harlantwood / ai-interviewer

AI Interviewer

TL;DR: AI interviews you, on whatever topics you want, creating audio archives and transcripts.

Status

Work in progress, not ready to try, developing in public, check back later 😎

Motivations

• I want an AI Interviewer, as a way to motivate myself to talk about a variety of topics that I'd like to share with the world.
• Experiment with AI in the browser via WASM.
• Contribute back to open source

Stack / Architecture

• SvelteKit - fullstack TS front + backend
• Deploy to Cloudflare Pages (at minimum)
• Audio recording via browser
• Audio transcription via
  • WASM LLM in browser or
  • LLM via API, eg OpenAI
• Next question generation via:
  • WASM LLM in browser or
  • LLM via API, eg OpenAI
• Save audio files to Cloudflare R2 (at minimum)

Features

Stage 1 - Text to Speech (AI) + Speech to Text (User)

• TTS (Text to Speech) via openAI in browser
• Works on cloudflare pages
• record user audio via browser
• STT user audio
  • Whisper WASM in browser
    • https://github.com/ggerganov/whisper.cpp/tree/master/examples/whisper.wasm
    • demo: https://whisper.ggerganov.com/
  • whisper API
  • Browser Speech API?
    • on chrome is sent to google servers
    • behind flag on Firefox

Stage 2 - LLM Question Generation

• feed convo to llm to determine next question
  • prefer: WASM llm - mistral7b or better
    • https://github.com/mlc-ai/web-llm
  • alt: gpt4[-turbo] api

Stage 3 - Audio Archives & Playback

• save audio (ai and human) to cf r2
• stitch together chunks in player to play back interview
• download full interview as mp3