This repo is for development for DeepSpeech on Edge Devices (such as RPi, RPi with neural network accelerators such as Intel Neural Compute Stick, Google Coral EdgeTPU
NOTE: This project is still on heavy development
The Convertion from .gb tensorflow file to .xml+.bin Intermediate Representation can be found here https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_DeepSpeech_From_Tensorflow.html
Note: When converting the model, remember to selecte FP16 for the data type as NCS doesn't support FP32 (the default value)
Memory: Due to limited memory on Neural Compute Stick and Complex speech to text model NCSDK Issue: mvNCCompile doesn't support transpose stage when compiling a graph model
- Resolve the memory issue
- Produce the real output after inference
Use NCSDK version instead.
Know more about NCSDK: https://movidius.github.io/ncsdk/index.html