Transcript Error Rate Reviewer

How to use Transcript Error Rate Reviewer

Review how two transcript excerpts differ without uploading their contents or running speech recognition. Choose the token and case rules, then inspect matches, substitutions, deletions and insertions against the exact reference denominator. The complete evidence connects every comparison token back to its original text position, including repeated words and omitted whitespace.

1. Paste the reference and hypothesis excerpts. They remain in the downloaded evidence, so use only content you intend to include in that report.
2. Choose word tokens, scalar characters or both. Review the case and CER whitespace policies before running; punctuation stays literal and no spelling is repaired.
3. Read the S/D/I counts and reference denominator, then inspect the ordered alignment and source spans. Download JSON or CSV for a complete review rather than treating a percentage alone as transcript quality.

Good to know

• A supplied-text edit review only: no audio, ASR, translation, semantic equivalence, truth grading or quality certification. It does not reproduce NIST/SCTK/jiwer scoring conventions or benchmark preprocessing. A reliable reference and suitable language-specific tokenization remain your responsibility.
• Words are maximal runs outside U+0009–000D, U+0020, U+00A0, U+1680, U+2000–200A, U+2028–2029, U+202F, U+205F, U+3000 and U+FEFF. Punctuation and U+0085/U+200B stay inside words. A language without whitespace word boundaries may need a different tokenizer. CER counts Unicode scalars, not visible graphemes: accents and joined emoji can contain several tokens.
• Exact case or ASCII A–Z lowercase only. There is no NFC normalization, accent removal, punctuation stripping, full Unicode case folding, word deduplication or silent sorting. Word comparison views join tokens with one space; CER views join retained scalars. Original strings, token source spans and omitted separator spans remain available to reconstruct the supplied text.
• Global insertion/deletion/substitution cost 1, exact-match cost 0. Backtrace from the final cell prefers an optimal diagonal, then deletion, then insertion. Only one optimum is emitted; tied alternatives can have different S/D/I decompositions. WER/CER is errors divided by reference tokens, so insertions can make it exceed 100%. An empty reference gives null rate, including when both sides are empty; counts are still reported.
• At most 10,000 UTF-16 units per input, 16,000 units and 64,000 UTF-8 bytes jointly, 2,000 tokens per side per metric, 1,200,000 DP cells jointly and a 2-second synchronous compute guard. Both complete JSON and CSV must each fit 2,000,000 UTF-16 units and UTF-8 bytes before either returns. Oversized, malformed-Unicode or over-budget runs return no partial evidence. Shorten the excerpt or choose one metric.
• Coordinates are zero-based, with excluded ends. Source spans include both UTF-16 and scalar positions; comparison-view spans use UTF-16 and alignment indexes address tokens. A textarea may normalize line endings before this tool receives its value. This is not a byte-preserving file import.
• JSON includes the original transcript text; it is not anonymous. CSV quotes all cells and adds one apostrophe to every nonempty source/comparison text cell, including normalized views, to reduce spreadsheet formula interpretation. Remove exactly that declared prefix only when deliberately recovering text. No spreadsheet importer compatibility is guaranteed. Inputs are processed locally without storage or network requests.