An AI interview assistant uses two kinds of AI model, not one. A speech-to-text model turns spoken words into text, and a large language model (LLM), such as a GPT, Claude or Gemini model, reads that text and writes the suggested answer, the next practice question or the feedback on your answer.
The model names matter less than most buyers expect. How quickly speech becomes text, what context the language model receives and how the tool shapes its output decide whether the help is usable. This guide explains the stack, what each choice changes, which vendors say what they use (checked October 2026) and a 10-minute test you can run on any tool. For the wider market picture, see our October 2026 report on the category.
What AI model does an AI interview assistant use?
Every tool in the category runs the same five-layer pipeline. Only two of the layers are AI models in the usual sense. The rest is engineering, and that's where products differ most.
| Layer | What it does | Common options | What it changes for you |
|---|---|---|---|
| Audio capture | Gets the sound: your microphone in practice, your computer's audio output on a call | Operating-system audio, a browser tab, a phone microphone | What the tool can hear at all |
| Speech-to-text (model 1) | Turns speech into text, often while the person is still talking | Whisper-family models, streaming APIs such as Deepgram and AssemblyAI, a call app's own captions | Caption speed and accuracy, especially on names and jargon |
| End-of-question detection | Decides the speaker has finished | Silence thresholds, voice-activity detection, turn-detection models | Whether help starts too early, cutting the question off, or too late |
| Language model (model 2) | Reads the question plus any context and writes the answer, question or score | GPT, Claude and Gemini models; open-weight models | Quality, structure, tone and how long you wait |
| Formatting and display | Shapes the output into bullets, a STAR outline, code or a score | The vendor's own instructions and interface | Whether you can use it at a glance |
In CoPilot Interview, for example, audio is streamed in short encrypted segments to a speech-to-text model and the question text goes to the AI model you chose. Nothing is kept on our servers beyond the request, and session files stay in your Documents folder. You see live captions of the question, then a structured answer in about 2 seconds after the question ends. Our explainer on how real-time transcription works goes deeper on the first half of the pipeline.
Which speech-to-text models do interview tools use?
Most vendors don't name their speech-to-text provider (see the table further down), but the options fall into four groups:
- Whisper-family models. OpenAI's Whisper is a general-purpose speech recognition model whose code and weights are released under the MIT license (Whisper on GitHub). It works through audio in 30-second windows rather than as a native stream, so live tools built on it cut audio into chunks. Many faster derivatives and hosted versions exist.
- Streaming speech-to-text APIs. Services such as Deepgram and AssemblyAI are built for live audio. They return partial text while someone is talking and signal when a turn has ended: AssemblyAI's streaming API marks a finished turn with an
end_of_turnflag (AssemblyAI streaming docs), and Deepgram calls the same idea endpointing (Deepgram endpointing docs). - Call-app captions. Zoom, Microsoft Teams and Google Meet can caption calls themselves. A tool that reads those captions skips a step but inherits that app's caption quality and language list.
- On-device models. Running speech-to-text on your own computer removes the network round trip, but your hardware limits the model size.
Speech-to-text is the first place an answer goes wrong. A misheard product name, acronym or person's name produces a fluent answer to the wrong question, and no language model can repair text it never received correctly. That's why the caption matters as much as the answer.
Which language models power AI interview copilots?
The second model writes what you read. Interview tools use the same language models as everyone else, reached through each provider's API:
- Hosted models from the large providers: GPT models from OpenAI, Claude models from Anthropic and Gemini models from Google. Each provider sells several tiers, from small, fast models to larger, slower ones.
- Open-weight models: models whose weights are published, which a vendor can host itself or you can run on your own machine.
Within one provider, the tier you pick changes more than the brand does:
| Choice | Speed | Accuracy | Tone and length |
|---|---|---|---|
| Small, fast model | First words arrive soonest | Good on definitions, common behavioral structure and standard coding patterns; more likely to slip on niche facts | Shorter and more generic |
| Larger model | Slower to start | Stronger on multi-step reasoning: design trade-offs, harder coding, long context | More thorough, often longer than you can say |
| Open-weight model you run yourself | Depends on your hardware | Depends on the model you pick | Fully under your control, as are setup and updates |
| Two models side by side | You read two answers instead of one | Disagreement shows you where to check | Two framings to choose from |
Tone comes mostly from the instructions the vendor wraps around the model (often called the system prompt) and from the context you supply, not from the model's brand. Two tools built on the same model can sound completely different.
Which vendors say which models they use?
We checked each vendor's own homepage in October 2026. Several name their language models; none of the competitors below names its speech-to-text provider.
Disclosure: we make CoPilot Interview.
