About Dalm

Dalm is a Mac app that records meetings, writes summaries from them, and types what you dictate into any app.

Why we build it

Most of what gets decided at work is said out loud: on a call, across a table, or into a microphone when talking is faster than typing. Then it is gone, or it sits in a recording nobody plays back. We build Dalm to keep what was said and make it useful twice: right after the meeting, as a summary you can act on, and months later, when you need to find who said what.

We would rather show how something works than describe it in large words. That is why this page describes how each number is measured before it shows any number, and why a figure we cannot back with a written protocol stays off it.

How we measure

Word error rate is the share of words a transcript gets wrong — left out, added or replaced — counted against a reference transcript. We build the reference from the agreement of two different speech models, so that no model is graded against its own output.

Processing speed is measured on whole recordings, not on one step: the time from the end of a recording to a transcript with the voices separated. It covers both steps, transcription and separating voices, and says nothing about how accurate the separation is.

Dictation latency is the time from letting go of the key to the text appearing in the field, reported at the median and at the 95th percentile.

Each of the three gets a figure here only together with its protocol: which Mac, how many runs, and how you could check it yourself. Until then the methods are all we publish.

What we don’t do

Dalm is made for one kind of computer: a Mac with Apple Silicon · macOS 15.0 or later. The full list is on system requirements.

We do not publish figures we have not measured with a method we could show you: battery life, memory use over a long meeting, or accuracy in languages we have not tested. Where your recordings and your text go, step by step, is set out on how it works.