First off, this is an awesome API that I would love to use in my app. The big problem I have, though, is that it's not really usable for real-time, low latency identification from short samples because:
1. The asynchronous callback method requires me to make constant polls to the operation result endpoint, which takes (from my measurement) about 1200ms in the ideal case, whereas I would really prefer results within 400-500 ms.

1. Each poll on the operation status costs me QPS, which triggers throttling if I poll to often

I would propose the following change to the speaker identification API:

• For long inputs (audio over 10+ seconds long), return HTTP 202 with an operation result URL, as it does today




• Otherwise, process the audio immediately and return HTTP 200 with the identification results. This should both reduce latency and the polling required to retrieve the results.