SFSpeechRecognizer on Monterrey error 102

Question

I have updated to macOS Monterrey and my code for SFSPeechRecognizer just broke. I get this error if I try to configure an offline speech recognizer for macOS

Error Domain=kLSRErrorDomain Code=102 "Failed to access assets" UserInfo={NSLocalizedDescription=Failed to access assets, NSUnderlyingError=0x6000003c5710 {Error Domain=kLSRErrorDomain Code=102 "No asset installed for language=es-ES" UserInfo={NSLocalizedDescription=No asset installed for language=es-ES}}}

Here is a code snippet from a demo project:

private func process(url: URL) throws {
    speech = SFSpeechRecognizer.init(locale: Locale(identifier: "es-ES"))
    speech.supportsOnDeviceRecognition = true
    let request = SFSpeechURLRecognitionRequest(url: url)
    request.requiresOnDeviceRecognition = true
    request.shouldReportPartialResults = false
    speech.recognitionTask(with: request) { result, error in
      guard let result = result else {
        if let error = error {
          print(error)
          return
        }
        return
      }

      if let error = error {
        print(error)
        return
      }

      if result.isFinal {
        print(result.bestTranscription.formattedString)
      }
    }
  }

I have tried with different languages (es-ES, en-US) and it says the same error each time.

Any idea on how to install these assets or how to fix this?

Answer 1

The issue is that you're forcing it to run offline.

request.requiresOnDeviceRecognition = true

This capability requires the A12 bionic, M1, or newer SoC, because these have a neural engine. On devices without a neural engine, you need to perform speech recognition on Apple's servers.

Answer 2

This suggestion offered in Apple's dev forums here worked, at least partially, for me:

After spending a lot of time, what worked for me was enabling and onboarding Siri for iOS. Then downloading models using Settings -> Translate -> (set on-device mode to true), then download models for the specific locale. Then run the app and waiting for some time for the models to download. (It doesn't work immediately).

I say it worked partially for me, because if I relaunch the app a few days later, the same error message appears. I need to leave the app running idle for quite a bunch of seconds (usually over a minute), which sort of makes it impossible to launch this to final users.