How to make captions, a transcript and chapters for a podcast or YouTube video

Transcribe once

Export your episode as an audio file (MP3 or M4A) or video file. Upload it using the helper from the video editing guide and ask for JSON output:

{
  "audio_url": "https://example.com/interview.mp3",
  "model": "oracle",
  "language": "auto",
  "word_timestamps": true,
  "output_format": "json"
}

If you're transcribing an interview or a panel, set include_diarization: true so you can see who is speaking.

Captions as an SRT file

Turn the words into cues and upload the .srt to YouTube Studio or your podcast host. See the SRT guide for the code and the API's own SRT output, which puts one word per cue.

Translated captions

To get translated captions, you can either run the first transcript through your LLM, or send a second job on the same file using an Oracle model with translate_to and output_format: "srt". Running a separate job for each language keeps the translated subtitles lined up with the audio.

A transcript for the show notes

If you just want plain text, pass output_format: "txt", or stitch the text fields from your JSON segments together with the speaker labels. Give it a read-through before you publish, since automated transcripts often slip up on people's names, product names, and numbers.

Chapters with timestamps

Feed the numbered segments to your LLM and ask it to pick the segment numbers where new chapters start. Don't ask the LLM to write the timestamps directly. Instead, look up the start times for those segment numbers in your code. You can use this prompt with any model:

Below is a podcast transcript, one line per segment: S<number> [start-end] text.
Pick 5 to 9 chapter starts. For each give the segment number and a title of at most six words.
Return JSON: [{"segment": 0, "title": "Intro"}]. The first chapter must be segment 0.

chapters.py

# chapters.py — YouTube chapter lines from (seconds, title) pairs
def stamp(seconds):
    m, s = divmod(int(seconds), 60)
    h, m = divmod(m, 60)
    return f"{h}:{m:02}:{s:02}" if h else f"{m}:{s:02}"


def chapter_lines(chapters):
    return "\n".join(f"{stamp(start)} {title}" for start, title in chapters)


print(chapter_lines([(0, "Intro"), (95, "Why captions matter"), (412, "Editing from a transcript")]))

If you paste a list of timestamps into your video description, YouTube automatically turns them into chapters. For it to work, the first timestamp has to be 0:00, you need at least three timestamps in order, and each chapter must run for at least ten seconds.

Pricing

Billed by the second with no minimum. Full rates: /llms/pricing.md.

Questions and answers