Does a local AI mock interview actually work?

Can a voice model really run on your own laptop?

Yes, and that's new. On 23 October 2026, Block unveiled Dessa, a full-duplex speech-to-speech model the company says runs locally on a MacBook Pro M4 Max or better, faster than real time. Dessa was trained from scratch on human conversational speech alone — no text corpus, no pretrained foundation — and because it listens and speaks at the same time, it copes with interruptions, laughter and mid-sentence changes of tone. Block hasn't disclosed a parameter count, benchmark scores or a release date; the research preview went first to a small group of AI safety researchers, with a public waitlist behind it. The lab lead called it a GPT-3 moment for speech. For anyone rehearsing interview answers, the interesting part isn't the uncanny realism — it's the word "locally".

Look at the hardware that claim rests on and it stops sounding exotic. The M4 Max starts at 36GB of unified memory and configures up to 128GB, with memory bandwidth of 546GB/s on the full die. Anything that fits inside that pool runs without a single network call. And you don't have to wait for a waitlist to try the idea: people have been stitching together local voice stacks on Apple Silicon for a while now — whisper.cpp with Metal acceleration for speech-to-text, Ollama hosting a 7B-to-14B model as the brain, and Kokoro-82M, an Apache-2.0 text-to-speech model, for the voice. Qwen3-Omni now has an MLX path too, benchmarked on an M4 Pro with 48GB at roughly one second to first audio using a 4-bit checkpoint.

So why does this matter for interview prep specifically? Because the thing most job seekers are missing isn't a list of questions — it's reps out loud, under time pressure, with someone cutting in. Reading your answer silently off a screen builds nothing. Saying it badly eight times until the fourth sentence stops rambling builds everything. Full-duplex behaviour is what makes that rehearsal feel like a real screen rather than a turn-based quiz: the model can say "mm-hm" while you're still talking, or jump in when you drift. Moshi, the earlier open full-duplex model, lands around 200ms of practical latency. That's conversational speed, and it changes how hard the practice feels.

What does a local AI mock interview do well?

Local voice practice wins on three things: privacy, unlimited repetition, and conversational pressure. Everything you say stays in your laptop's memory, so there's no vendor transcript of you rehearsing why you left your last job. You can run the same question twenty times at 11pm the night before a 9am screen without watching a per-minute meter. And because the model answers in under a second, you feel the real discomfort of a pause — the silence that makes candidates pile three unnecessary sentences onto a finished answer. Those are not small advantages. They're precisely the conditions a paid human coach can't give you at 11pm, and a cloud service gives you only with a logged transcript.

The privacy angle deserves more weight than it usually gets. Think about what actually comes out of your mouth in a mock interview: the real reason you were made redundant, the salary number you're hoping for, the gap you're nervous about, the project you signed an NDA over, the visa status you'd rather disclose on your own terms, the manager you clashed with. Plenty of people self-censor in cloud practice tools and then rehearse a sanitised version of their story — which is the one answer they'll never need. On-device inference removes that hesitation completely. Nothing leaves the machine, nothing lands in a log, nothing gets reviewed for model training. You can be blunt in rehearsal so you can be composed in the room.

There's a practical edge too: offline AI interview practice works on a plane, in a basement office with terrible wifi, or in the twenty minutes before a video call when you don't trust the hotel connection. It also doesn't rate-limit you mid-session, which matters because the most useful mock interview is the boring one — the same three questions, hammered until your opening line is fifteen words instead of ninety. My honest opinion: the human-sounding voice is the least valuable part of this story. A slightly robotic voice asking the exact question that will sink you beats a beautiful voice asking "tell me about yourself".

How do you feed it your CV and the job ad?

A local voice model only asks questions worth answering if you hand it two documents first: your actual CV and the job description you're targeting. Left to its own devices, it will default to the generic warm-up questions you've already over-rehearsed. With your CV in context, it can ask why your title changed in 2023, what you personally shipped versus what the team shipped, and how you'd explain eleven months out of work. With the job ad in context, it can ask about the three requirements you only half-meet. That pairing — your document plus the target role — is the whole difference between theatre and preparation.

