vgene/ASR-Inference-for-Edge

Introduction

Automated speech recognition inference for edge devices including Jetson TX1 and Altera Arria 10, etc.

This project modified from following repo. I also used that repo to complete the training process. This modified project is specificly designed for inference process on edge devices. Redundant dependencies are removed.

https://github.com/zzw922cn/Automatic_Speech_Recognition

Make sure you are using Python 3.5!

File Description

Requirements

tensorflow-gpu edit numpy scipy edit_distance

Preparation

Need to install sox sudo apt-get install sox

FOR TX1:

Install Lapack: sudo apt-get install liblapack-dev
Install requirements: pip install -r requirements-tx1.txt
Install TensorFlow: https://github.com/jetsonhacks/installTensorFlowJetsonTX/tree/master/TX1

FOR AMD64:

pip install -r requirements.txt

Usage

Configure args in libri_inference.py, including audio file path and model path.

Run: python3 libri_inference.py

Procedure

Flac to wav

Channels : 1 Sample Rate : 16000 Precision : 16-bit Bit Rate : 256k Sample Encoding: 16-bit Signed Integer PCM Endian Type : little Reverse Nibbles: no Reverse Bits : no
txt split
to lower case
getfeature

import scipy.io.wavfile as wav
from sklearn import preprocessing
from calcmfcc import calcfeat_delta_delta

(rate,sig)= wav.read(fullFilename)
# mode = 'mfcc','fbank'
# feature_len =13
# win_step = 0.01
# win_len = 0.02
feat = calcfeat_delta_delta(sig,rate,
    win_length=win_len,win_step=win_step,mode=mode,feature_len=feature_len)
feat = preprocessing.scale(feat)
feat = np.transpose(feat)
#print(feat.shape)

vgene / ASR-Inference-for-Edge