Engineering notes · Updated August 25, 2026

A dictation app is judged on the one time it loses your words.

Nine hundred perfect dictations don’t matter if the app loses one. The moment a spoken thought disappears, trust is gone. So Alder is engineered around a single promise: once you said it, it survives — through microphone failures, interruptions, and text fields that refuse to accept the result.

The microphone is the first thing that breaks

The speech models were never our biggest source of production issues. Audio capture was. A Bluetooth headset can connect, report itself ready, and then never deliver a single audio buffer. A microphone can disconnect mid-sentence. macOS can route audio somewhere unexpected the moment another device appears.

Alder’s answer is to trust evidence, not status: recording doesn’t count as started until audio actually arrives, input levels are normalized across devices, and a device disconnect is detected instead of silently producing an empty take. If the signal is silence or noise, Alder says so — it will not hand you an empty result and call it success.

Interruptions must not destroy the take

People quit apps mid-sentence. Macs sleep. Sessions get interrupted. Alder writes dictation audio crash-safely while you speak, so an interrupted session leaves a recoverable recording instead of nothing. On the next launch, Alder notices the orphaned take and offers to process it — the thought you spoke yesterday is still there today.

The promise is simple: you spoke it, you keep it. Everything else is implementation detail.

Bad output is blocked before it reaches you

Reliability isn’t only about keeping audio — it’s about refusing to deliver garbage. Local speech models can produce pathological output on silence or noise. Alder quality-gates the transcript before insertion: output that fails those checks is blocked rather than typed into your document. Cleanup is held to the same standard — it removes high-confidence fillers and repeats, and preserves your meaning instead of rewriting it.

Delivery is verified, never assumed

The last step is the most delicate: putting text where you meant it. Before inserting, Alder checks that the app where you started dictating still has focus. If you switched apps mid-transcription, the result is not typed into the wrong window — it stays on the result card, visible and ready to place. If a field blocks insertion entirely, the result remains available to copy, and your history keeps every take.

None of this shows up in a feature list, and that’s the point — it’s the engineering that makes the feature list true. For how the models themselves are chosen and delivered, read Why Alder runs multiple speech models and The hard part of on-device AI isn’t inference.