gistfile1.py

import simplejson as json
import itertools
LABELS = ['howard', 'leonard', 'non_speech',
          'other', 'penny', 'raj', 'sheldon']

def loss(gold, pred, durations=None):

    # convert '0' to 'howard', '1' to 'leonard', etc...
    gold = [LABELS[int(k)] for k in gold]
    pred = [LABELS[int(k)] for k in pred]

    # use segment durations if provided
    # otherwise artifically set them to 1s (or 250ms, it does not matter)
    if durations is None:
        durations = [1.] * len(gold)

    denominator = 0.
    numerator = 0.

    for g, p, d in itertools.izip(gold, pred, durations):

        # correctly classified as non speech --> no error
        if g == 'non_speech' and p == 'non_speech':
            continue
        
        # increment denominator with duration of 'speech' segments
        if g != 'non_speech':
            denominator += d

        # increment error duration in case of error
        numerator += (g != p) * d

    return 100 * numerator / denominator

X = np.load('....X.npy')
durations = X[:, 0]

pred, gold = json.load(open('ep=1.json'))

approxIER = loss(gold, pred, durations=durations)