BrassTranscripts Launches AI Transcription Research Index
BrassTranscripts has published the Research Index — an openly accessible, hand-curated collection of 28 primary sources on AI transcription quality, organized into five categories and annotated for the people who actually build and evaluate transcription systems. It is live now at brasstranscripts.com/research, with no account or payment required to read any of it.
The Research Index answers a question that vendor marketing usually skips: what does the actual evidence say about AI transcription accuracy? Instead of a single headline percentage, the index links to the peer-reviewed papers, open benchmarks, and documented tools that underpin every credible WER number — plus two studies BrassTranscripts published from its own data.
Quick Navigation
- What the Research Index Is
- Five Categories, 28 Primary Sources
- A Strict Inclusion Bar
- BrassTranscripts Published Its Own Research, Too
- Who the Research Index Is For
- Frequently Asked Questions
What the Research Index Is
The BrassTranscripts Research Index is the evidence layer beneath the product — every entry is a primary document with a plain-language annotation explaining what builders should do with it. It treats AI transcription accuracy as a measurable, sourced question rather than a marketing slogan, and it links directly to the original papers and benchmarks so any claim can be checked at the source.
Most transcription marketing collapses a complex picture into one number. The index does the opposite: it gives you the papers behind the number, the benchmarks the number came from, and the conditions under which that number holds. You can start at the Research Index pillar page and drill into whichever quality dimension matters for your audio.
Five Categories, 28 Primary Sources
The BrassTranscripts Research Index organizes 28 curated entries across five categories that map to the real quality dimensions of AI transcription: accuracy, speaker identification, language coverage, audio robustness, and benchmark methodology. Each category is its own page with its own annotated entries and a focused FAQ.
The five categories:
- Transcription Accuracy — the foundational speech-recognition architecture, the long-form inference paper, the Open ASR Leaderboard, the Artificial Analysis benchmark, LibriSpeech as the canonical floor, and BrassTranscripts' own accuracy investigation.
- Speaker Diarization — the neural diarization research, overlap-aware methods, and the ETH Zurich multi-model benchmark that shows what speaker identification actually costs in error terms.
- Multilingual Speech — FLEURS, Common Voice, Earnings-22, peer-reviewed accent research, the Indian-English (Svarah) and African-English (AfriSpeech-200) benchmarks, and BrassTranscripts' own 30-language demand study.
- Audio Quality — reverberation benchmarks, signal-to-noise thresholds, the CHiME far-field challenge, and a practical WER measurement library.
- ASR Benchmarks — the Open ASR Leaderboard long-form track, Artificial Analysis, MLPerf Inference v5.1, and the NIST scoring toolkit that underlies virtually every published WER number.
For readers who want the applied version of this evidence, the 99-language AI transcription guide and the complete speaker identification guide translate the research into practical expectations.
A Strict Inclusion Bar
Every entry in the BrassTranscripts Research Index must clear three rules: it must be a primary source, it must be current, and it must support a concrete builder decision — no marketing benchmarks, no competitor product comparisons, and no auto-aggregated link lists. This is why the index stays small and hand-maintained rather than scraping arXiv abstracts into an undifferentiated feed.
The bar is deliberately exclusionary. Vendor white papers don't qualify because they aren't independent. Paywalled papers without a preprint don't qualify because readers can't verify them. Abandoned tools don't qualify because builders shouldn't adopt them. What survives is a short, trustworthy reading list — and each surviving entry carries a hand-written note on what to do with it, not a restatement of the abstract. Competitor comparisons live separately in pages like the AI transcription accuracy benchmark comparison, keeping the Research Index focused on the literature and open-source ecosystem.
BrassTranscripts Published Its Own Research, Too
Two entries in the Research Index are first-party BrassTranscripts data studies, labeled as such so their provenance is unambiguous. The first traces the widely repeated "98% accuracy" claim to a single MLCommons LibriSpeech result and documents a real-world accuracy range of roughly 33% to 97.9%; the second analyzes six months of production data to map where global transcription demand actually concentrates.
