@gotcos/glasses-server 6.1.0

package/server/lib/fuzzy-correct.ts

@@ -0,0 +1,286 @@
+// Deterministic fuzzy correction for transcribed text.
+// Matches words against a canonical list of proper nouns (enrolled speakers +
+// profile vocabulary) and replaces close misses using length-scaled Levenshtein.
+//
+// Used by:
+//   - /api/transcribe       (short-form dictation / messaging)
+//   - /api/meeting/save     (batch post-processing, Phase 3.5 — via meeting.ts)
+//
+// Zero LLM cost, ~5ms on typical dictation length.
+//
+// SAFETY RULES (revised after QA 2026-04-10):
+//   1. Candidates: any 5+ char word (case-insensitive). This supports lowercase
+//      dictation output from Whisper (short /transcribe path) AND mixed-case
+//      proper nouns like product names and brands.
+//   2. Stop-word list blocks common English (prevents "right"→"Wright",
+//      "through"→"Brough"). Must be comprehensive — this is the primary defense.
+//   3. Tight distance thresholds (1 for 5-8 char, 2 for 9-11, 3 for 12+)
+//   4. Length difference ≤ 2 max (prevents long→short collapses)
+//   5. Target must also be a proper noun (starts uppercase in canonical list)
+//   6. Canonical min length 4 (allows short 4-character names as targets)
+
+import { getOwnerSpeakerLabel } from './profile.js'
+
+/** Compute Levenshtein edit distance between two lowercase strings. */
+export function levenshtein(a: string, b: string): number {
+  if (a === b) return 0
+  if (a.length === 0) return b.length
+  if (b.length === 0) return a.length
+  const m = a.length, n = b.length
+  let prev = new Array(n + 1).fill(0)
+  let curr = new Array(n + 1).fill(0)
+  for (let j = 0; j <= n; j++) prev[j] = j
+  for (let i = 1; i <= m; i++) {
+    curr[0] = i
+    for (let j = 1; j <= n; j++) {
+      const cost = a[i - 1] === b[j - 1] ? 0 : 1
+      curr[j] = Math.min(
+        prev[j] + 1,           // deletion
+        curr[j - 1] + 1,       // insertion
+        prev[j - 1] + cost,    // substitution
+      )
+    }
+    ;[prev, curr] = [curr, prev]
+  }
+  return prev[n]
+}
+
+// Common English words that must NEVER be fuzzy-replaced with proper nouns.
+// Deduplicated, alphabetically sorted for maintainability.
+// Covers: conversational fillers, verbs, nouns, function words, and common
+// first names (which the LLM correction pass can disambiguate contextually).
+// Real false positives this prevents:
+//   "Alright"→"Wright", "through"→"Brough", "details"→"Retail",
+//   "Shelley"→"Kelley", "right"→"Wright", "Christian"→"Cristian"
+const COMMON_ENGLISH_WORDS = new Set<string>([
+  // A
+  'about', 'above', 'absolutely', 'actually', 'admit', 'adopt', 'adult',
+  'after', 'again', 'against', 'agent', 'agree', 'ahead', 'album', 'alert',
+  'alike', 'alive', 'allow', 'allright', 'alone', 'along', 'aloud', 'alright',
+  'also', 'alter', 'although', 'always', 'among', 'amount', 'anger', 'angle',
+  'angry', 'another', 'answer', 'answers', 'anyone', 'anything', 'apart',
+  'around', 'array', 'arrive', 'arrow', 'aside', 'asked', 'asking', 'aspect',
+  // B
+  'basic', 'basically', 'beach', 'because', 'become', 'been', 'before', 'began',
+  'begin', 'behind', 'being', 'believe', 'below', 'bench', 'beside', 'better',
+  'between', 'beyond', 'birth', 'black', 'blame', 'blank', 'blast', 'blind',
+  'block', 'blood', 'board', 'booth', 'boost', 'bound', 'brain', 'brand',
+  'bread', 'break', 'breed', 'brief', 'bring', 'broad', 'broke', 'broken',
+  'brother', 'brought', 'brown', 'brush', 'built', 'bunch', 'business', 'burst',
+  // C
