# D54. Ship Intel Macs with media tools, without local AI (17 Oct 2026)

*Revised on 19 Oct 2026: refusing the word aligner is remembered for the whole app (per language), and the aligner is never offered right after the voices' offer; it waits for the next local generation in a later opening. The rest of this entry is kept as it was decided.*

Release a Mac Intel build to learn whether schools and other authors form a significant user base. Do not maintain an older ONNX Runtime branch for this platform: local voice generation (Pocket TTS), forced word alignment (wav2vec2) and local transcription (Parakeet) are unavailable on darwin/x64, even if old models or a working native binding remain. The app explains this in downloaded elements, offers no installation or local voice selection, and the core refuses installation and inference too (CLI/MCP included). Models already downloaded can still be removed. Apple Silicon and Windows x64 retain local AI; recordings, existing narration audio and word times, editing, playback, media conversion, publishing and explicitly chosen online services remain available on Intel. No new usage collection is introduced by this change.

The existing checksummed prebuilt Intel FFmpeg/ffprobe remains pinned independently of the other platforms. Its supplier stopping new Intel builds does not make the shipped tools stop working. Reassess this pin for a concrete decoding defect, compatibility issue or necessary fix, rather than automatically rebuilding FFmpeg solely to follow releases. This supersedes the Intel-motivated runtime constraint in D25; upgrading ONNX Runtime itself still needs validation of the three local models on the supported platforms.

Before a requested local voice generation, the app offers the missing word aligner once per language and window, with an explicit choice to continue using estimated word times. Installing voices from the first-use offer leads to the same alignment offer before generation resumes. The aligner analyzes the produced recording; it does not generate or modify the voice waveform. Pocket's retry checks reject silence, non-finite samples and generations without an end marker; after all three attempts fail, no take is written. These checks do not establish intelligible speech: a loud whistle can still pass them. The reported Henri/Marc take from diagnostic-reseau on a 2019/2020 Intel Mac has not been reproduced or attributed to a native-runtime defect.

*Revisit if:* Intel usage justifies maintaining local inference, another optional local element does not need ONNX Runtime, or a measured quality/performance benefit calls for a newer runtime or a stronger speech-quality check.