (whitespace), but the logic change is small:

Here is a diff ignoring whitespace:

# git show -w HEAD
diff --git a/nerd-dictation b/nerd-dictation
index e5d8196..c4adc49 100755
--- a/nerd-dictation
+++ b/nerd-dictation
@@ -944,22 +944,30 @@ def text_from_vosk_pipe(

     vosk.SetLogLevel(-1)

+    # Allow for loading the model to take some time:
+    if verbose >= 1:
+        sys.stderr.write("Loading model...\n")
+    model = vosk.Model(vosk_model_dir)
+
+    rec = None
+
+    def vosk_load_recognizer() -> None:
+        nonlocal vosk_grammar_file
+        nonlocal rec
+
         if not vosk_grammar_file:
             grammar_json = ""
         else:
             with open(vosk_grammar_file, encoding="utf-8") as fh:
                 grammar_json = fh.read()

-    # Allow for loading the model to take some time:
-    if verbose >= 1:
-        sys.stderr.write("Loading model...\n")
-    model = vosk.Model(vosk_model_dir)
-
         if grammar_json == "":
             rec = vosk.KaldiRecognizer(model, sample_rate)
         else:
             rec = vosk.KaldiRecognizer(model, sample_rate, grammar_json)

+    vosk_load_recognizer()
+
     if verbose >= 1:
         sys.stderr.write("Model loaded.\n")

@@ -1135,6 +1143,7 @@ def text_from_vosk_pipe(
         if verbose >= 1:
             sys.stderr.write("Reload.\n")
         process_fn("")
+        vosk_load_recognizer()

     import signal

[KJ7LNW]

@ideasman42,

Just realized I still have mypy issues to work out on this PR, but I'm not sure what to do.

Scoping rec = None outside of the new vosk_load_recognizer() function causes mypy to complain as follows because None is not properly typed:

nerd-dictation:1052: error: Item "None" of "Optional[Any]" has no attribute "FinalResult"
nerd-dictation:1069: error: Item "None" of "Optional[Any]" has no attribute "PartialResult"
nerd-dictation:1100: error: Item "None" of "Optional[Any]" has no attribute "Reset"
nerd-dictation:1208: error: Item "None" of "Optional[Any]" has no attribute "AcceptWaveform"

I tried a few things but I'm not very familiar with Python's typing. For example, changing rec = None to rec = vosk.KaldiRecognizer(model, sample_rate) and then still calling vosk_load_recognizer() to load the JSON grammar "fixes" the typing issue, but the extra call to vosk.KaldiRecognizer() seems unnecessary.

Do you know what should be done here?

[ideasman42]

In this case mypy needs to know that it's not None.

• If there is no chance of it being None, you can do assert rec is not None to quiet the warning.
• Or if there is no reason for the value to be None, remove the Optional[..] type.

[KJ7LNW]

The only reason I set rec = None is to scope outside of the vosk_load_recognizer() function.

Questions:

• Is there a way to scope rec without assigning it to None?
• How do you remove the Optional[..] type?
• I've pushed a commit with a few assert lines that fix the mypy issue so you can see it. Do you think is the best way to go here? If so then let me know and I will squash+repush.