Video transcription
Video transcription is the process of turning the speech in a video into written text, with timestamps that show when each word or line was said. It can be done by AI speech-to-text, by human transcribers, or by AI followed by a human review pass. A finished transcript has punctuation, timestamps and often speaker labels. It can be read as a document, searched, or cut into caption files such as WebVTT and SRT. For a lecture or a news interview, the transcript is often a second product that sits under the video.
How does video transcription work?
A video file is a container, such as MP4 or a set of HLS segments, holding separate video and audio tracks. Transcription only needs the audio, so the first step is to demux it, which means pulling the audio track out of the container, then decode it and resample it to the format the speech model expects.
The speech-to-text model converts the audio into words and assigns each word a start and end time. Speaker diarization groups the voices, so the transcript can say Speaker 1 and Speaker 2, or real names if you map them. A formatting pass, called inverse text normalization, adds punctuation and writes spoken numbers and dates as digits.
The output is usually a structured file, often JSON, with every word and its timing. From that one file you can render a plain text document, a caption track, a search index or the input to a summary. Keeping the structured version, not just the text, is what lets later features point back to an exact second.
Video transcription vs. captions
A transcript is the complete text of what was said, laid out for reading. It has paragraphs, speaker labels and no limit on line length. Captions are the same words cut into short timed lines, usually one or two lines on screen at a time, kept on screen long enough to read and placed so they do not cover faces or on-screen text.
Captions also follow rules a transcript does not need. Each cue has a start and end time measured against the video timeline, lines break at natural phrase boundaries, and sounds that matter to meaning are described in brackets.
There is also a choice of style. A verbatim transcript keeps every um, false start and repetition, which matters for legal records and interview review. A clean read drops them, which suits study notes and articles. Captions usually sit close to verbatim so the text matches what viewers hear.
What affects video transcription accuracy?
Most transcription errors start in the audio, not the model. A lavalier or headset microphone gives far cleaner text than a laptop mic across a room. Music beds under speech, crowd noise at a match and echo in a lecture hall all compete with the voice. When two people talk over each other, both sentences suffer.
Encoding choices matter too. Audio compressed to a very low bitrate, or recorded on a phone over a weak connection and then re-encoded, loses the detail that separates similar sounds. If you control capture, record each speaker on a separate channel and keep the source audio at a sensible quality.
Vocabulary is the other big factor. Names, product terms, acronyms and code words are the errors people notice. Interview and proctoring platforms care most about speaker labels; course platforms care about subject terms. Spot-check those words on real recordings before you trust a transcript at scale.
AI vs. human video transcription
AI transcription returns text in minutes, costs a small amount per minute of audio and handles a whole library without scheduling anyone. On clear speech in a common language it reads well. Its weak spots are heavy accents, overlapping speakers, poor audio and rare words, where it can produce confident text that is wrong.
Human transcription is slower and priced much higher per minute, but a trained transcriber uses context to resolve unclear words, marks inaudible sections honestly and follows a style guide. It is still used for court records, published interviews and other text that people will quote.
Most teams now run a hybrid. AI transcribes everything as it is uploaded, and a person reviews only the recordings that will be published or relied on. The review is faster when the tool shows the transcript next to the video and lets an editor fix a word without re-timing the whole file.
How does FastPix handle video transcription?
FastPix transcribes video inside in-video AI, in parallel with encoding. The same upload, whether it comes through the resumable web, iOS, Android or Flutter SDKs or is pulled from S3, GCS, Azure Blob or an HTTPS URL, starts both jobs, so the transcript is ready alongside the playable video rather than after a second pipeline.
The result is one timestamped JSON per video. It carries the transcript with word-level timestamps and speaker diarization, alongside chapters, people and brands, scenes and on-screen text. Transcription covers 50+ languages. FastPix turns the transcript into captions and subtitles in WebVTT and SRT, with multi-language output and editable cue points, and offers a searchable transcript export. Webhooks fire on upload, encode and ready, so your backend knows when to fetch the results.
For recordings that should not be public, such as interviews or recorded reviews, the video can sit behind a signed playback ID with an expiring token, and you decide where the exported transcript is kept. Where data-location rules apply to the video itself, FastPix can set the storage region on request; ask which regions are available. FastPix does not offer human transcription or a human review service.
Transcription is billed per minute of video processed, on the same bill as encoding, with no token billing. Encoding the video on Standard is free up to 4K. To test, run a handful of your own lectures or interviews through the free plan's subtitle and AI sample minutes; no card is needed.
By the numbers
- About a third of the 680,000 hours of audio used to train OpenAI's Whisper model is in languages other than English. (Introducing Whisper (OpenAI))
