Model choice is a measurement problem
Alder’s background is data science, not model worship. We don’t assume the newest or largest speech model is automatically the best. Candidate engines are run on the same recordings, and the results are segmented by language — because an engine that wins on English narration can lose badly on Chinese, and an engine that handles one language beautifully may not exist for another at a size a laptop can run.
We also evaluate mixed-language speech separately. Many real people speak the way they think — an English sentence with a product name in Japanese, a Chinese sentence that borrows English engineering vocabulary. Mainstream speech benchmarks are largely monolingual and Western-centric, so a model’s leaderboard rank simply doesn’t predict how it behaves when languages switch mid-sentence. If your users code-switch, you have to test code-switching yourself.
These are controlled Apple M1 Max measurements using identical, clean synthetic 16 kHz audio. Speed is model decode time, not the complete dictation workflow; real-world results vary by voice, accent, microphone, and environment.
Quality is the gate; size and speed break ties
The selection rule is deliberately boring. An engine must first clear a quality bar for a language — measured on the same controlled recordings, not marketing claims. Only engines that clear that bar compete on the second axis: end-to-end latency, memory footprint, and download size. A faster engine never wins by being worse, and a bigger engine never wins by being bigger.
No single model won everywhere. So we stopped asking “which model is best?” and started asking “which model is best for this language?”
Users declare languages. Engines are derived state.
The most important product decision followed from the measurements: nobody should have to pick a model. In Alder you declare the languages you speak — the engine becomes derived state. A universal multilingual engine covers all 100 language routes; where a specialized engine measurably beats it for your language profile, Alder routes there instead. If a specialized route is ever unavailable, recognition degrades gracefully to the universal engine rather than failing.
For fixed-language profiles, the choice follows the languages you declare. With Automatic Language enabled, a lightweight acoustic check can choose a supported route for each recording. Both paths stay on your Mac, and a universal engine remains ready whenever a specialized route is unavailable or does not clear its quality gate.
What this means when you dictate
- Short takes in a fixed language stay stable, because the engine was chosen for that language.
- Mixed-language speech goes to the engine that handled code-switching best in our tests.
- Recognition quality and speed can improve per language over time without you re-configuring anything.
- Everything above runs on your Mac — model routing never requires a server.
More on how the models reach your Mac in The hard part of on-device AI isn’t inference, and on the full pipeline in What is Alder?