Updates to the assistants overall abilities, like always listening, and only responding on messages when activated.

.env.example

@@ -10,6 +10,8 @@ ELEVENLABS_VOICE_ID=
 
 # Wake phrases, comma-separated (matched on the transcript, at the start of what's said)
 #WAKE_WORDS=hey bolt,bolt,thunderbolt,hey thunderbolt,ok bolt,okay bolt,hey assistant
+# Words that count anywhere in what's said; the whole sentence becomes the request
+#ANYWHERE_WAKE_WORDS=assistant
 
 # What Bolt may do on this machine when he relays a command:
 #   off | read (default: list and read files) | all (shell commands and file changes too)

README.md

@@ -73,6 +73,8 @@ See `.env.example`. The ones you're most likely to change:
 - `VAD_RMS_THRESHOLD`: raise it if background noise keeps getting transcribed, lower it if you
   have to raise your voice.
 - `WAKE_WORDS`: the phrases, comma-separated.
+- `ANYWHERE_WAKE_WORDS`: words that count anywhere in a sentence, not just at the start
+  (default `assistant`: "what's the weather, assistant?"). Blank to turn off.
 - `COMMAND_WAIT_SEC`, `FOLLOW_UP_SEC`: how long it waits after "Hey Bolt", or after asking you
   something.
 

bolt_assistant/assistant.py

@@ -70,6 +70,7 @@ class Assistant:
         log: Callable[[str], None] = print,
         clock: Callable[[], float] = time.monotonic,
         wake_words=None,
+        anywhere_words=None,
         verbose: bool = False,
     ):
         self._transcription = transcription
@@ -79,6 +80,7 @@ class Assistant:
         self.log, self.clock, self.verbose = log, clock, verbose
         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.segmenter = Segmenter(config.VAD_RMS_THRESHOLD, config.VAD_START_FRAMES,
                                    config.VAD_SILENCE_END_SEC, config.MIN_UTTERANCE_SEC,
                                    config.MAX_UTTERANCE_SEC)
@@ -152,7 +154,7 @@ class Assistant:
         if not text:
             return
         self._check_deadline()
-        woke, rest = find_wake(text, self.wake_words)
+        woke, rest = find_wake(text, self.wake_words, self.anywhere_words)
         if self.state == "idle":
             if not woke:
                 if self.verbose:

bolt_assistant/audio.py

@@ -30,7 +30,42 @@ def open_mic(on_frame: Callable[[bytes], None]):
 
 def play(chunks: Iterable[bytes], sample_rate: int = config.TTS_SAMPLE_RATE,
          should_stop: Callable[[], bool] = lambda: False) -> bool:
-    """Play PCM16 mono chunks as they arrive. False if stopped early."""
+    """Play PCM16 mono chunks as they arrive. False if stopped early.
+
+    Through `pacat`, not sounddevice: PortAudio only sees raw ALSA devices,
+    so a Bluetooth or PipeWire sink is invisible to it and the audio goes
+    somewhere nobody can hear. pacat talks to the audio server, so it
+    follows the system's default sink like every other player.
+    """
+    import shutil
+    import subprocess
+
+    if shutil.which("pacat") is None:
+        return _play_sounddevice(chunks, sample_rate, should_stop)
+
+    proc = subprocess.Popen(
+        ["pacat", "--rate", str(sample_rate), "--channels", "1",
+         "--format", "s16le", "--latency-msec", "60"],
+        stdin=subprocess.PIPE, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
+    )
+    try:
+        for chunk in chunks:
+            if should_stop():
+                return False
+            proc.stdin.write(chunk)
+        proc.stdin.close()
+        proc.wait(timeout=10)
+    except (BrokenPipeError, OSError):
+        return False
+    finally:
+        if proc.poll() is None:
+            proc.kill()
+    return True
+
+
+def _play_sounddevice(chunks: Iterable[bytes], sample_rate: int,
+                      should_stop: Callable[[], bool]) -> bool:
+    """Fallback for a machine without pacat."""
     import sounddevice as sd
 
     with sd.RawOutputStream(samplerate=sample_rate, channels=1, dtype="int16",

bolt_assistant/config.py

@@ -68,6 +68,10 @@ WAKE_WORDS = [w.strip().lower() for w in os.environ.get(
     "hey assistant,ok assistant,okay assistant,"
     "hey bolts,hey volt,hey bold,hey boat,okay volt,hey thunder bolt,thunder bolt",
 ).split(",") if w.strip()]
+# These count anywhere in what was said, not just at the start ("what's the
+# weather, assistant?"), and the whole sentence is sent as the request.
+ANYWHERE_WAKE_WORDS = [w.strip().lower() for w in os.environ.get(
+    "ANYWHERE_WAKE_WORDS", "assistant").split(",") if w.strip()]
 
