NLU on iOS

This is a companion to the NLU concept guide, which discusses the NLU subsystem holistically. Here we’ll talk about usage issues specific to the iOS client library.


The Spokestack NLU most commonly uses the following SpeechConfiguration properties:

  • nluVocabularyPath: The bundle path to the model vocabulary
  • nluModelPath: The bundle path to the nlu model
  • nluModelMetadataPath: The bundle path to the nlu model metadata

The following properties are configurable, but should not be changed unless a different NLU model or vocabulary is being used (i.e., if you’re changing these, you already know why you’re doing so):

  • nluTerminatorTokenIndex
  • nluPaddingTokenIndex
  • nluMaxTokenLength


As mentioned in the Getting Started guide, initializing the Spokestack NLU is like initializing the TTS component. Both a delegate callback interface and a publisher interface are provided. The call to classify is the same for either interface, but the result and DispatchQueue that the classification runs on are handled differently.


The controller initialization is straightforward—simply tell the SpeechConfiguration where to find the three files required for the Spokestack NLU model you’re using (here we assume that they’re contained in the app’s bundle with hardcoded names), and then initialize the controller.

class NLU {
    public let config = SpeechConfiguration()
    config.nluVocabularyPath = Bundle(for: type(of: self)).path(forResource: "vocab", ofType: "txt")
    config.nluModelPath = Bundle(for: type(of: self)).path(forResource: "nlu", ofType: "tflite")
   config.nluModelMetadataPath = Bundle(for: type(of: self)).path(forResource: "nlu", ofType: "json")
    public let nlu = try! NLUTensorflow(self, configuration: config)

1) Delegate

A delegate interface is as simple as:

extension NLU:  NLUDelegate {

    nlu.classify(utterance: "turn the lights on in the kitchen")

    func classification(result: NLUResult) {
        // Utilize the result intent and slot(s) as appropriate

    func failure(error: Error) {
        // respond appropriately to an error in classification

    func didTrace(_ trace: String) {
        // respond appropriately to a debug tracing message


When it comes time to classify an utterance using the delegate interface, Spokestack’s NLU does all the heavy lifting on a configurable async DispatchQueue. Set SpeechConfiguration.delegateDispatchQueue to a DispatchQueue with your desired QoS. The classify function will run and call your delegate in that queue. By default NLUTensorflow.classify will run with a QoS of userInitiated.

2) Publisher (Combine)

The publisher interface to the classifier can take advantage of a simpler setup, and is able to classify multiple utterances with a single call. Note that the call to classify uses subscribe(on:) to ensure it is not blocking the UI thread.

// using Combine

let utterances = ["turn the lights on in the kitchen"]

let _ = nlu.classify(utterances: utterances)
    .subscribe(on: .userInitiated))
    .sink(receiveCompletion: { completion in
        switch completion {
            case .failure(let error):
                // respond appropriately to an error in classification
            case .finished:
                // respond appropriately to finished classification
    }, receiveValue: { results in
        let _ ={
            // Utilize the result's intents and slots as appropriate
Something missing here?Edit this doc!Questions? Visit our forum