+  'calling', 'cannot', 'catch', 'caught', 'change', 'changed', 'chances',
+  'cheap', 'check', 'chest', 'chief', 'child', 'children', 'choice', 'choose',
+  'chose', 'christian', 'civil', 'claim', 'class', 'clean', 'clear', 'climb',
+  'clock', 'close', 'cloud', 'coach', 'coast', 'coming', 'company', 'could',
+  'count', 'course', 'court', 'cover', 'craft', 'crash', 'crazy', 'cream',
+  'create', 'created', 'crime', 'cross', 'crowd', 'crown', 'crude', 'curve',
+  'customer', 'customers', 'cycle',
+  // D
+  'daily', 'dance', 'dated', 'dealt', 'death', 'debut', 'definitely', 'delay',
+  'depth', 'despite', 'detail', 'details', 'doing', 'double', 'doubt', 'dozen',
+  'draft', 'drawn', 'dream', 'dress', 'drink', 'drive', 'driven', 'drove',
+  'during', 'dying',
+  // E
+  'early', 'earth', 'eight', 'either', 'email', 'emails', 'empty', 'enemy',
+  'enjoy', 'enough', 'entry', 'equal', 'error', 'especially', 'even', 'evening',
+  'event', 'every', 'everyone', 'everything', 'exactly', 'example', 'examples',
+  'except', 'extra',
+  // F
+  'faith', 'fancy', 'fault', 'feeling', 'fewer', 'field', 'fifth', 'final',
+  'finally', 'finish', 'finished', 'first', 'floor', 'focus', 'follow',
+  'followed', 'follows', 'following', 'force', 'forth', 'found', 'frame',
+  'fresh', 'friend', 'front',
+  // G
+  'getting', 'given', 'giving', 'going', 'gonna', 'gotta', 'grade', 'grant',
+  'great', 'green', 'group', 'groups', 'guess', 'guest', 'guide',
+  // H
+  'happened', 'happy', 'hard', 'having', 'head', 'heard', 'heavy', 'helping',
+  'helped', 'honestly', 'hope', 'hour', 'house', 'however',
+  // I
+  'ideal', 'image', 'important', 'inner', 'inside', 'instead', 'interest',
+  'issue', 'itself',
+  // J
+  'joined', 'judge',
+  // K
+  'keep', 'kept', 'kind', 'knew', 'know', 'known', 'knows',
+  // L
+  'large', 'later', 'latest', 'laugh', 'lead', 'learn', 'least', 'leave',
+  'left', 'less', 'level', 'levels', 'letter', 'light', 'line', 'listen',
+  'listened', 'literally', 'little', 'local', 'long', 'longer', 'look',
+  'looked', 'looking', 'lower',
+  // M
+  'making', 'many', 'matter', 'matters', 'maybe', 'mean', 'meaning', 'media',
+  'meeting', 'meetings', 'member', 'message', 'messages', 'might', 'mind',
+  'minor', 'model', 'money', 'month', 'months', 'moral', 'morning', 'mother',
+  'mouse', 'moved', 'moving', 'much', 'music', 'must', 'myself',
+  // N
+  'name', 'names', 'nature', 'near', 'need', 'needed', 'never', 'news', 'next',
+  'night', 'nobody', 'north', 'note', 'noted', 'notes', 'nothing', 'notice',
+  'number', 'numbers',
+  // O
+  'obvious', 'obviously', 'offer', 'office', 'offices', 'often', 'okay',
+  'once', 'only', 'onto', 'open', 'order', 'orders', 'other', 'outer',
+  'outside', 'over',
+  // P
+  'paper', 'party', 'people', 'perhaps', 'person', 'phone', 'phones', 'photo',
+  'piece', 'place', 'plan', 'plant', 'plate', 'point', 'points', 'poor',
+  'possible', 'power', 'press', 'pretty', 'price', 'probably', 'problem',
+  'problems', 'product', 'products', 'program', 'project', 'projects', 'proud',
+  'public',
+  // Q
+  'queen', 'question', 'questions', 'quick', 'quiet', 'quite',
+  // R
+  'rather', 'reach', 'reached', 'ready', 'really', 'reason', 'reasons',
+  'recent', 'recently', 'report', 'reports', 'respect', 'result', 'results',
+  'retail', 'return', 'right', 'rights', 'rough', 'round', 'rural',
+  // S
+  'sales', 'same', 'school', 'seeing', 'seemed', 'seems', 'sense', 'sent',
+  'seven', 'several', 'shall', 'share', 'sharp', 'shelley', 'short', 'should',
+  'showing', 'shown', 'sides', 'since', 'small', 'smile', 'social', 'some',
