Spaces:
Runtime error
title: WER
emoji: 🤗
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 3.0.2
app_file: app.py
pinned: false
tags:
- evaluate
- metric
Metric Card for WER
Metric description
Word error rate (WER) is a common metric of the performance of an automatic speech recognition (ASR) system.
The general difficulty of measuring the performance of ASR systems lies in the fact that the recognized word sequence can have a different length from the reference word sequence (supposedly the correct one). The WER is derived from the Levenshtein distance, working at the word level.
This problem is solved by first aligning the recognized word sequence with the reference (spoken) word sequence using dynamic string alignment. Examination of this issue is seen through a theory called the power law that states the correlation between perplexity and word error rate (see this article for further information).
Word error rate can then be computed as:
WER = (S + D + I) / N = (S + D + I) / (S + D + C)
where
S is the number of substitutions,
D is the number of deletions,
I is the number of insertions,
C is the number of correct words,
N is the number of words in the reference (N=S+D+C).
How to use
The metric takes two inputs: references (a list of references for each speech input) and predictions (a list of transcriptions to score).
from evaluate import load
wer = load("wer")
wer_score = wer.compute(predictions=predictions, references=references)
Output values
This metric outputs a float representing the word error rate.
print(wer_score)
0.5
This value indicates the average number of errors per reference word.
The lower the value, the better the performance of the ASR system, with a WER of 0 being a perfect score.
Values from popular papers
This metric is highly dependent on the content and quality of the dataset, and therefore users can expect very different values for the same model but on different datasets.
For example, datasets such as LibriSpeech report a WER in the 1.8-3.3 range, whereas ASR models evaluated on Timit report a WER in the 8.3-20.4 range. See the leaderboards for LibriSpeech and Timit for the most recent values.
Examples
Perfect match between prediction and reference:
from evaluate import load
wer = load("wer")
predictions = ["hello world", "good night moon"]
references = ["hello world", "good night moon"]
wer_score = wer.compute(predictions=predictions, references=references)
print(wer_score)
0.0
Partial match between prediction and reference:
from evaluate import load
wer = load("wer")
predictions = ["this is the prediction", "there is an other sample"]
references = ["this is the reference", "there is another one"]
wer_score = wer.compute(predictions=predictions, references=references)
print(wer_score)
0.5
No match between prediction and reference:
from evaluate import load
wer = load("wer")
predictions = ["hello world", "good night moon"]
references = ["hi everyone", "have a great day"]
wer_score = wer.compute(predictions=predictions, references=references)
print(wer_score)
1.0
Limitations and bias
WER is a valuable tool for comparing different systems as well as for evaluating improvements within one system. This kind of measurement, however, provides no details on the nature of translation errors and further work is therefore required to identify the main source(s) of error and to focus any research effort.
Citation
@inproceedings{woodard1982,
author = {Woodard, J.P. and Nelson, J.T.,
year = {1982},
journal = Ẅorkshop on standardisation for speech I/O technology, Naval Air Development Center, Warminster, PA},
title = {An information theoretic measure of speech recognition performance}
}
@inproceedings{morris2004,
author = {Morris, Andrew and Maier, Viktoria and Green, Phil},
year = {2004},
month = {01},
pages = {},
title = {From WER and RIL to MER and WIL: improved evaluation measures for connected speech recognition.}
}