How much latency does an AI interview assistant add?

An AI interview assistant is a pipeline, and each stage adds a small amount of delay:

• Audio capture — the app continuously buffers microphone and system audio. This is usually negligible (tens of milliseconds).
• Speech-to-text (STT) — streaming transcription returns partial results as the interviewer is still speaking, so you see text appear with a short lag rather than waiting for the full sentence.
• Language model inference — once the question is recognized, the model has to generate an answer. This is normally the largest single component of latency and scales with how long the answer is and how much context (resume, job description, prior turns) is included.
• Network round trips — calls to cloud STT and LLM providers depend on your connection quality and physical distance to the provider's region.

So the honest answer to "how much latency" is: it's the sum of those stages, not a single number.

As a rough mental model for any modern AI interview assistant, including SubcueAI:

• First transcript words appear within roughly a second of the interviewer speaking, because streaming STT emits partial results.
• First tokens of an answer usually start arriving a second or two after the question finishes — this is the figure that matters most, because you can start reading immediately.
• Full answer takes longer to finish streaming, but you don't have to wait for it to finish before you start speaking.

These ranges assume a stable broadband connection. On a weak Wi-Fi connection, a congested coffee-shop network, or while sharing your screen and running heavy apps, every stage gets slower.

SubcueAI is a native desktop app for macOS and Windows with dual audio capture (your mic plus the meeting's system audio) and a local floating overlay. A few design choices help keep perceived latency low:

• Capturing system audio directly avoids re-recording speakers through your microphone, which keeps transcription cleaner and reduces the need for retries.
• Streaming transcription and streaming answers mean you see useful content before the full response is finished.
• The overlay renders locally on your machine, so updating the UI doesn't depend on a browser or a meeting bot joining the call.

You can read more about the architecture on the overview page or the tutorial.

Most of the latency you'll notice in practice comes from your own setup, not the assistant. Practical things that help:

• Use a wired connection or a strong 5 GHz Wi-Fi signal rather than a marginal one.
• Quit heavy background apps (large IDEs indexing, video editors, big browser sessions) before the interview.
• Close other tabs and apps that are streaming audio or video.
• Do a dry run beforehand so you know how the timing actually feels — see the tutorial.

It's also worth being realistic: an AI assistant is not instant. Treat it as a hint layer you glance at, not a teleprompter you read word-for-word.