Refactor wake phrase handling: introduce REQUIRE_WAKE_WORD configuration, update documentation, and adjust assistant behavior for improved user experience.

bolt_assistant/__main__.py

@@ -1,4 +1,5 @@
-"""Bolt Assistant: say "Hey Bolt", ask, hear the answer.
+"""Bolt Assistant: talk while the mic is open, hear the answer. (Setting
+REQUIRE_WAKE_WORD=true puts "Hey Bolt" back in front of each request.)
 
     python -m bolt_assistant                 listen on the default mic
     python -m bolt_assistant --verbose       also show what was heard but ignored
@@ -89,7 +90,8 @@ def main(argv=None) -> int:
         return 0
 
     if "--type" in argv:  # the whole pipeline minus the microphone
-        print("Type what you'd say (including the wake phrase). Ctrl+C to quit.")
+        hint = " (including the wake phrase)" if assistant.wake_required else ""
+        print(f"Type what you'd say{hint}. Ctrl+C to quit.")
         try:
             while True:
                 assistant.heard(input("say> "))

bolt_assistant/assistant.py

@@ -1,7 +1,10 @@
-"""The assistant: hear the wake phrase, take the request, answer out loud.
+"""The assistant: hear a request, answer out loud.
 
     idle      every utterance the local detector passes on is transcribed by
-              the server; only ones that open with a wake phrase count
+              the server and, by default, is itself the request — the mic
+              being open is the activation. With REQUIRE_WAKE_WORD on, only
+              ones that open with a wake phrase count instead
+                "what's the time"            -> ask "what's the time"
                 "Hey Bolt, what's the time"  -> ask "what's the time"
                 "Hey Bolt"                   -> chime, go to `command`
     command   the next thing said (within COMMAND_WAIT_SEC) is the request
@@ -88,6 +91,7 @@ class Assistant:
         clock: Callable[[], float] = time.monotonic,
         wake_words=None,
         anywhere_words=None,
+        wake_required: Optional[bool] = None,
         verbose: bool = False,
     ):
         self._transcription = transcription
@@ -98,6 +102,7 @@ class Assistant:
         self.tools = tools or Tools(log=log)
         self.wake_words = wake_words or config.WAKE_WORDS
         self.anywhere_words = config.ANYWHERE_WAKE_WORDS if anywhere_words is None else anywhere_words
+        self.wake_required = config.REQUIRE_WAKE_WORD if wake_required is None else wake_required
         self.segmenter = Segmenter(config.VAD_RMS_THRESHOLD, config.VAD_START_FRAMES,
                                    config.VAD_SILENCE_END_SEC, config.MIN_UTTERANCE_SEC,
                                    config.MAX_UTTERANCE_SEC)
@@ -132,6 +137,8 @@ class Assistant:
             self.handle_transcripts()
 
     def listening_banner(self) -> str:
+        if not self.wake_required:
+            return "Listening - anything you say is a request. Ctrl+C to quit."
         return 'Listening for "Hey Bolt" (or "Thunderbolt", "Hey Assistant"). Ctrl+C to quit.'
 
     def process(self, frame: bytes) -> None:
@@ -210,6 +217,12 @@ class Assistant:
         self._check_deadline()
         woke, rest = find_wake(text, self.wake_words, self.anywhere_words)
         if self.state == "idle":
+            if not self.wake_required and not woke:
+                # Always listening: the mic being open is the activation, so
+                # whatever it managed to hear is the request.
+                self._follow_ups = 0
+                self.ask(text)
+                return
             if not woke:
                 if self.verbose:
                     self.log(f"  (not for me) {text}")

bolt_assistant/config.py

@@ -56,11 +56,18 @@ MIC_DEVICE = _device(os.environ.get("MIC_DEVICE", ""))
 SPEAKER_DEVICE = _device(os.environ.get("SPEAKER_DEVICE", ""))
 
 # ── wake phrases ────────────────────────────────────────────────────────────
+# With REQUIRE_WAKE_WORD off (the default) the mic itself is the activation:
+# every utterance the local detector passes on is treated as a request, and
+# whatever is said near an open mic gets answered. Set it to true to go back
+# to needing a phrase before each request.
+REQUIRE_WAKE_WORD = _flag("REQUIRE_WAKE_WORD", "false")
+
 # Heard through the server's speech-to-text, so these are matched against
 # *text*. A phrase counts only near the start of what was said ("hey bolt,
 # what's the time"), not buried in a sentence ("tighten that bolt"). Speech
 # recognisers often write "Bolt" as "bolts"/"volt"/"bold"; those spellings
-# only count after "hey"/"ok", where they can't be ordinary words.
+# only count after "hey"/"ok", where they can't be ordinary words. Only used
+# when REQUIRE_WAKE_WORD is on (a phrase said anyway is still stripped).
 
 WAKE_WORDS = [w.strip().lower() for w in os.environ.get(
     "WAKE_WORDS",