Hi,
I tried GPT-Live in the ChatGPT app on iPhone, and I’m building a hardware device, a wearable with its own microphone and speaker, that I plan to connect to the GPT-Live API once it’s released (I’ve signed up on the notification form). I’m designing the audio pipeline now, so I’d like to understand the expected client-side contract for echo before freezing the hardware.
The issue: when the model’s spoken response plays out of my device’s speaker, some of that audio inevitably leaks back into the microphone. And since GPT-Live is full-duplex (listening while it speaks), its own voice would be present in the uplink continuously, not just at turn boundaries. My questions:
- Will GPT-Live handle this server-side, i.e., recognize and tolerate its own response audio in the incoming mic stream? Or will it assume the uplink is already echo-free?
- In the ChatGPT app, is GPT-Live relying on the phone OS’s built-in acoustic echo cancellation (voice-processing audio session)? If so, I’ll assume a custom device with no OS-level AEC must implement its own.
For context, my device won’t have a phone in the audio path: mic audio will go straight from my hardware to the endpoint. So if echo handling is a client-side responsibility, I’ll implement AEC in firmware, and I’d like to confirm that’s the expected architecture.
Thanks!
Ram