run nue-asr with a long wav file on CPU

runnue.py

import nue_asr
import librosa
import numpy as np
import sys

# pip
#   torch
#   transformers
#   git+https://github.com/rinnakk/nue-asr

# git clone https://huggingface.co/rinna/nue-asr
modelName ='./nue-asr'
tokenizerName = './nue-asr'

myDevice = 'cpu'
wavFileName = './voice_s16le16kHz.wav'

# usage : python3 runnue.py [wavfile] [nueDir]
if len(sys.argv)>2:
	modelName = sys.argv[2]
	tokenizerName = modelName

if len(sys.argv)>1:
	wavFileName = sys.argv[1]



def simpleVAD(snd , tMin=1, tMax=16, sr=16000, wms=4, factTh=0.5):
	nMin = tMin/wms*1000//1
	nMax = tMax/wms*1000//1

	vRmsAll = (snd**2).mean()**0.5
	vThresh = vRmsAll*factTh

	lenWin = sr*wms//1000
	nWin = ((len(snd))//lenWin)
	nmod = len(snd) - nWin*lenWin

	vAryRms = np.zeros(nWin+1)
	vAryRms[:nWin] = (snd[:-nmod].reshape(nWin,lenWin)**2).mean(1)**0.5
	vAryRms[nWin] = (snd[len(snd)-nmod:]**2).mean()**0.5

	vflagValid = (vAryRms > vThresh)*1
	vStFlag = np.zeros(len(vflagValid))
	vStFlag[0] = vflagValid[0]
	vEndFlag = np.zeros(len(vflagValid))
	indxFlg = np.array(len(vflagValid))
	vStFlag[1:] = ((vflagValid[1:] - vflagValid[:-1]) == 1)*1
	vEndFlag[1:] = ((vflagValid[1:] - vflagValid[:-1]) == -1)*1

	aposStart=[]
	aposStop=[]
	bflagAct = False
	pPrevSt = 0

	for k,v in enumerate(vflagValid):
		if bflagAct == False:
			if vStFlag[k] ==1:
				bflagAct = True
				aposStart.append(k*lenWin)
				pPrevSt = k
			continue
		if bflagAct == True:
			if (k-pPrevSt) < nMin:
				continue
			if (k-pPrevSt) >= nMax:
				bflagAct = False
				aposStop.append(k*lenWin)
				continue
			if vEndFlag[k] ==1:
				bflagAct = False
				aposStop.append(k*lenWin)

	if len(aposStart) > len(aposStop):
		aposStop.append(len(snd))

	retList =[]
	for k in range(len(aposStart)):
		retList.append([aposStart[k],aposStop[k]])

	return retList


sampleRate=16000
snd,sr = librosa.load(wavFileName, sr=sampleRate)

periodSecMin = 5
periodSecMax = 16
windowMilliSec = 4
threshSoundRms = 0.5
segSp = simpleVAD(snd,periodSecMin,periodSecMax,sampleRate,windowMilliSec,threshSoundRms)

lstSnd =[]
for c in segSp:
	lstSnd.append(snd[c[0]:c[1]])


model = nue_asr.load_model(modelName, device= myDevice,fp16=False)
tokenizer = nue_asr.load_tokenizer(tokenizerName)
# result = nue_asr.transcribe(model, tokenizer, wavFileName)
# print(result.text)

for snda in lstSnd:
	result = nue_asr.transcribe(model, tokenizer, snda)
	print(result.text)