yukarisonz / edgedeepspeech

This repo is for development for DeepSpeech on Edge Devices (such as RPi, RPi with accelerators)

EdgeDeepSpeech

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)

Current Issue

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

TODO

• Resolve the memory issue
• Produce the real output after inference

UPDATE

Use NCSDK version instead.

Know more about NCSDK: https://movidius.github.io/ncsdk/index.html

edgedeepspeech's Issues

Terminate without any exception on windows Machine

Memory Issue on RPi

When run it on Neural Compute Stick, it got memory issue. It might because of the limited memory on NCS and complex speech-to-text model

Log:
$ python3 main.py -m output_graph.xml -i LDC93S1.wav -d MYRIAD
*** Error in `python3': malloc(): memory corruption (fast): 0x0153bb90 ***
Aborted