My customer is planning to use i.MX RT for running a TensorFlow model for Voice/Audio Recognition but
I have noticed that eIQ Toolkit and eIQ Portal are suited for Image Classification/detection.
Could you please advise what is the correct process to works with the eIQ toolkit to handle Audio/Voice as input?
Waiting for your kind responses, Thanks in advance.
MCUXpresso SDK has tensorflow lite for microcontroller (TFLm) examples. If your model does not contain operators that is not supported by TFLm, in general it could be inferenced, and note that you may need some pre-processing and post-processing code besides the model inference.
An important note is:in the eIQ example, it configures TFLm to include only the operators needed by the example model (mobilenet) to save flash size, it is VERY LIKELY that your model use other operators so the safest way is modify the code in "model.cpp" to use "tflite::AllOpsResolver" instead of MODEL_GetOpsResolver() (in model_mobilenet_ops_micro.cpp). After you know the exact operators required by your model, you can also specify your version of MODEL_GetOpsResolver() to only include needed operators to save flash.
Currently eIQ Portal can only be used to generate vision based models for classification/detection. Audio/voice models would need to be created using TensorFlow or Pytorch, an example of which can be found here: https://www.tensorflow.org/tutorials/audio/simple_audio
Once the model is created, then eIQ can be used to run the model like found in the i.MX and i.MX RT examples.
There is any update regards my question above?
in addition to that, could the eIQ toolkit import a model that has been already trained to handle Audio, can it be possible to run via MCU TensorFlow lite micro inference?
Waiting for your kind feedback, Thanks in advance