def processTrainingSet(self, basedir="sounds/train/", signalword="oknavsa", savedir="data/"):
clips = [clip for clip in os.listdir(basedir) if clip.endswith(".wav")]
try:
with open("sounds/train/thresh.dat","r") as fh:
lines = [line.strip() for line in fh.readlines() if len(line)>1]
bestThresh = {line.split(",")[0]:int(line.split(",")[1]) for line in lines}
except:
bestThresh = {}
tr = Trigger()
self.YXtot = []
fnames = []
durations = []
self.keywordDurations = []
with warnings.catch_warnings():
warnings.filterwarnings("ignore",category=Warning)
for clip in clips:
if clip in bestThresh: tr.setParams({"THRESHOLD": bestThresh[clip]})
else: tr.setParams({"THRESHOLD": 600})
if clip.lower().startswith(signalword): isSignal = True
elif clip.lower().startswith("random"): isSignal = False
elif clip.lower().startswith("background"): isSignal = False
else: continue
tr.readWav(basedir+clip)
subsamples = tr.getSubsamples()
framerate = tr.getFramerate()
if self.verbosity > 1: print "Loading clip %s (isSignal: %i) ==> %i subsamples" % (clip, isSignal, len(subsamples))
for ss in subsamples:
duration = self.getSampleDuration(ss, framerate)
if isSignal: self.keywordDurations.append(duration)
self.YXtot.append( self.getFeatures(ss,framerate,isSignal) )
fnames.append(clip)
durations.append(duration)
self.YXtot = np.array(self.YXtot)
self.keywordDurations = np.array(self.keywordDurations)
outputname = "%simagedata_%i_%i.npy" % (savedir,self.Nfreq,self.Ntime)
outputname_meta = "%smetadata_%i_%i.npy" % (savedir,self.Nfreq,self.Ntime)
np.save(outputname, self.YXtot)
np.save(outputname_meta, self.keywordDurations)
if self.verbosity > 1: print "made %s and %s" % (outputname, outputname_meta)
idx_test, YXtest = self.trainAndTest()
return np.array(fnames)[idx_test], np.array(durations)[idx_test], YXtest
parser = Parser()
if haveLeds:
from Lights import Lights
led = Lights()
led.stop()
tr = Trigger(verbose=True, led=led, AMBIENT_MULT = 1.9)
else:
tr = Trigger(verbose=True)
# proc.processTrainingSet(basedir="sounds/train/", signalword="oknavsa", savedir="data/")
# proc.processTrainingSet(basedir="16khz/", signalword="oknavsa", savedir="data/")
proc.loadTrainData("data/imagedata_15_15.npy")
#if not in this range, we want to not fingerprint it to save time and trouble
lower,upper = proc.getKeywordDurationRange()
tr.setParams({"MIN_WORD_TIME": lower, "MAX_WORD_TIME": upper})
tr.setParams({"THRESHOLD": 150})
# tr.getAmbientLevels(duration=0.7)
print "Now will score realtime audio"
def myCallback(trigger, data, data_raw):
print
framerate = trigger.getFramerate()
if not trigger.hasSaidKeyword():
t0 = time.time()
confidence = proc.getKeywordProbability(data, framerate)
print "took %.2fms to classify %.2fs clip, score: %.2f" \
% (1000.0*(time.time()-t0), 1.0*len(data)/framerate, confidence)
