Representative speech
Document speaker regions, recording conditions, content domains, code-switching, and consent rather than presenting one accent as all Indian English.
Back to tools
Research data pending
Indian English Subtitle Accuracy Benchmark
This is a methodology shell, not a published benchmark. No accuracy scores, rankings, or comparative claims are available yet.
Version 0.1Prepared 19 August 2026Author and editor: CapsAI research team
Proposed scope
What a responsible study must define
Transparent ground truth
Use trained human transcription and an adjudication process for names, numbers, punctuation, and ambiguous speech.
Multiple useful measures
Report word error rate alongside proper-noun, numeric, segmentation, and subtitle-readability observations.
Reproducible reporting
Publish sample definitions, exclusions, tool versions, evaluation scripts, uncertainty, and limitations with the results.
Dataset design
Accent, language, and recording categories
The final study must publish its sampling rationale and cannot claim to represent all Indian English from a narrow speaker set.
Regional coverage
Multiple regions and documented speaker backgrounds
Code-switching
English-only and English mixed with Indian languages
Content domains
Creator, education, business, interview, and instructional speech
Audio conditions
Studio, phone, room noise, music, compression, and overlapping speech
Test methodology
Proposed evaluation sequence
- 1Create consented, documented evaluation audio that is not part of system tuning.
- 2Produce human reference transcripts with independent review and adjudication.
- 3Run every compared system with disclosed settings and version information.
- 4Normalize only pre-declared formatting differences before scoring.
- 5Publish aggregate and category-level observations with uncertainty and limitations.
Scoring framework
More than one headline metric
Word error rate
Standard insertion, deletion, and substitution measurement after documented normalization.
Proper-noun accuracy
Separate review of names, brands, products, and places.
Number accuracy
Dates, currency, percentages, measurements, and other meaning-sensitive figures.
Subtitle segmentation
Readable cue boundaries, line breaks, duration, overlap, and reading speed.
Code-switch handling
Errors around transitions between Indian English and other languages.
Sample output component
How evaluated examples will be disclosed
Reference transcript
Pending human-adjudicated text
System output
Pending evaluated system text
Error annotation
Pending substitution, deletion, insertion, name, number, and segmentation labels
| System | Version | Overall WER | Names | Numbers | Status |
|---|---|---|---|---|---|
| CapsAI | Pending | - | - | - | Data pending |
| Comparison system A | Pending | - | - | - | Data pending |
| Comparison system B | Pending | - | - | - | Data pending |
Methodology disclosure
The eventual report must disclose dataset construction, consent, exclusions, human-reference process, normalization, scoring software, system versions, conflicts of interest, uncertainty, limitations, and any data unavailable for independent inspection. CapsAI evaluating its own system is a conflict that must remain visible.
Reference standards
Publication gate
This page will remain noindex and outside the XML sitemap until a completed study includes evidence, methodology, limitations, review, and results that support every public claim.
Start with 3 free minutes
Test CapsAI with Your Own Video
Upload your own Indian English video and try CapsAI's subtitle workflow. Start with 3 free minutes.