Kai Hirota

2 min read402 words

Callback

Job interview practice with AI: listen to a simulated interview, or practise with an AI interviewer and get feedback on every answer.

Callback started as an experiment with voice AI, comparing speech-to-text and text-to-speech against voice-to-voice models. It became an app for practising job interviews: you upload your resume, paste the link to the job, and rehearse the interview.

callback.from81.app

Callback on iPhone: the chat that sets up your record; a job with its interview rounds; and a round with its questions

Callback on iPhone: the transcript of a rehearsal; the report with its scorecard and missing facts; and the resume built from your record

2 ways to rehearse#

The 2 ways match the 2 approaches I was comparing.

Mock interview, voice to voice. You answer out loud, and an AI interviewer follows up on what you say in real time. It runs on OpenAI's gpt-live-1, a voice-to-voice model. The audio goes from your phone straight to OpenAI and never passes through Callback's server.

Simulated interview, text then speech. For when you can't talk: a text model writes the interview turn by turn, with Callback answering as you from your resume, and Cartesia's text-to-speech speaks it. Each turn starts playing as soon as its first audio is ready, and you can steer the interview while you listen.

Feedback on every answer#

Each rehearsal ends in a report: whether you would pass and why, a rating for every answer, and a stronger way to say it. The report separates 2 kinds of gap: answers a better telling would fix, and facts missing from your record, which no telling fixes. For the second kind, it asks you the question that fills the gap. Callback works in 8 languages.

Choosing the voice model#

Before choosing the model for mock interviews, I compared gpt-live-1 with xAI's grok-voice-3.0. A scripted candidate, an AI that answers from a fixed record and speaks through text-to-speech, sat about 60 full interviews with gpt-live-1 and 21 with grok-voice-3.0. During each call the server sent the interviewer steering notes, such as moving to another topic or wrapping up.

grok-voice-3.0gpt-live-1
Time from the candidate stopping to the interviewer's reply (median)1.04 s1.96 s
Pauses of 3 to 8 s mid-sentence where the interviewer took the turnabout 8 of 103 of 14
Steering notes followed2 to 3 of 54 of 4
Record of the call kept on the servernoyes

Grok replies faster, and it opens and ends a call on its own. GPT-Live lets a candidate pause to think, follows the steering notes, and gives the server its own record of the call. An interviewer should let the candidate think, so mock interviews run on GPT-Live.