Mechanically, this is easy in a glue-code stack and hard in a pure speech-to-speech model. If you're running whisper.cpp, Ollama and Kokoro, there's a text layer in the middle, so you paste your CV and the job ad straight into the system prompt and tell the model to interview you as a hiring manager for that specific role. A model like Dessa, trained end-to-end on audio with no text channel at all, has nowhere obvious to put a document — which is exactly why the uncanny-voice demos are a research preview and not an interview-prep product yet. For now, the architecture that sounds least impressive is the one you can actually steer.

The harder problem is knowing which questions to feed it, and that's where your CV analysis does the work. Paste the job description into a free CV check against the role and you get back five category scores out of 100, the strengths a recruiter will latch onto, and — on the paid Full Analysis — a list of missing skills, blind spots and the specific reasons you might be rejected. Read that list again as an interview question bank, because that's what it is. Every gap a screening system flags is a question a recruiter will ask out loud. Hand those to your local voice model verbatim and the rehearsal stops being comfortable.

Where does local voice practice fall short?

It falls short at scoring and at probing. Speech-to-speech models lose measurable reasoning ability compared with the text models they're derived from — researchers in the field describe this plainly as an intelligence gap, and it shows up exactly where interview coaching needs judgement. Kyutai's own model card for Moshi says the model has limited abilities for complex tasks, can't use tools, and suits casual conversation, basic facts and roleplay. That's a fair description of what you get. It will ask you a competency question and sound convinced by whatever you say. It won't notice that your STAR answer had no measurable result, or that you described a team outcome as a personal one.

The context limits bite too. Moshi cuts off conversations running past roughly five minutes, and its knowledge base covers only 2018 to 2023 — so a fifty-minute panel simulation, or any question about how hiring works this year, is simply outside its range. Expect repetition and drift in longer sessions. Local text models have their own staleness problem, which has become a live issue now that even frontier assistants are being trained away from the live web; it's worth understanding which CV advice goes stale fastest before you trust any model on current market norms. Treat the voice model as a sparring partner with no memory, not a coach with a plan.

The follow-up probing gap is the one that costs you interviews. A real hiring manager hears "I improved deployment reliability" and asks: by how much, measured how, over what period, and what did you personally change? Three layers deep is where candidates fall apart, and it's where a small speech-trained model loses the thread. Run the same transcript through a capable text model afterwards and you'll get those follow-ups, properly ordered, plus an honest read on which claims sound unverifiable. The split is clean: voice for delivery, text for interrogation.

What does the right interview prep stack look like?

Split the job three ways. Use a local voice model for delivery reps — pacing, filler words, interruption recovery, saying your pivot story out loud until it's forty seconds instead of four minutes. Use a text model for content: building the question bank from your CV and the job ad, scoring each answer against what the role demands, and generating the follow-ups a sharp interviewer would ask. Use a human for the things neither can judge: whether you came across as someone they'd want in a 4pm crisis meeting. Most candidates invert this and spend their time reading answers they've never once spoken aloud.

A workable week looks like this. Monday, run your CV against the exact job description and write down every flagged gap. Tuesday, draft tight answers in text and cut each to under ninety seconds. Wednesday through Friday, speak them — to a local voice model if you've got the hardware and patience to wire whisper.cpp and Ollama together, or to a hosted voice model if you'd rather skip the setup. Our walkthrough of running voice mock interviews built from your CV covers the hosted route, including ten prompts worth stealing. Over the weekend, record one full run and listen back. You'll hear problems no model flags.

One rule holds regardless of which voice you practise with: never bring it into the real interview. Live AI assistance during a screen is detectable, increasingly screened for, and a fast route to being blacklisted by a company you wanted. And a last piece of perspective — if you're reaching for mock interviews, make sure you're actually getting interviews first. A perfect spoken answer is worthless if your CV never clears automated screening, which is where most candidates are genuinely stuck. Fix the document, earn the conversation, then rehearse it until the words come out clean. In that order.

Frequently asked questions