← Hemant Bansal

Bugs I've loved

A running list of the strange ones. Most only showed up on real phones.

  1. The microphone that crashed with earbuds in

    Our speech-to-text library crashed on Android, but only with Bluetooth audio connected. Fixed with two null checks inside the library, using an unmerged fix from a GitHub issue. Read the note

    fixed · Android · native

  2. Eight seconds of silence

    The voice agent worked, but every step waited for the one before it. Streaming each step, then moving to the OpenAI Realtime API, brought it to about 1.5 seconds. Read the note

    fixed · latency

  3. Human voices, slower answers

    Swapping robotic device voices for ElevenLabs and PlayHT made the agent sound great and made it slow again. Streaming the audio fixed it.

    fixed · trade-off

  4. "Am I online?"

    The connection status we relied on was not accurate enough for a live voice call. I wrote a custom useNetwork hook that tracks the connection in real time.

    fixed · React hook

  5. Long sentences, noisy rooms, wrong language

    Android speech-to-text crashed on long recordings, struggled with background noise and sometimes guessed the wrong language. No single fix: I worked through them one at a time.

    fixed, slowly · Android

  6. A beta with opinions

    Running the OpenAI Realtime API in beta for months meant handling noise, conversations that switch languages halfway, and a model that did not always follow its instructions.

    tamed · AI

  7. The alert that had to be heard

    Not a bug, but the same kind of fight: making critical alerts ring on a phone set to Do Not Disturb. Read the note

    shipped · Android