Speech Recognition Accuracy Calculator

Calculator Inputs

Enter your corpus totals or batch evaluation counts. The form stays single-column overall, while inputs adapt to screen size.

Reference Words

Total words in the reference transcript.

Substitutions

Words recognized as the wrong words.

Deletions

Reference words missed entirely.

Insertions

Extra words added by the recognizer.

Reference Characters

Useful for fine-grained text quality review.

Character Errors

Character insertions, deletions, and substitutions.

Total Sentences

Number of evaluated sentences.

Correct Sentences

Sentences transcribed with no errors.

Audio Duration (minutes)

Reference audio length in minutes.

Transcription Time (minutes)

System runtime for the same batch.

Average Confidence (%)

Average confidence score reported by your model.

Reset

Example Data Table

This example shows how a speech recognition batch could be evaluated.

Reference Words	Substitutions	Deletions	Insertions	Reference Characters	Character Errors	Total Sentences	Correct Sentences	Audio Min	Runtime Min	Confidence	WER	Word Accuracy
100	4	3	2	550	18	20	17	10	7	92%	9.00%	91.00%
250	11	8	7	1400	46	40	31	25	18	89%	10.40%	89.60%

Formula Used

This calculator combines common speech recognition evaluation measures. It works well for corpus-level benchmarking, model comparison, and deployment readiness checks.

Word Error Rate (WER) = ((Substitutions + Deletions + Insertions) ÷ Reference Words) × 100
Word Accuracy = ((Reference Words - Substitutions - Deletions - Insertions) ÷ Reference Words) × 100
Correct Words = Reference Words - Substitutions - Deletions
Recognized Words = Reference Words - Deletions + Insertions
Precision = (Correct Words ÷ Recognized Words) × 100
Recall = (Correct Words ÷ Reference Words) × 100
F1 Score = 2 × Precision × Recall ÷ (Precision + Recall)
Character Error Rate (CER) = (Character Errors ÷ Reference Characters) × 100
Sentence Accuracy = (Correct Sentences ÷ Total Sentences) × 100
Real-Time Factor (RTF) = Transcription Time ÷ Audio Duration
Confidence-Adjusted Accuracy = max(Word Accuracy, 0) × (Average Confidence ÷ 100)

How to Use This Calculator

Count the total reference words in your ground-truth transcript.
Enter substitutions, deletions, and insertions from your alignment output.
Add character totals if you also want CER tracking.
Enter sentence totals to assess whole-sentence correctness.
Provide audio duration and transcription runtime for speed analysis.
Enter average confidence to compare raw accuracy with model certainty.
Click Calculate Accuracy to show results above the form.
Use the chart and downloads for reporting, audits, or model reviews.

FAQs

1) What does WER tell me?

WER shows how many word-level mistakes exist relative to the reference transcript. Lower values mean cleaner recognition and usually better user experience.

2) Why use CER if I already have WER?

CER is useful when spelling matters, such as names, codes, captions, and multilingual text. It catches smaller mistakes that word-level metrics can hide.

3) Is higher sentence accuracy always better?

Yes, but it is stricter than WER. A sentence with one small mistake still counts as incorrect, so sentence accuracy can drop faster than word accuracy.

4) What is a good real-time factor?

An RTF below 1.00 means the system processes audio faster than playback length. That is often desirable for live or near-live applications.

5) Why can word accuracy be lower than confidence?

Confidence reflects model certainty, not truth. A system can be very confident and still wrong, which is why both measures should be reviewed together.

6) Can this calculator compare two ASR systems?

Yes. Run the same evaluation batch for each system and compare WER, CER, sentence accuracy, speed, and confidence-adjusted accuracy side by side.

7) Should insertions matter as much as deletions?

They matter differently by use case. Deletions remove content, while insertions add false content. WER includes both because each can harm transcript quality.

8) Is the composite score a standard industry metric?

No. It is a practical blended score for dashboarding. Use it for internal comparison, while still reporting standard metrics like WER and CER.