Bolt Assistant: wake-phrase voice assistant for Bolt

bolt_assistant/wake.py

@@ -0,0 +1,50 @@
+"""Finding the wake phrase in a transcript, and what was said after it.
+
+    >>> find_wake("Hey Bolt, what's the weather?", WAKE)
+    (True, "what's the weather?")
+    >>> find_wake("tighten that bolt", WAKE)
+    (False, '')
+
+Rules, all aimed at not waking up on ordinary conversation:
+
+- The phrase has to **open** what was said. Only filler may come before it
+  ("um, hey Bolt", "okay so Bolt"), at most two words of it, so "tighten
+  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.
+"""
+
+from __future__ import annotations
+
+import re
+from typing import Iterable, Tuple
+
+_WORD = re.compile(r"[a-z0-9']+")
+_LEAD_SLACK = 2  # filler words allowed before the wake phrase
+_FILLER = {"um", "uh", "erm", "er", "so", "oh", "ok", "okay", "well", "and", "right",
+           "alright", "yeah", "hmm", "hm", "ah", "hi", "hey", "yo"}
+
+
+def _words(text: str) -> list:
+    """Lower-case words with their character spans in the original text."""
+    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)."""
+    words = _words(transcript or "")
+    if not words:
+        return False, ""
+    candidates = sorted({tuple(_WORD.findall(p.lower())) for p in phrases if p.strip()},
+                        key=len, reverse=True)
+    for start in range(min(_LEAD_SLACK + 1, len(words))):
+        if any(word not in _FILLER for word, _, _ in words[:start]):
+            break  # a real word came first: this isn't addressed to Bolt
+        for phrase in candidates:
+            span = words[start:start + len(phrase)]
+            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 False, ""