OCR extremely slow / returns no text on macOS 27 Golden Gate (Vision VNRecognizeTextRequest regression)

[Jason-K]

Summary

Since upgrading to macOS 27 Golden Gate, macshot's OCR has become extremely slow and unreliable. The cause is an OS-side regression in Apple's Vision framework: the legacy VNRecognizeTextRequest at its default revision (revision 3) with .accurate recognition. The API behaves differently on 27, so the app needs a workaround.

Environment

• macOS 27.0 Golden Gate
• Apple Silicon (M4 Pro, 24 GB)

Behavior

• The first .accurate request in each process stalls for about 15–30 seconds while the system compiles the model through the Neural Engine (E5RT).
• Later requests in the same process return zero results with no error, or on some machines throw E5RT ... Code: 13, until the process restarts.

Reproduction (outside macshot)

A standalone PyObjC test ran sequential requests in one process against the same screenshot:

Request	Time	Result
.accurate, default revision, #1	26.9 s	0 text blocks
.accurate, default revision, #2–3	0.13 s	0 text blocks
.accurate, VNRecognizeTextRequestRevision2	0.12 s	3 text blocks ✅
.fast, default revision	0.07 s	6 text blocks ✅

Proposed fix

TRex hit the same problem and fixed it in amebalabs/TRex#95. The same regression is also reported in m-tkg/clipkun and jfarcand/mirroir-mcp#36.

1. On macOS 15+, use the modern Swift Vision API (RecognizeTextRequest). It isn't affected and keeps full language support. Leave the existing path in place for macOS 12.3–14.
2. Wrap each OCR attempt in a timeout of about 10 s. On timeout, error, or an empty result, retry once at .fast so a stalled model compile can't hang the OCR window.
3. Fallback if the legacy API must stay: pin request.revision = VNRecognizeTextRequestRevision2. It's correct and fast on 27, but lacks some languages such as Cyrillic.

import Vision

func recognizeText(in cgImage: CGImage) async throws -> String {
    if #available(macOS 15.0, *) {
        var request = RecognizeTextRequest()
        request.recognitionLevel = .accurate
        request.usesLanguageCorrection = true
        let observations = try await request.perform(on: cgImage)
        return observations
            .compactMap { $0.topCandidates(1).first?.string }
            .joined(separator: "\n")
    } else {
        let request = VNRecognizeTextRequest()
        request.recognitionLevel = .accurate
        request.usesLanguageCorrection = true
        let handler = VNImageRequestHandler(cgImage: cgImage)
        try handler.perform([request])
        return (request.results ?? [])
            .compactMap { $0.topCandidates(1).first?.string }
            .joined(separator: "\n")
    }
}

Features that depend on text recognition, such as auto-redact PII, the OCR/QR results window, and OCR-translate, are likely affected too.

I'm happy to test a build.

[sw33tLie]

It really sucks we have to do something here instead of just having Apple ship a hot fix for this, but yeah I noticed it too...will investigate, thank you

[sw33tLie]

Fixed in the latest beta. Thanks!