The accuracy investigation is the more pointed of the two. Read it in full in AI Transcription Accuracy Claims: An Investigation — it walks through OpenAI's own documentation, four peer-reviewed studies, and independent benchmarks to show that the "98%" figure is a clean-audiobook result extracted from context, while phone-call audio documents accuracy as low as 46–57%. The point isn't that AI transcription is bad; it's that accuracy depends on conditions, which is exactly why BrassTranscripts offers a 30-word preview before you pay.
The second study, Global AI Transcription Trends 2026, is built from 515 paid jobs and 252 hours of audio across 30 languages. It found that non-English jobs make up 37% of volume, Portuguese leads non-English demand, and one small-population language — Norwegian Nynorsk — ranked third by total hours on the strength of long, many-speaker institutional recordings. It's a working demonstration that "99+ languages" reflects real, paid, uneven demand rather than a marketing checkbox.
Who the Research Index Is For
The BrassTranscripts Research Index is built for three audiences: developers choosing or evaluating an ASR system, journalists fact-checking transcription accuracy claims, and researchers who need primary sources without wading through marketing. Each annotation is written for someone who has to make a decision, not someone who wants a summary.
If you're a builder, sort the Open ASR Leaderboard by the Earnings-22 long-form column rather than LibriSpeech before you trust any production accuracy claim. If you're a journalist, the accuracy category gives you citable, condition-specific numbers instead of a single vendor figure. And if you're recording audio you'll need transcribed, the audio quality research and the practical audio quality tips for perfect transcription explain which recording problems actually move WER. Teams turning transcripts into downstream deliverables can pair the index with the AI Prompt Guide and the speaker identification feature page.
Frequently Asked Questions
What is the BrassTranscripts Research Index?
The BrassTranscripts Research Index is an openly accessible, curated index of 28 primary sources on AI transcription quality, organized into five categories: transcription accuracy, speaker diarization, multilingual speech, audio quality, and ASR benchmarks. Each entry links to the original paper, benchmark, or tool and includes a short annotation explaining what builders should do with the finding. It is available at brasstranscripts.com/research.
Do I need an account to read the Research Index?
No. The BrassTranscripts Research Index requires no account, no email, and no payment to read — every category page and every entry is openly accessible at brasstranscripts.com/research. The index exists so builders, journalists, and researchers can verify the evidence behind AI transcription accuracy claims directly.
Does the Research Index include BrassTranscripts' own research?
Yes. Alongside peer-reviewed papers and independent benchmarks, the index includes two first-party BrassTranscripts data studies: an investigation tracing the widely repeated "98% accuracy" claim to its single LibriSpeech origin, and a production-data analysis of language demand across 30 languages. Both are labeled as first-party in their bylines so the provenance is clear.
What sources does the Research Index exclude?
The BrassTranscripts Research Index excludes vendor marketing benchmarks, competitor product comparisons, paywalled papers with no preprint, abandoned tools, and auto-aggregated content. Every entry must be a primary source, must be current, and must support a concrete answer to a builder question — three inclusion rules applied by hand to each entry.
Why did BrassTranscripts build a research index instead of just publishing accuracy numbers?
BrassTranscripts built the Research Index because a single accuracy percentage is misleading without the conditions that produced it — documented real-world accuracy for AI transcription spans roughly 33% to 97.9% depending on audio quality, accent, and language. The index points to the primary evidence so readers can evaluate any accuracy claim, including BrassTranscripts' own, against the source.
About BrassTranscripts
BrassTranscripts is a professional AI transcription service offering automatic speaker identification, support for 99+ languages with automatic detection, and four output formats (TXT, SRT, VTT, JSON) with every transcription. Files are priced per batch with no subscription — $2.50 for files up to 15 minutes and $6.00 for files 16–120 minutes — and every transcript includes a 30-word preview before payment so users can verify quality on their own audio first. Explore the evidence base at the Research Index, or start transcribing in about 30 seconds.