Inference

Listen provides three types of inference methods: inference, async inference, and batch. We assume that we have already finished the initialization and audio recording process.

val listen = Listen(context)
listen.load("SDK_KEY", "MODEL_ASSETS_PATH")

val audioSamples = getAudioSamplesForMic() // get audio samples from mic using AudioRecord

Simple Inference

The simplest and most straightforward method is synchronous inference.

// audioSamples must have fixed (pre-defined) sample rate and must be larger than or equal to the minimum input size, 
// which are available from listen.getAudioParams().sampleRate and listen.getAudioParams().minInputSize, respectively.
val results = listen.inference(audioSamples)

Please note that this simple inference method inference() runs multiple inferences with variable-length audio samples which is at least minInpustSize. You can get the minimum audio sample input size via getAudioParams().minInputSize property.

Asynchronous Inference

If we want to detect a specific sound event in real-time, an asynchronous manner is better for sound event detection, especially because we cannot estimate when the sound event occurs. Asynchronous detection is performed in the following manner:

  1. Register a listener using setAsyncInferenceListener()

  2. Pass audio samples recorded in real-time to inferenceAsync(audioSamples)

  3. Listener or Kotlin coroutine will be invoked when the registered sound event is detected