Transcribe audio to subtitles with Claude Code

1. Have the audio ready as a file

   Transcription takes an mp3, a wav, or the soundtrack of a video. Our input was the MiniMax voiceover from the text to speech comparison, saved in the project's outputs/ folder. If what you have is a video, say so — the agent pulls the audio track out locally first. If it is a TikTok, YouTube, or Instagram link, the download service fetches the clip for $0.01 per run before transcription starts.

2. Ask for subtitles, not just text

   Say what the transcript is for. "Transcribe outputs/sneaker-voiceover.mp3 with timestamps so I can use it as subtitles" tells the agent to keep the segment timings and write an .srt file; "give me the text" gets you a paragraph. The service returns both, so the choice costs nothing extra.

3. Pick segment or word timestamps

   Segment timestamps give you subtitle lines — a phrase at a time, which is what a video editor expects. Word timestamps give the start and end of every word, for word-by-word captions or for cutting a clip on a specific word. Both are the same service at the same price; the agent asks which one you want and runs it.

4. Proofread the proper nouns

   Transcription is accurate on ordinary words and guesses on names. On the demo file it heard every word of the script correctly and only changed one piece of punctuation. On an earlier run it wrote "ActionWay" for "Actionway". Read the product names, people's names, and numbers before the subtitles ship.

5. Drop the .srt next to the video

   The agent writes the file into outputs/ with a name matching the audio. Most video editors, and every major video platform, accept an SRT file as a subtitle track directly. If your editor wants VTT instead, ask — converting timed text between formats is ordinary local work the agent does for free.

The run, end to end

1
00:00:00,192 --> 00:00:04,080
Meet the everyday white sneaker: a clean leather upper,

2
00:00:04,200 --> 00:00:06,016
a gum sole that grips wet pavement,

3
00:00:06,320 --> 00:00:09,940
and a fit that works from the morning commute to the late train home.

4
00:00:10,240 --> 00:00:11,300
No logos,

5
00:00:11,340 --> 00:00:11,936
no noise,

6
00:00:12,320 --> 00:00:15,120
just a shoe that goes with everything you already own.

7
00:00:15,320 --> 00:00:15,900
Order today,

8
00:00:15,900 --> 00:00:19,104
and your first pair ships in three days with free returns.

That is a real session, unedited. Every word of the script came back; the one change is a colon after "sneaker" where the script had a full stop. Segments 4, 5, and 7 are under a second long — the model breaks lines at the pauses in the read — so if your editor prefers longer lines, ask the agent to merge segments shorter than two seconds, which it does locally at no cost. The run went through Actionway's media understanding capability, and the Skill installed with the CLI is what gave Claude Code a service that can listen.

When the transcript comes back wrong

Most errors are one of three kinds, and each has a cheap fix. A brand name spelled phonetically: correct it once and ask the agent to apply the same correction across the file. A line that breaks in the wrong place: ask for segments shorter than two seconds to be merged with their neighbours. A file that returns nothing: the audio is usually too quiet or has music over the voice, and the noise-reduction option on the call, or a louder export from your editor, fixes it on the second run — one more $0.10 per run, not a new workflow.

What else the same service does

The price is $0.10 per run whatever the input. Options on the call include the language, a filter that skips silence, a noise-reduction pass, and a cap on segment length so subtitle lines stay short. Timed text chains naturally with the rest of the catalog: transcribe an hour-long recording, find the three best minutes, cut them — the podcast clips use case is that loop — or transcribe a competitor's ad after the download service fetches it. There is no subscription; one balance covers every capability in the catalog.

Set it up once

npm install --global @actionway/cli@latest
actionway init
mkdir -p "$HOME/.claude/skills"
cp -R "$HOME/.actionway/skill/actionway" "$HOME/.claude/skills/actionway"

You need Node.js 20 or newer and an Actionway account; after sign-in the agent tells you your wallet balance and, if it is empty, links you straight to top-up. Prefer a plugin or a Remote MCP server to the command line? Both end at the same catalog and the same wallet — the get started page lists the steps per client side by side.