Word timings

Each KittenTTSResult can include model-predicted timings:

type KittenWordTiming = {
  wordIndex: number;
  word: string;
  startTime: number;
  endTime: number;
};

const result = await tts.generate('Highlight these words as they play.');

await tts.play(result, {
  onPlaybackStart: () => {
    const startedAt = Date.now();
    const timer = setInterval(() => {
      const seconds = (Date.now() - startedAt) / 1000;
      const active = result.wordTimings.find(
        word => seconds >= word.startTime && seconds < word.endTime,
      );
      setActiveWordIndex(active?.wordIndex ?? null);
    }, 50);
  },
});

Accuracy and scope

Timings come from predicted phoneme durations. They are useful for read-aloud highlighting but are not forced alignment.

• Generate one sentence or short paragraph for the most stable mapping.
• wordIndex is local to the generated result.
• Timings may be empty if the model did not return duration data.
• Timings may be unavailable when input is split across multiple inference chunks.
• Start clocks from actual playback, not from generation completion.

For long content, use generateStreaming() and update highlighting separately for each yielded result.