This has been greatly improved using 2 updates:

1) User responses are live transcribed removing the time for transcription upon completing the full response

2) interviewer text-to-speech models have been improved - so the interviewer audio is much quicker after their response has been generated.