| Tool | Language models named on its site | Can you choose? | Speech-to-text provider named? |
|---|---|---|---|
| CoPilot Interview | Best choice: 9 AI models: fast free models on Free and Standard, ChatGPT and Gemini models on Pro, plus Claude on Pro+ | Yes; Dual AI Mode shows two models' answers side by side on paid plans | No; audio goes in short encrypted segments to a speech-to-text model |
| Parakeet AI | Models from OpenAI, Anthropic and Google | Yes, from a model dropdown when you create or edit a session | No |
| Interviews Chat | Models from OpenAI, Anthropic, Google and other providers | Yes; it says you can switch between models | No |
| LinkJob AI | GPT, Claude and Gemini | Yes, depending on plan; the free trial has basic models | No |
| Natively AI | Your own API key (OpenAI, Anthropic, OpenRouter or a custom endpoint), local models through Ollama, or Natively's hosted plans | Yes, you supply the model | No |
| Verve AI | Names models from OpenAI, Google, Anthropic and Meta | Not stated | No |
| Final Round AI | Not stated on its homepage | Not stated | No |
| LockedIn AI | Not stated on its homepage | Not stated | No |
| Sensei AI | Not stated; says you can tune the tone and style of responses | Not stated | No |
Credit where it's due: Parakeet AI and Interviews Chat are the most open about model choice among the paid tools, and Natively AI gives the most control if you're comfortable managing your own keys or local models. Our own limits: CoPilot Interview's full app is desktop only (Windows and macOS), the Android companion is a beta, there's no human coach, and the free plan is 15 minutes a day on the free models. Vendors change line-ups often, so confirm on each site before you buy. For head-to-heads, see CoPilot Interview vs Parakeet AI and CoPilot Interview vs LinkJob AI; for reading two answers at once, see how Dual AI Mode works.
Why does context matter more than the model name?
A language model can only use what it receives. In an interview tool that's three things: the question text, the vendor's instructions and whatever context you supplied. Give the strongest model a misheard question and no context, and you get a polished generic answer.
An example we wrote. The question: "Tell me about a project you're proud of."
- No context, any model: a template. Pick a project, describe the challenge, your role and the result. Correct and empty.
- With one résumé line and the job description: the résumé says you "moved nightly reporting to an incremental pipeline, cutting runtime from 3 hours to 25 minutes", and the posting stresses data reliability. The outline now leads with that project and ties the result to reliability.
The second answer comes from better input, not a better model. CoPilot Interview's Mock Interview builds its questions from your résumé and the job description for exactly this reason. Our guide to how these tools use your résumé and job description shows what to paste and what to leave out, and our guide to making the answers sound like you covers length, format and tone instructions.
How can you compare models in 10 minutes?
A feature page can't tell you which model suits you; a short test can. Use the tool's practice mode, or ask a friend to read you questions on a test call with the headset you'll actually use.
- Pick five questions (2 minutes). One definition, one behavioral question, one coding explanation, one design trade-off and one that contains a product name or acronym from your field.
- Run them on the default model (4 minutes). For each, note whether the caption is right, how long the first useful line takes and whether you'd say the answer out loud.
- Switch models and rerun the two hardest (3 minutes). Usually the design trade-off and the jargon question.
- Score it (1 minute). Fill in the grid below and keep the model that wins on the questions you're most likely to get.
| Check | Model A | Model B |
|---|---|---|
| Caption correct on the jargon question | ||
| Seconds to the first useful line | ||
| Used your résumé or story where relevant | ||
| Factual or complexity errors | ||
| Short enough to say in under 2 minutes | ||
| Something you'd say in your own words |
Treat the caption row as a test of the speech-to-text model and the rest as a test of the language model. If captions fail, switching language models won't help; fix your audio first. If the delay is the problem, our breakdown of where the seconds go explains each stage. Follow the rules of the employer or platform you're interviewing with — many people use it only to practice, which works on the free plan.
Practice it: learn a model's habits where mistakes are free
The safest place to learn how a model answers is a practice round. CoPilot Interview's Mock Interview builds the round from your résumé and the job description, reads each question aloud, scores your spoken answer out of 10 with what to fix and a model answer, and ends with a report card. The browser AI mock interview needs nothing installed. Both work on the free plan of the desktop AI interview assistant we make: 15 minutes every day, no account and no credit card.
If you're curious how the pieces fit together in code, our developer walkthrough for a mock-interview tool builds one from public APIs, and ChatGPT vs Claude vs Gemini for interview prep compares the three chatbots directly.
Compare models on your own questions
Download CoPilot Interview free for Windows or macOS: 15 minutes a day, all four modes and the free AI models, with no account or credit card. Pro adds ChatGPT and Gemini models; Pro+ adds Claude.
Download CoPilot Interview freeFAQ
What AI model does an AI interview assistant use?
Two kinds. A speech-to-text model turns the spoken question into text, and a large language model, such as a GPT, Claude or Gemini model, writes the answer or the feedback. Several tools let you choose the language model; few name their speech-to-text provider.
Does a bigger AI model make a better interview tool?
Not by itself. Larger models help with multi-step questions such as design trade-offs, but they start slower and write longer answers. A misheard question or missing context hurts more than a smaller model does, so check the captions and supply your résumé and the job description first.
Do interview tools use ChatGPT?
Several use OpenAI's GPT models through its API, and some name them on their sites. That isn't the same as the ChatGPT app: an interview tool adds transcription, its own instructions and an interface built for interviews. CoPilot Interview offers ChatGPT and Gemini models on Pro and adds Claude on Pro+.
Which speech-to-text model is the most accurate for interviews?
There's no single answer. Accuracy depends on your audio, accents, background noise and vocabulary. Test with questions and terms from your own field, and read the caption before any suggested answer.
Which AI models does CoPilot Interview use?
CoPilot Interview has 9 AI models: fast free models on the Free and Standard plans, ChatGPT and Gemini models on Pro, plus Claude on Pro+. Dual AI Mode shows two models' answers side by side on paid plans.
Is CoPilot Interview made by OpenAI, Anthropic or Google?
No. CoPilot Interview is made by NeuraSense AI LLC and is not affiliated with OpenAI, Anthropic, Google, Microsoft Copilot or GitHub Copilot. Its Pro and Pro+ plans give you access to ChatGPT, Gemini and Claude models.