This is great to see. Evals on voice are hard - we only have evals on text based prompting, but it doesn't fully capture everything. Excited to give this a try.