Most mock interview tools rely on text chatbots or static questionnaires. But in a real technical or managerial interview, how you articulate ideas verbally under time constraints is what gets evaluated. We built PracticeOnline to simulate the authentic conversational dynamics of live video calls with conversational AI.
1. The Conversational Latency Problem
Human conversational turn-taking has an average latency of approximately 200–300 milliseconds. When an AI interviewer takes 3 to 5 seconds to process user speech and respond, the conversation feels robotic, breaking the candidate's immersion and flow.
To deliver an authentic interview experience, our entire pipeline—from the moment the candidate stops speaking to the first byte of synthesized interviewer voice—had to complete in under 800ms. Here is the architecture we engineered to achieve sub-500ms response times.
2. Client-Side Audio Pipeline: Web Audio API & VAD
Instead of sending massive uncompressed audio streams to backend servers continuously, we process user audio directly in the browser using the Web Audio API and a calibrated Voice Activity Detection (VAD) loop:
We attach an AudioContext and AnalyserNode to the candidate's microphone input stream.
We analyze RMS energy levels every 50ms. When speech volume drops below the calibrated noise floor for a hangover delay (typically 2.5 to 3.5 seconds), the recorder automatically finalizes the audio blob.
The client buffers audio chunks into an efficient MediaRecorder stream, minimizing upload payload sizes down to <100KB per response.
3. Speech-to-Text: Groq Whisper-Large-v3-Turbo
Traditional cloud transcription APIs often introduce 1.5–3.0 seconds of latency. We route recorded audio blobs to Groq's low-latency LPU infrastructure running whisper-large-v3-turbo.
By injecting the target Job Description (JD) and role context into the Whisper prompt metadata, our speech-to-text pipeline accurately transcribes domain-specific acronyms (e.g. Kubernetes, Kafka, Redux, PostgreSQL, ACID, CI/CD) with transcription latency consistently under 200ms.
4. Intent Classification & Adaptive Follow-Up Probes
Rather than simply advancing to the next pre-written question blindly, a real hiring manager actively listens to your answer and determines whether to probe deeper or clarify misunderstandings. Our LLM inference engine runs an intent classification heuristic on every transcript:
If the candidate asks to repeat or clarify, the AI rephrases without deducting score points.
If the answer is shallow, the AI asks a targeted follow-up to test underlying depth.
Scores the response across 4 rubrics and dynamically generates the next progressive question.
5. Real-Time Vocal Synthesis & Diagnostic Reporting
The AI synthesizes natural vocal audio using low-latency streaming TTS and plays it directly through the browser audio bus. Meanwhile, background workers calculate communication speed (WPM), filler word density (um, like, basically, actually), and STAR method structural completeness.
By eliminating scheduling barriers and high hourly coaching costs, voice AI mock interviews give every candidate access to rigorous, repeatable interview preparation on PracticeOnline.
