ENGINEERING NOTES
Long meetings are a queueing problem, too
A short speech-to-text demo can look straightforward: capture audio, transcribe it, send text to a language model and show the result. In a long conversation, the next piece of speech arrives before that sequence has finished.
People do not wait for the API
If every new segment starts another request, work can pile up. A fixed queue limit creates another problem: what happens to the words beyond the limit? Latency, memory, request volume and lost context are connected.
My TalkEcho changes use one active audio batch and one pending batch that can grow as new segments arrive. When the pending batch becomes active, its transcripts are combined into one AI request. This is an attempt to catch up without creating a separate request for every arriving segment.
Retries need to have a reason
A temporary connection failure is different from an invalid API key. The project retries transient speech-to-text failures once and puts bounds on request duration. For AI streaming, retrying is limited to failures before any response text appears.
Those limits are a way of keeping the tool understandable under pressure. A captioning tool is useful when it helps someone follow a conversation, not when its internal queue becomes another thing to manage.
TalkEcho is my personal fork of Pluely. The original project provides the foundation; the public repository records the upstream credits and the Windows, dictation and request-handling work.
Implementation context: TalkEcho meeting request resilience.
Prepared for this portfolio with AI assistance, based on the public project descriptions.