The author

TranscribeBench is written and maintained by one person, a Principal ML engineer, who publishes under the site's name.

Why this site

I have spent more than ten years building machine-learning systems, much of it on video and audio streams: I created Videoflow, an open-source framework for video stream processing, and I was the lead ML platform engineer for a computer-vision platform that ran 5,000 streams. Measuring a model on data it has never seen, with a metric written down before the results come in, is the habit that work taught me. TranscribeBench applies it to a question people ask before they pay for transcription: how accurate is it on audio like mine, and what will it cost?

Most "best transcription software" lists answer that with a vendor's own accuracy claim. TranscribeBench measures instead: openly licensed recordings with human transcripts, the word error rate computed by code you can run yourself (the WER calculator is the same code), and the data published. Where something has not been measured, the site says so.

How the site is funded

The tools and the benchmark data are free. The site may earn referral or affiliate commissions from some vendors. Every such link is labeled "(paid link)" where it appears. Today none are active, so every vendor link is a plain link. The affiliate disclosure lists every relationship and the rules that keep commissions out of the results. There is no advertising and no sponsored content.

Independence

TranscribeBench is not affiliated with any transcription, captioning or podcast vendor, with OpenAI, or with the universities that publish the speech corpora it uses. Vendors cannot pay for placement or ranking, or to be left out of a test. The site publishes no reviews or testimonials.

Contact

Email [email protected] to report a wrong price, a scoring mistake, a clip whose reference transcript is wrong, or a bug in a tool. Corrections are dated on the page they affect.

Also available as Markdown.