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.
This small safety set covered those controlled cases, not every possible utterance. In the product, cleanup guards reject unsafe transformations and keep the original transcript instead.
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.