File size: 6,228 Bytes
e8b2823
f21e916
0a485f3
f21e916
 
e8b2823
 
 
 
f21e916
0a485f3
 
8dfbd14
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ff842da
0a485f3
f21e916
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
05297cf
f21e916
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
---
title: WER
emoji: 🤗 
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 3.0.2
app_file: app.py
pinned: false
tags:
- evaluate
- metric
description: >-
  Word error rate (WER) is a common metric of the performance of an automatic speech recognition system.
  
  The general difficulty of measuring performance 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 instead of the phoneme level. The 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.
  
  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.
  
  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).
  
  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.
---

# 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](https://en.wikipedia.org/wiki/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](https://huggingface.co/metrics/perplexity) and word error rate (see [this article](https://www.cs.cmu.edu/~roni/papers/eval-metrics-bntuw-9802.pdf) 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).


```python
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](https://huggingface.co/datasets/librispeech_asr) report a WER in the 1.8-3.3 range, whereas ASR models evaluated on [Timit](https://huggingface.co/datasets/timit_asr) report a WER in the 8.3-20.4 range. 
See the leaderboards for [LibriSpeech](https://paperswithcode.com/sota/speech-recognition-on-librispeech-test-clean) and [Timit](https://paperswithcode.com/sota/speech-recognition-on-timit) for the most recent values.

## Examples 

Perfect match between prediction and reference:

```python
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:

```python
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:

```python
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

```bibtex
@inproceedings{woodard1982,
author = {Woodard, J.P. and Nelson, J.T.,
year = {1982},
journal = {Workshop on standardisation for speech I/O technology, Naval Air Development Center, Warminster, PA},
title = {An information theoretic measure of speech recognition performance}
}
```

```bibtex
@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.}
}
```

## Further References 

- [Word Error Rate -- Wikipedia](https://en.wikipedia.org/wiki/Word_error_rate)
- [Hugging Face Tasks -- Automatic Speech Recognition](https://huggingface.co/tasks/automatic-speech-recognition)