The Short Answer
For a live job interview, the best free speech-to-text API in our November 2026 test was Groq's Whisper Large v3 Turbo. The interviewer's question was written out 0.74 seconds after they stopped speaking (median). With live captions on, the first words were on screen 1.2 seconds after they started. It handled all 25 answer-or-not decisions right. Its free tier allows 20 requests a minute and 2,000 a day for each Whisper model.
Gemini (also free) is a good backup. It was slower (about 2.5 seconds per question) and needed guards at Google's busy hours. OpenAI and xAI (both paid) were fast and accurate too.
Since Ghotly AI 4.3.3, the app listens with these keys, so you need no separate speech-to-text account. The keys you add for answers also hear the interviewer.
How We Tested
- One interview, 143 seconds, 25 lines. It had technical and HR questions, coding tasks, "okay", "hmm" and "thanks", small talk, a question cut in two by a long pause, and background noise in four places. The voices were text-to-speech (Kokoro) with US, UK and Indian English accents.
- Played in real time through Ghotly AI's own listening code: voice detection on the PC, one request per sentence, and the same filters the app uses. We used real API keys on 21 and 22 November 2026, in the Indian evening.
- Measured:
- the words wrong (word error rate);
- how soon after the interviewer stopped the whole question was written;
- with live captions, how soon after they started the first words showed;
- whether the app would answer each line or rightly leave it ("okay", "thanks").
Synthetic voices are cleaner than a real call, so read the numbers as a comparison between providers, not a promise for every call.
The Results
| Speech-to-text | Free? | Question ready after the interviewer stops (median) | First words on screen after they start | Words wrong | Answer-or-not right |
|---|---|---|---|---|---|
| Groq Whisper Large v3 Turbo | Yes | 0.74 s | 1.2 s | 15.7% | 25/25 |
| xAI Grok Voice Transcribe | Paid | 0.87 s | 1.2 s | 11.9% | 25/25 |
| OpenAI GPT Transcribe | Paid | 1.16 s | 1.6 s | 10.5% | 25/25 |
| Gemini 3.1 Flash-Lite | Yes | about 2.5 s | 5.6 s (no live captions) | 10.7–15.4% | 24–25/25 |
Most of the wrong words came from two hard lines: the synthetic voice saying "bit.ly", and a "hmm" written out as a word. Every real question came through. The answer-or-not decision is what matters in an interview: the app must answer the question and stay quiet on "okay, thanks".
What Went Wrong, and How Ghotly AI Handles It
- Gemini answered instead of writing. Once it replied "I'm not sure what you mean" to a question it should have transcribed. Ghotly AI now tells it, as a system instruction, that it is a speech-to-text engine and never an assistant, and drops chat-style replies.
- Gemini at busy hours. We saw "high demand" (503) errors, and requests that hung for 12 seconds and then answered at once when asked again. The app now:
- asks Gemini's other free model with the same key when one is busy;
- waits about 5 seconds for a sentence before asking once more.
- The word list read back. We give the model the role's technologies so it spells them right. Once, on a noise, Gemini wrote that list itself. Text that is only the hint words is now dropped.
- Whisper writes words on noise. We saw "Thank you." on silence and "Rekord" on a fan. Ghotly AI drops Whisper's known silence lines, and drops a word or two that Whisper itself scores as unsure (average log-probability below −1.0).
- Free limits. Live captions send the sentence so far every 1.5 seconds. On the same Groq model, they used up the 20-a-minute limit the questions need. Groq counts limits per model: in our check, the 21st request to Turbo in a minute was refused while Large v3 still answered. So Ghotly AI asks the second Whisper model for captions.
The Best Free Setup
- A free Groq key (console.groq.com). It listens fastest, and it also writes answers.
- A free Gemini key (aistudio.google.com) as the backup for when Groq's free limit for the day is used up.
- In Ghotly AI, open Settings → API Keys and paste both. Settings → Listening then shows "Listening with Groq · Whisper Large v3 Turbo (free) · backup: Gemini". The home screen's audio test has a Test listening step.
Step by step, with pictures: the free API keys guide. Everything that changed: the 4.3.3 release notes.
Common Questions
What is the best free speech-to-text API for live interviews?
In our November 2026 test it was Groq's Whisper Large v3 Turbo: free, with the question written out 0.74 s after the interviewer stopped, and 25 of 25 lines handled right.
Is Groq Whisper free?
Yes, with limits. A free Groq key allows 20 requests a minute and 2,000 a day for each Whisper model, as Groq's own response headers showed in November 2026. Paid tiers allow more.
Can Gemini do speech-to-text?
Yes. Gemini models take audio and write it down, and Gemini 3.1 Flash-Lite does it on a free key. In our test it took about 2.5 s per question. At busy hours it needed retries, and once it answered instead of writing.
How fast can live captions be without a streaming service?
About a second behind the speaker. Ghotly AI sends the sentence so far every 1.5 seconds while the interviewer speaks. In our test the first words showed 1.2 s (Groq, xAI) to 1.6 s (OpenAI) after the interviewer started.
Do I still need a Deepgram key for Ghotly AI?
No. Since version 4.3.3 your Groq, Gemini, xAI, OpenAI or OpenRouter key listens. Claude and NVIDIA keys answer but cannot hear, so add a free Groq or Gemini key with them.
