The Short Answer
Ghotly AI 4.3.3 is a free update for the Windows app. The biggest change: the app no longer needs a Deepgram key. It hears the interviewer with the AI keys you already have, so one free Groq or Gemini key is enough to listen and to answer.
To get it, open the app, let the update download, then close the app so it installs. Or download the installer from the Download page.
How the companies compare when they listen, with our own measurements: free speech-to-text for live interviews.
See all of 4.3.3 in the new 9-minute video guide: the install, one free key, then a real video call with live captions, an answer in seconds, and Ghotly AI kept out of the screen share.
Listening Without Deepgram
- Your keys listen: Groq (Whisper, free and the fastest), Gemini (free), xAI, OpenAI and OpenRouter (a free audio model, or Whisper with credit).
- Automatic uses your keys in this order: Groq, Gemini, xAI, OpenAI, OpenRouter. If one refuses (a free limit, a wrong key, a network drop), the next takes over by itself and the listening strip says so. The words stay in the order they were said.
- Settings → Listening shows who listens, for example "Listening with Groq · Whisper Large v3 Turbo (free) · backup: Gemini". You can choose a company, a model and the interview language: English, Hindi, Marathi, Tamil, Telugu, Kannada, Bengali, Gujarati, or Detect for a Hindi-English mix.
- Live captions: "Interviewer speaking…" shows the moment a voice starts, and the words appear while they speak, about a second behind (when Groq, OpenAI or xAI listens). On Groq the captions use a second Whisper model, so they never use up the limits the questions need. Settings → Listening → Live captions turns them off.
- Only speech is sent: each sentence goes as soon as the interviewer pauses. Silence, a fan or a cough is not sent, which also saves your free limits. Lines that speech-to-text sometimes invents in silence ("Thank you for watching") are dropped.
- Your own microphone (Settings → More) is heard the same way.
- Claude and NVIDIA keys still answer, but cannot hear: add a free Groq or Gemini key with them. The home screen says so before you start.
- The audio test on the home screen has a new Test listening step: it shows which company answered and how fast.
- An old Deepgram key is not used any more, and nothing sends it anywhere.
Knowing When to Answer
- Small talk and thank-yous are left alone: "okay", "great, thanks", "excellent, thank you", "good answer", "hello?", "are you there?", "give me one second", "let me share my screen" and "see you soon" are not answered, nor "okay" and "hmm" however speech-to-text spells them ("O-key", "okkk", "Hmmmm"), the Hindi and Marathi ones ("achha", "theek hai", "haan ji", "barobar"), your own name ("Okay Riya"), "go ahead" or "take your time", and "sorry?" or "can you repeat that?". We checked over 200 such lines; "Okay, thanks. Now tell me about Docker." still gets its answer.
- A quick check for unclear lines: a line that may or may not need an answer (a statement such as "We use Kafka here", or a bare topic) gets a yes-or-no check with your Groq key (the fastest; about two thirds of a second, at most 1.5) before the app decides. Without a Groq key the app's own rules decide, as before. Settings → Auto answer → Smart answering turns it off.
- It learns you: when you send a line by hand that the app skipped, lines like it are answered by themselves, in later interviews too. What it learned stays on your PC; Settings → Auto answer shows it and can forget it. It never learns to answer "okay" or "thanks".
- Long questions: when the interviewer talks for a while, the app answers the last question in what they said (the rest goes along as context), and the question box shows at most 3 lines, so the answer is always in view.
Faster, Better Answers About You
- Ready answers: with resume details in your profile, the app prepares the answers to the common questions about you when listening starts (one request with your own key): tell me about yourself, your main project, strengths and weaknesses, why this role, why you are leaving, and the biggest challenge you faced. When one of them is asked, its answer is on screen at once. Answer again asks the AI afresh, and specific or technical questions always go to the AI.
- Answer length: a new setting (Short, Medium, Detailed) under Settings → Auto answer.
- Your Custom Instructions win: a rule such as "keep every answer to 6 lines" or "max 80 words" now replaces the app's own length, and the answer is shaped to fit it.
- Code only when it is asked for: "What is a closure?" gets an explanation in words; "Write a function to..." or "Write a SQL query..." gets the code.
- Quicker replies: the part of each request that never changes comes first, so providers that reuse repeated text (Gemini, OpenAI) can answer sooner; a question that ends with a question mark is sent as soon as its words are in; and OpenAI's GPT-6 Luna answers without thinking first.
- xAI (Grok): the new first model, Grok 4.20 Fast, starts an answer in about a second; the thinking models (4.3 to 4.7) are asked to think briefly, so they start in seconds instead of 20 or more.
- Gemini at busy hours: when Gemini's bigger models give no words for 5 seconds, the same key asks Gemini Flash-Lite (it was 15 seconds), and keeps doing so for 10 minutes instead of waiting each time. Gemini also listens sentence by sentence now.
Fixes
- NVIDIA: "No content in NVIDIA response" is fixed. Nemotron models are asked to answer without long thinking, and an empty answer tries the next NVIDIA model before it says so.
- NVIDIA: an "Authorization failed" reply now says that the key is not valid.
Common Questions
Do I still need a Deepgram key?
No. From 4.3.3 your Groq, Gemini, xAI, OpenAI or OpenRouter key listens to the interviewer. Versions before 4.3.3 still need Deepgram, so update the app.
Which key should I add for listening?
Groq: it is free and the fastest. Add a free Gemini key as well, as a backup for when Groq's free limit is reached. Our free API keys guide shows each step.
I only have a Claude or NVIDIA key. Will it work?
Those keys answer but cannot hear the call. Add a free Groq or Gemini key and the app listens with it; the answers can still come from Claude or NVIDIA.
Does the interview audio go to Ghotly AI?
No. The audio goes from your PC to the company that listens (the key shown in Settings → Listening), the same way your questions go to your AI provider. See the privacy policy.
