Phonon 31M: open TTS running on Apple's Neural Engine via Swift
Post on a Swift port of the open Phonon 31M TTS model, using CoreML on the Neural Engine, with 103 voices and latency and speed figures stated by the author.
Source: Phonon 31M: modelo TTS aberto rodando na Apple Neural Engine em Swift (github.com). Text prepared with AI from this source.
What happened and what to do
The post presents a Swift port of the open text-to-speech model Phonon 31M, running through CoreML on Apple's Neural Engine. According to the author, the implementation offers 103 voices and a text normalizer ported from Gradium, with about 20 ms to first audio and about 20x real time on an M5 Pro. The code and model are linked in the post itself.
For a company with Apple apps or devices, the case points to fully local speech synthesis, without sending text to external services. In practice, this means checking whether the quality, languages and voices fit the product, measuring latency and power use on real hardware, and defining a text normalization pipeline for numbers, dates and domain acronyms. A pilot can validate these points before any production integration.
How the consultancy can help
Diagnosis of where synthesized voice would add value in your Apple product and which data would leave the device or stay on it. Then, a scoped implementation of a local TTS prototype, with benchmarks for latency, quality and power use on your target hardware, and support for operating model updates.
Next step
Send a short description of your case, including platform, usage volume and privacy requirements, and receive a scoped proposal with timelines and contracting options.
Let's assess your project
I reply with scope and a proposal.