A request for a “video transcript” can mean several different jobs. Choosing the operation first makes the output, failure handling and eventual bill easier to understand.
Retrieving an existing track
Caption retrieval reads an available track. Its language, timestamps and origin belong to that track. The track may have been written by a person or generated automatically. Retrieval does not independently verify every word against the audio.
The youtube-transcript-api project documents language selection and transcript types. The official YouTube caption-download operation has authorization requirements. Public playback alone is not proof that your account can download any video’s captions through that API.
Creating text from audio
Speech recognition processes audio to produce text. It is a separate workload, even when an existing caption track is unavailable. Input access, language support, speaker handling, timestamps and audio quality affect what a useful result means. A caption-access error says nothing about the accuracy of a speech recognition model.
For an evaluation, use audio you are permitted to process and a checked reference transcript. Record the language and normalization rules before calculating word error rate. Without a reference, you can inspect formatting and completeness but should not report a measured accuracy percentage.
Translation and summary are further steps
Translation changes the output language; summarization selects and rewrites information. Neither is a verbatim transcript of the original speech. Store the original language, requested output language and transformation separately. A summary must not silently replace a transcript in a field that promises the spoken words.
Choose an output contract
Before integrating, answer four questions:
- Do you need the available caption track or newly recognized speech?
- Must the language match the audio, or is translation requested?
- Do you need timestamps, speaker labels or plain text?
- What should happen when the requested output is unavailable?
Preserve the distinction between missing captions, denied access, a failed request and a completed result. See the transcript troubleshooting guide for a diagnosis path and the cost guide for differences between media duration and runtime.
TrueFetch’s video tasks remain planned. This explanation does not establish support for a particular language, site or transcription model.