Step by Step
A
ASR pipeline — from audio signal to text
Automatic Speech Recognition converts an audio signal into features (like a mel spectrogram), processes them through a model, and produces a final text transcript.
Example: converting a raw audio recording into a mel spectrogram representation, then feeding that into a model that outputs the corresponding text transcript.
S
Substitutions, Deletions, Insertions — the three WER error types
Word Error Rate is calculated as the sum of Substitutions (wrong word), Deletions (missing word), and Insertions (extra word), divided by the total number of words in the reference transcript.
Example: if a transcript substitutes one word, deletes another, and inserts an extra word, all three of these errors get counted in the WER numerator.
R
Recent breakthrough — Whisper's near-human accuracy
OpenAI's Whisper (2022) was trained on 680,000 hours of multilingual audio, achieving near-human accuracy on English, while being open-source and free to run locally.
Example: Whisper being used to accurately transcribe English audio recordings with accuracy approaching that of a human transcriber, all while being freely available to run without cost.
Applied Walkthrough
1
A speech recognition system transcribes an audio clip, but its output contains one word substituted incorrectly, one word missing entirely, and one extra word inserted that wasn't actually spoken.
2
Ask: how does WER account for these three different kinds of errors? All three (substitution, deletion, insertion) get summed together in the WER numerator.
3
Dividing this sum by the total number of words in the correct reference transcript produces the final WER score for this transcription.
4
A lower WER indicates better transcription accuracy — this is the standard metric used across the field to evaluate and compare different speech recognition systems, including comparing against benchmarks like Whisper's near-human English accuracy.
Exam Application
Exams test whether you can correctly compute WER given a described set of substitution, deletion, and insertion errors, and whether you know Whisper as a landmark 2022 open-source ASR model achieving near-human accuracy on English.
⚠ Common Trap
The most common trap is forgetting that WER counts three distinct error types (substitutions, deletions, AND insertions), not just one. All three must be summed together and divided by the total word count in the reference transcript to correctly calculate WER.
✓ Quick Self-Check
1. What does ASR stand for?
Automatic Speech Recognition.
Tap to reveal / hide
2. What three types of errors does WER account for?
Substitutions (wrong word), Deletions (missing word), and Insertions (extra word).
Tap to reveal / hide
3. What is the formula for WER?
(Substitutions + Deletions + Insertions) divided by the Total number of words in the reference transcript.
Tap to reveal / hide
4. Is a lower or higher WER better?
Lower is better — it indicates fewer transcription errors.
Tap to reveal / hide
5. What notable achievement did OpenAI's Whisper (2022) accomplish?
Near-human accuracy on English speech recognition, trained on 680,000 hours of multilingual audio, and released as open-source.
Tap to reveal / hide