Give a spoken instruction room to finish

The basic voice workflow was straightforward: capture speech, send the instruction to an assistant and play the response. The harder requirement was keeping that interaction usable on a phone, where people pause mid-sentence, change audio routes and receive calls.

PitchAI extended the native interface with controls around the captured speech. The work included manual and automatic submission, transcript buffering and recovery when the audio or transcription connection changed.

In automatic mode, completed speech enters a four-second composition window. If speech resumes, the pending submission is cancelled. The next completed portion can extend the instruction before it is sent.

A spoken send command can bypass that wait. This gives the interface two ways to handle intent: leave room for another sentence, or let the speaker explicitly finish the request.

Before restarting transcription, the client moves the latest unfinished transcript into its capture buffer. It compares that text with what is already buffered so an overlapping partial can be merged rather than blindly appended.

A late final transcript can also contain words that were submitted earlier. The client checks recently submitted speech and removes a matching prefix while retaining newly appended words. That distinction matters when the transcript provider revises a whole sentence after the user has already sent its first part.

Later interruption work addressed phone calls. Recovery tracks whether an interruption is active, whether restoration is pending and how many attempts have failed. Retry delays increase to a fixed ceiling, and reconnect sounds remain suppressed until restoration succeeds.

The native app and its Watch companion were signed, installed and checked on physical devices. Delivery records include a complete speech-to-assistant-to-audio turn; the interruption release also records resumed microphone frames after installation.

The resulting product does more than wrap a microphone around a chat box. It gives spoken instructions a composition window, preserves pending text across transcription restarts and makes recovery a defined part of the interaction. Those mechanisms reduce specific failure risks; they are not a guarantee of perfect transcription or uninterrupted service.