Parakeet Words

1// Merge Parakeet-MLX token timestamps into word timestamps.
2//
3// parakeet-mlx JSON emits SUB-WORD tokens (" H", "ello", ...) with per-token
4// start/end. Captions + transcript-cut need WORD timestamps, so join tokens
5// into words on the space boundary: a token whose text starts with a space
6// (or the very first token) begins a new word; the rest append. Output matches
7// the { words: [{ text, start, end }] } shape the rest of media-use consumes
8// (see words.mjs / cutlist.mjs).
9
10export function mergeTokensToWords(parakeet) {
11  const sentences = Array.isArray(parakeet?.sentences) ? parakeet.sentences : [];
12  const words = [];
13  for (const s of sentences) {
14    for (const t of s.tokens ?? []) {
15      const raw = typeof t.text === "string" ? t.text : "";
16      const startsWord = raw.startsWith(" ") || words.length === 0;
17      if (startsWord) {
18        words.push({ text: raw.trim(), start: t.start, end: t.end });
19      } else {
20        const w = words[words.length - 1];
21        w.text += raw;
22        w.end = t.end;
23      }
24    }
25  }
26  return { text: (parakeet?.text ?? "").trim(), words: words.filter((w) => w.text.length > 0) };
27}