Word Error Rate (WER)

The most common measure of speech recognition accuracy is known as word error rate or WER.

Calculating WER

Human vs. machine

To detect an error in transcriptions produced by a machine, we need the expected, correct, or gold standard transcription so a comparison can be made.

Note that WER calculation assumes a human transcription is 100% accurate, even though humans have about a 5% WER themselves.

See the example below. The example is completely made up. The results from the API are a fabrication.