So I built speaklanguageonline.com - a voice call with an AI that speaks Thai (or Vietnamese), listens to your attempts, and gives you one gentle correction at a time. No typing, no flashcards, just talking.
Tech:
OpenAI Realtime API (WebRTC) for speech-to-speech Next.js 14 + Vercel No transcription step - the model processes audio directly, which preserves tone (critical for Thai's 5 tones)
What makes it different from ChatGPT voice:
- Tuned for slow, patient corrections (not conversational chat) - One correction per turn (anxiety-inducing to get 5 things wrong at once) - Corrections explained in your native language - 3-minute session cap to keep it focused
Pricing: Credits, not subscriptions. You pay only for minutes used. I hate subscription guilt as much as you do.
Current state: Thai and Vietnamese work well. Adding Spanish, Hindi, Mandarin soon — OpenAI's model handles them but quality varies.
What I learned:
Realtime API latency is ~300-500ms which feels natural for conversation Prompting for "one correction only" took way more iteration than expected Tonal languages need explicit instruction to focus on tone mistakes
Would love feedback, especially from anyone who's built voice-first apps or is learning a language.