+  'someone', 'something', 'sometimes', 'soon', 'sorry', 'sound', 'south',
+  'space', 'speak', 'spoke', 'sports', 'staff', 'stage', 'stand', 'start',
+  'started', 'state', 'still', 'stood', 'stop', 'stopping', 'story', 'stuff',
+  'style', 'subject', 'success', 'such', 'suggest', 'support', 'sure',
+  'system', 'systems',
+  // T
+  'table', 'taking', 'talking', 'taught', 'teach', 'team', 'teams', 'telling',
+  'tells', 'terms', 'thanks', 'thank', 'their', 'them', 'then', 'there',
+  'these', 'thing', 'things', 'think', 'third', 'this', 'those', 'though',
+  'thought', 'three', 'through', 'throughout', 'today', 'together', 'told',
+  'tomorrow', 'tonight', 'took', 'total', 'touch', 'toward', 'towards', 'town',
+  'trade', 'train', 'tried', 'trouble', 'truly', 'trust', 'truth', 'trying',
+  'turn', 'turned', 'turning', 'twelve', 'type',
+  // U
+  'under', 'underneath', 'understand', 'unique', 'until', 'upon', 'usually',
+  // V
+  'value', 'various', 'very', 'visit', 'voice',
+  // W
+  'wait', 'wanna', 'want', 'wanted', 'watch', 'water', 'ways', 'week', 'weeks',
+  'weekend', 'well', 'went', 'were', 'what', 'when', 'where', 'whether',
+  'which', 'while', 'white', 'whole', 'whose', 'wife', 'will', 'wind', 'wish',
+  'with', 'within', 'without', 'women', 'word', 'words', 'work', 'worked',
+  'working', 'world', 'worry', 'worse', 'worst', 'would', 'wrong',
+  // Y
+  'yeah', 'year', 'years', 'yesterday', 'young', 'your', 'yours', 'yourself',
+
+  // ── Common first names ──
+  // Whisper frequently capitalizes these at any position. Leave them alone —
+  // LLM correction pass handles context-aware disambiguation vs enrolled speakers.
+  'aaron', 'albert', 'alexander', 'alice', 'amanda', 'amber', 'amy', 'andrea',
+  'angela', 'ann', 'anne', 'anthony', 'arthur', 'ashley', 'barbara', 'becky',
+  'benjamin', 'betty', 'brandon', 'brenda', 'brian', 'bruce', 'carl', 'carlos',
+  'carol', 'carolyn', 'catherine', 'charles', 'cheryl', 'christina', 'christine',
+  'christopher', 'cynthia', 'daniel', 'david', 'dawn', 'deborah', 'denise',
+  'dennis', 'diana', 'diane', 'donna', 'douglas', 'dylan', 'edward', 'elizabeth',
+  'emily', 'emma', 'eric', 'eugene', 'evelyn', 'frank', 'gary', 'george',
+  'gloria', 'grace', 'gregory', 'hannah', 'harold', 'heather', 'helen', 'henry',
+  'howard', 'jack', 'jackson', 'jacob', 'jacqueline', 'james', 'jane', 'janet',
+  'janice', 'jason', 'jeffrey', 'jennifer', 'jerry', 'jessica', 'joan', 'john',
+  'johnny', 'jonathan', 'jordan', 'jose', 'joseph', 'joshua', 'joyce', 'judith',
+  'julia', 'julie', 'justin', 'karen', 'katherine', 'kathleen', 'kathryn',
+  'keith', 'kelly', 'kenneth', 'kevin', 'kimberly', 'larry', 'laura', 'lauren',
+  'lawrence', 'linda', 'lisa', 'lori', 'louis', 'margaret', 'maria', 'marie',
+  'mark', 'martha', 'martin', 'mary', 'matthew', 'megan', 'melissa', 'michael',
+  'michelle', 'monica', 'nancy', 'natalie', 'nicholas', 'nicole', 'olivia',
+  'pamela', 'patricia', 'patrick', 'paul', 'peter', 'philip', 'rachel', 'ralph',
+  'randy', 'raymond', 'rebecca', 'richard', 'robert', 'roger', 'ronald', 'rose',
+  'roy', 'ruby', 'russell', 'ruth', 'ryan', 'samantha', 'samuel', 'sandra',
+  'sara', 'sarah', 'scott', 'sean', 'shannon', 'sharon', 'shirley', 'stephanie',
+  'stephen', 'steven', 'susan', 'teresa', 'theresa', 'thomas', 'timothy',
+  'tina', 'tony', 'tyler', 'victoria', 'virginia', 'walter', 'wayne', 'william',
+  'willie', 'zachary',
+])
+
+/**
+ * Fuzzy-match transcript words against a canonical list of proper nouns.
