Building InterviewRoom: Live Voice Interviews on a Free-Tier Budget

• Challenge: A spoken interview has to feel like a conversation, but every turn pays for speech-to-text, a language model and speech synthesis. A free product needs both responsiveness and a predictable bill.
• Solution: A streaming turn pipeline over WebSockets, speech detection in the browser, sentence-by-sentence synthesis, interruption support, and usage ceilings checked on every turn.
• Outcome: InterviewRoom is live for candidates practising and for hiring teams sending interviews by invite link. When a budget runs out, it degrades in a planned order instead of failing.

Most interview practice tools are quizzes: read a question, type an answer. Real interviews are spoken, and the hard part is the follow-up question that probes what you just said. We wanted a product where you speak to an interviewer, it responds to your actual answer, and you leave with a transcript and a score.

Voice makes every weakness obvious. Silence after you finish speaking feels like a dropped call. An interviewer that keeps talking over you feels like a recording. And because the practice tier is free, every minute of usage is paid by us, so cost had to be designed in from the first commit.

The browser runs the open-source Silero voice activity detection model and streams raw audio to the API only while the candidate is speaking. The API is Node.js with Express and the ws library, with Postgres through Drizzle for interviews, transcripts and daily usage. Speech-to-text is Whisper on Groq. Replies come from a gpt-4o-mini deployment on Azure OpenAI, and the voice is Azure neural text-to-speech.

Groq has no streaming transcription, so live captions work by re-transcribing the growing utterance at most once per interval, with one request in flight per session. Captions get a fixed share of the daily speech-to-text budget, so they can never starve the final transcription of an answer.

The model reply is streamed and cut into sentences, handling abbreviations such as "Dr." and "e.g.". Each sentence goes to speech synthesis as soon as it is complete, so the candidate hears the first sentence while later ones are still being written. Lines that repeat across sessions, like greetings and transitions, are served from an in-memory audio cache keyed by voice and text.

Interruptions use a generation counter. When the candidate starts talking, the session increments the counter, aborts in-flight model and synthesis requests, and tells the browser to stop playback. Any late response tagged with an old generation is discarded. The interview stage is a pure function of elapsed time and planned duration, which keeps pacing predictable and easy to test.

InterviewRoom is live at interviewroom.mecverse.com. Candidates practise company-style interview loops for free, and hiring teams send candidates an invite link and review the transcript and score afterwards.

The cost model held up because limits were part of the design rather than added after launch. When the daily voice budget runs out, the interview continues using the browser's built-in speech synthesis. When transcription or model capacity runs out, the session ends with a clear message rather than a broken one. We wrote up the reasoning in Put a Ceiling on Every AI Feature Before You Launch and Talking to an AI Should Not Feel Like Leaving a Voicemail.