We trained two on-device AI models in-house for Detail and Subwave: one cleans audio, one finds filler words, and both run entirely on your iPhone at zero variable cost.
Two new on-device AI models are shipping this week inside Detail and Subwave. We trained both of them in-house, and both are small enough to run entirely on your iPhone, which means your recording never leaves your device. Clear enhances the sound of a recording. Uhm finds all the filler words you didn't even notice you were saying.
Enhance a 10 minute recording in 7 seconds
A year ago, building these models ourselves would not have made any sense. The benchmark for audio enhancement was Dolby, a server-side service that costs us thousands of euros a year on Detail alone and requires uploading and retrieving every recording to process it. Per-minute pricing on audio enhancement limits what you can offer inside a product like Subwave, where every post might get cleaned and we don't charge for the product yet. So we built our own.
Four things became reachable for a small team over the past year. Open source base models we could build on. Training data we already had from years of Detail recordings. Hardware fast enough to train models overnight. And tools like Claude doing the engineering scaffolding around the model itself, including orchestrating training runs, preparing data, and running the pipelines.
With all four within reach of one person, the bottleneck stops being compute and becomes taste.
AI features are commoditizing fast. Anyone can ship a transcription service, pull a generic audio enhancement off the shelf, or call an API to remove filler words. What doesn't commoditize is the taste. The hundreds of small decisions about what the final output should sound like, where the user gets to make a choice, and where the product should just do the right thing.
When you train your own model, every parameter and every piece of training data becomes a lever in your product.
For our audio enhancement model, the most important parameter wasn't speed or size. Most audio enhancement tools produce a result that sounds flat or neutral, broadcast-saved in a thin way. That's not what we want a Detail recording to sound like.
A Detail recording should sound warm and present, like you're in the room, closer to a podcast studio than an actual phone call.
So the clean reference that our model learns from is itself lightly enhanced and loudness-normalized, shaped to the sound we want it to reach. The model learns the sound of Detail, not the sound of clean speech in general.
The iteration loop looked a lot more like product iteration than research. I would train overnight, listen back in the morning to tens of Detail recordings, compare the new model against the previous best, against Dolby, and against the unprocessed source. To do this, I built an audition tool and set up a blind A/B test with outputs labeled blindly A or B. We'd pick the version that sounded best on the kinds of recordings we typically find in our app, then rinse and repeat.
For filler word detection, I evaluated hundreds of clips by hand to validate quality and improve it. The signal that ended a training run was the blind audition the next morning, not a number in the model card or a line in an API doc.
We ended up with a 5 MB model running on your iPhone, competitive with a server-side enhancement that charges by the minute, at zero variable cost to us. Our filler word detection model processes a 57-minute interview in under 30 seconds.
iPhone 17 Pro benchmarks for Clear, on-device audio enhancement
Because the models are ours and cost us nothing to run, we can make product choices that wouldn't be possible if we were paying by the minute.
In Detail, auto-edit runs both models without you asking. Fillers come out, audio gets cleaned, and you get a usable edit on the first pass. If you want manual control, it's there: step through fillers one by one in transcript view, keep the ones that feel like you, or move a slider on the audio enhancement anywhere from nearly no processing to a podcast studio sound.
In Subwave, every post gets cleaned by default if you want it to. The cost is zero, so nothing pushes against making the good version the default.
What we're shipping in our products is not the model. It's the sound of a Detail or Subwave recording, and the editing taste of a post. The model is how we get there without paying a per-minute tax to a generic service that produces the same output for every app.
The interesting move in AI right now isn't bigger generic models. It's swapping the large generic model for a small local one, trained on the data you have, tuned to the product you actually ship.
Building AI models became a product design challenge.