+ *
+ * RULES (all must pass):
+ *   - Candidate word length ≥ 5 chars (any case — supports lowercase dictation
+ *     AND mixed-case names like GitHub/OpenAI)
+ *   - Word must NOT be in COMMON_ENGLISH_WORDS stop-list (primary defense)
+ *   - Target must start with uppercase (canonical proper noun)
+ *   - Target token length ≥ 4 chars (allows short names like Anna, Ryan, Luke as targets)
+ *   - Length difference ≤ 2 chars
+ *   - Distance within length-scaled threshold:
+ *       5-8 chars  → 1
+ *       9-11 chars → 2
+ *       12+ chars  → 3
+ *
+ * Preserves the original word's first-letter case: capitalizes the replacement
+ * when the source was capitalized, otherwise lowercases it. This avoids
+ * creating sentence-middle capitalizations where none existed.
+ */
+export function applyFuzzyCorrections(
+  text: string,
+  targets: string[],
+): { text: string; replacements: number } {
+  if (targets.length === 0 || !text) return { text, replacements: 0 }
+
+  // Keep only targets that look like proper nouns.
+  // Drop speaker tokens and entries that don't start with uppercase.
+  const owner = getOwnerSpeakerLabel()
+  const canonical = Array.from(new Set(
+    targets
+      .filter(t => {
+        if (!t || t.length < 4) return false
+        if (t === 'Ext' || t === 'Unknown' || t === owner) return false
+        return /^[A-Z]/.test(t.trim())
+      })
+      .map(t => t.trim()),
+  ))
+  if (canonical.length === 0) return { text, replacements: 0 }
+
+  const targetLowerSet = new Set(canonical.map(t => t.toLowerCase()))
+  let replacements = 0
+  const fuzzyLog: Array<{ from: string; to: string }> = []
+
+  // Match any 5+ char word (case-insensitive). Stop-list + strict distance
+  // thresholds are the primary defenses against false positives.
+  const result = text.replace(/\b[A-Za-z][A-Za-z'-]{4,}\b/g, (word) => {
+    const lower = word.toLowerCase()
+
+    // Skip common English words (primary false-positive defense)
+    if (COMMON_ENGLISH_WORDS.has(lower)) return word
+
+    // Skip exact matches (already correct)
+    if (targetLowerSet.has(lower)) return word
+
+    // Length-scaled distance threshold
+    const len = word.length
+    const maxDist = len <= 8 ? 1 : len <= 11 ? 2 : 3
+
+    let bestTarget: string | null = null
+    let bestDist = maxDist + 1
+
+    for (const target of canonical) {
+      // Split multi-word targets — match on individual tokens
+      const tokens = target.split(/\s+/)
+      for (const token of tokens) {
+        if (token.length < 4) continue
+        // Length difference must be small to prevent collapses
+        if (Math.abs(token.length - len) > 2) continue
+        const dist = levenshtein(lower, token.toLowerCase())
+        if (dist > 0 && dist < bestDist && dist <= maxDist) {
+          bestDist = dist
+          bestTarget = token
+        }
+      }
+    }
+
+    if (bestTarget) {
+      replacements++
+      fuzzyLog.push({ from: word, to: bestTarget })
+      // Preserve original capitalization pattern
+      const wasCapitalized = /^[A-Z]/.test(word)
+      return wasCapitalized ? bestTarget : bestTarget.toLowerCase()
+    }
+    return word
+  })
+
+  if (fuzzyLog.length > 0) {
+    const preview = fuzzyLog.slice(0, 5).map(x => `${x.from}→${x.to}`).join(', ')
+    console.log(`[fuzzy-correct] Applied ${replacements} correction(s): ${preview}${fuzzyLog.length > 5 ? '...' : ''}`)
+  }
+
+  return { text: result, replacements }
+}