Transcription

Types of transcription

Type	What it contains	Typical use
Verbatim	Every word, including hesitations, false starts and filler words	Legal records, research interviews
Clean (intelligent) verbatim	The words spoken, without fillers and repetitions	Marketing video, webinars, customer stories
Timed transcript	Text split into short cues with start and end times	Closed captions, subtitles, search inside video
Edited transcript	Rewritten for reading, with headings	Blog posts, case studies made from a recording

A timed transcript can always be turned into a plain one by removing the timings. The reverse requires aligning the text with the audio again.

How automatic transcription works

Automatic speech recognition (ASR) systems run a recording through a short pipeline:

1. Audio extraction. The sound track is separated from the video and converted to a standard format, usually mono at a fixed sample rate.
2. Voice activity detection. Passages with no speech, such as music beds or silence, are identified and skipped.
3. Recognition. A neural model converts the speech into words and detects the language. Current models are trained on many languages at once.
4. Alignment. Each segment of text is given start and end times, which is what turns a transcript into captions.

How accurate is transcription?

Accuracy is usually measured as word error rate (WER): the share of words substituted, deleted or inserted compared with a correct reference transcript. On clear speech recorded with a good microphone, current models make few errors. Accuracy drops with:

• background music and crowd noise at events;
• several people talking over each other on a panel;
• strong accents or switching between languages mid-sentence;
• brand names, product codes, acronyms and internal jargon, which the model has rarely or never seen.

The last point matters most for marketing teams. A transcript that writes a product name three different ways is still searchable by the other words, but it should not be published as captions until someone has corrected it. Systems that accept a list of preferred spellings reduce this kind of error.

Video and audio transcription in DAM work

The main reason to transcribe a media library is search. Without a transcript, a video is found only by its title, tags and description. With one, every spoken word becomes searchable text:

• A product marketer searches for "carbon footprint" and finds the three customer interviews where a client mentions it, not only the video titled "Sustainability".
• A legal reviewer checks every recorded webinar that mentions a discontinued offer before the archive is opened to partners.
• A content team finds a quotable sentence from a CEO keynote and turns it into a social post.
• A podcast archive becomes searchable by topic, guest and product.

Transcripts also produce captions and subtitles for accessibility, give translators a source text, and give the metadata team material for summaries and keywords. WCAG 2.2 success criterion 1.2.1 asks for a text alternative, typically a transcript, for prerecorded audio-only content such as podcasts.

Common mistakes:

• Treating an automatic transcript as an approved quote. Check the recording before quoting a customer.
• Transcribing only new uploads, which leaves the archive unsearchable.
• Translating from an uncorrected automatic transcript, which carries its errors into every language.

Transcription in Freedam

• Video and audio files are transcribed automatically at upload, with no setting per file. The language is detected automatically and stored on the transcript.
• Transcription runs on Freedam's own servers; the audio is not sent to an outside transcription service.
• Videos without speech, or with music only, are marked No speech detected instead of getting an empty transcript.
• On videos, the transcript becomes a caption track. The transcript editor shows the video next to the timed text, highlights the line being spoken, and saves corrections back to the track. Transcripts download as VTT, SRT or plain text. See video transcripts.
• Transcripts of both video and audio are added to the search index, so a phrase spoken in a recording finds it in the gallery search. A transcript translated into another language makes the asset findable in that language. See search.
• Administrators list brand and product names on the Glossary page, and those spellings are given to the transcription model for every new transcript.

Frequently asked questions

What is the difference between a transcript and captions?

A transcript is the text of what is said, read on its own. Captions are the same text split into short timed cues and shown over the video, and they also describe relevant sounds for viewers who cannot hear them.

Can I search a video by the words spoken in it?

Yes, when the video has a transcript and the library indexes it. The transcript is treated like any other text field, so a search for a phrase returns the videos in which it is spoken.

Which languages can automatic transcription handle?

Current speech recognition models are multilingual and detect the language themselves. Accuracy is highest for widely spoken languages and for recordings in a single language.