Speech Recognition Metrics: Word Error Rate, CER & Real-Time Factor
Score a speech-to-text engine hands-on: WER from alignment counts, a word-by-word comparison table, the Levenshtein edit-distance table, character error rate with MID, WER by accent and noise, real-time factor and latency percentiles, custom-vocabulary impact and cost per audio hour, in 8 graded labs.
The first lesson is free — no account needed.
About this course
A speech recognition metrics course that scores one fictional start-up's meeting-transcription engine the way evaluation teams do, in a spreadsheet, for ML and speech engineers, product managers buying transcription and data science students. You turn substitution, deletion and insertion counts into word error rate and pool a corpus WER, label a word-by-word comparison with a nested IF and EXACT and compare a normalised with a strict WER, fill a Levenshtein edit-distance table with one copied MIN formula, split strings with MID for the character error rate, break WER down by accent and noise and see why the pooled WER differs from an average, calculate the real-time factor and P50 to P99 latency, measure what a custom vocabulary really gains and pick the cheapest model per audio hour that meets a WER target. Synthetic data. The first lab is free.
Lessons
- Word Error Rate: (S + D + I) / N from Alignment CountsFree
- Word-by-Word Comparison: Labelling C, S, D and I Pro
- Edit Distance by Dynamic Programming: the Levenshtein Table Pro
- Character Error Rate: Splitting Strings with MID Pro
- WER by Accent and Noise: SUMIFS, Pooled vs Average Pro
- Real-Time Factor & Latency Percentiles Pro
- Custom Vocabulary: Measuring the Before/After Impact Pro
- Model Comparison: Cost per Audio Hour vs Accuracy Pro