Speech Recognition Metrics: Word Error Rate, CER & Real-Time Factor

What you'll build

What you will be able to do

Each lab ends with a concrete result that you build yourself and the checker confirms, cell by cell.

• Turn alignment counts for eight test utterances into errors, WER and hypothesis length, pool them into a corpus WER and count the utterances above a quality threshold.
• Label every aligned word pair as correct, substitution, deletion or insertion with a nested IF, flag exact matches with EXACT, count each error type and compare a normalised WER with a strict, case-sensitive one.
• Fill a word-level Levenshtein table with one copied MIN formula, read the minimum number of edits from its corner and turn it into the WER of a voice command.
• Split a reference and a hypothesis into single characters with MID, flag every differing character with EXACT, calculate the character error rate and compare it with the word error rate of the same pair.
• Break the WER down by accent and by noise condition with SUMIFS, compare the pooled WER with the plain average of recording WERs, measure the noise penalty and chart the accent gap.
• Calculate the real-time factor of eight batch jobs and the pooled RTF, read P50 to P99 streaming latencies with PERCENTILE.INC, check the share within a latency target and estimate how many live streams one GPU can carry.
• Measure how a custom-vocabulary boost changes the error rate of eight domain terms, express the gain as a relative reduction, and check what it does to the WER of the whole test set once new false hits are counted.
• Turn RTF and GPU prices into a cost per audio hour and a monthly bill, find the cheapest model that meets a WER target with MINIFS and INDEX/MATCH, and chart the cost of the five candidates.