 # ── listening (local voice-activity detection, before anything is sent) ────
 # Nothing leaves the machine until this says someone is talking: silence and

bolt_assistant/wake.py

@@ -12,6 +12,9 @@ Rules, all aimed at not waking up on ordinary conversation:
   that bolt" and "the bolt on the door" never count.
 - Whole words only: "boltage" is not "bolt".
 - Longest phrase wins, so "hey thunderbolt" isn't read as "hey" + garbage.
+- Except *anywhere* words ("assistant"): one of those anywhere in what was
+  said counts, and the whole transcript is the request ("set a timer,
+  assistant" asks "set a timer, assistant").
 """
 
 from __future__ import annotations
@@ -30,10 +33,11 @@ def _words(text: str) -> list:
     return [(m.group(0), m.start(), m.end()) for m in _WORD.finditer(text.lower())]
 
 
-def find_wake(transcript: str, phrases: Iterable[str]) -> Tuple[bool, str]:
-    """(heard, remainder): whether a wake phrase opens *transcript*, and the
-    request that followed it in the same breath ("" if it was just the
-    phrase)."""
+def find_wake(transcript: str, phrases: Iterable[str],
+              anywhere: Iterable[str] = ()) -> Tuple[bool, str]:
+    """(heard, remainder): whether a wake phrase opens *transcript* (or an
+    *anywhere* word appears in it), and the request that came with it in the
+    same breath ("" if it was just the phrase)."""
     words = _words(transcript or "")
     if not words:
         return False, ""
@@ -47,4 +51,15 @@ def find_wake(transcript: str, phrases: Iterable[str]) -> Tuple[bool, str]:
             if len(span) == len(phrase) and all(w == p for (w, _, _), p in zip(span, phrase)):
                 rest = transcript[span[-1][2]:]
                 return True, rest.lstrip(" \t,.;:!?-—").strip()
+    return _find_anywhere(transcript, words, anywhere)
+
+
+def _find_anywhere(transcript: str, words: list, anywhere: Iterable[str]) -> Tuple[bool, str]:
+    said = [w for w, _, _ in words]
+    for phrase in {tuple(_WORD.findall(p.lower())) for p in anywhere if p.strip()}:
+        n = len(phrase)
+        if any(tuple(said[i:i + n]) == phrase for i in range(len(said) - n + 1)):
+            if all(w in phrase or w in _FILLER for w in said):
+                return True, ""  # just the word on its own: wait for the request
+            return True, transcript.strip()
     return False, ""

tests/test_assistant.py

@@ -44,7 +44,6 @@ W = config.WAKE_WORDS
     ("tighten that bolt", (False, "")),
     ("the bolt on the door", (False, "")),
     ("I was bold", (False, "")),                       # the mishearing alone isn't a wake word
-    ("assistant set a timer", (False, "")),            # "assistant" needs the "hey"
     ("boltage is high", (False, "")),
     ("", (False, "")),
 ])
@@ -52,6 +51,20 @@ def test_find_wake(said, expected):
     assert find_wake(said, W) == expected
 
 
+@pytest.mark.parametrize("said, expected", [
+    ("assistant set a timer", (True, "assistant set a timer")),
+    ("What's the weather, assistant?", (True, "What's the weather, assistant?")),
+    ("can the assistant turn it up", (True, "can the assistant turn it up")),
+    ("Hey assistant. Set a timer", (True, "Set a timer")),  # at the start: stripped as before
+    ("Assistant.", (True, "")),                             # on its own: wait for the request
+    ("um, assistant", (True, "")),
+    ("my assistants are busy", (False, "")),                # whole words only
+    ("tighten that bolt", (False, "")),
+])
+def test_find_wake_anywhere(said, expected):
+    assert find_wake(said, W, config.ANYWHERE_WAKE_WORDS) == expected
+
+
 # ── voice-activity detection ────────────────────────────────────────────────
 
 def seg(**kw):