- Home
- Skills
- Productivity
- Whisper Transcription Workflows
More screenshots
Works with the AI Agents you already use
Whisper Transcription Workflows
Transcribe audio with Whisper the way that actually works: right model size, faster-whisper recipe, subtitles, long-audio chunking, and a QC pass that catches errors.
$5
Whisper Transcription Workflows
Example session with this skill installed
Transcribe a 90-minute two-person interview recorded in a café, and give me subtitles for the video version.
- Read your context and instructions
- Compiled the whisper transcription workflows
Here is your setup and what you will get.
Model and settings: café noise plus two voices means go one size up — medium on CPU, turbo if you have a GPU. Set the language explicitly (skip auto-detect; it can commit to the wrong language over a noisy intro), beam_size 10 for the noise, vad_filter on so the quiet stretches between answers do not get filled with invented speech. Seed initial_prompt with both speakers' names spelled correctly plus the interview topic — that is your biggest accuracy lever for names.
Chunking: 90 minutes is past the single-pass comfort zone. Split into six 15-minute chunks with 30 seconds of overlap, transcribe each separately, then stitch and trim the one duplicated boundary sentence per join during QC.
Speaker labels: Whisper hears a single voice stream, so it cannot say who said what. If you need Interviewer/Guest labels, add a diarization pass (pyannote) and merge by timestamp. Flagging that now so it is not a surprise at delivery.
Deliverables: interview.txt (the full transcript), interview.srt for the video, and a short QC log. QC on this file means: search every instance of both names and any numbers against the audio, delete fluent sentences that appear where the café went quiet, and watch the first two minutes of video with the SRT loaded to check caption timing.
Expected accuracy: names and numbers still get a correction pass — they are the top misheard category — but with this setup a three-minute spot-check should land in the 3–6 errors-per-minute band: fixable in QC, no re-run needed.
Connects securely to your tools. The creator never sees your data.
About this skill
Whisper is easy to run and easy to run badly — wrong model size, hallucinated sentences over silence, guest names spelled three ways. This skill is the full workflow around the model: a model-selection decision table (draft vs. CPU-only vs. hard audio), a faster-whisper recipe that runs several times quicker than the reference package, the three settings that matter more than model size (explicit language, a seeded initial prompt, VAD), SRT/VTT subtitle generation, word-level timestamps for karaoke and lyric sync, a chunking protocol that keeps long recordings from degrading, a hard-audio playbook for music beds, café noise, crosstalk, and quiet speakers, and a mechanical QC pass with an errors-per-minute acceptance rubric that tells you whether to ship, fix, or re-run at a larger model.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Productivity.
- Batch processing only — no real-time or live captioning.
- No built-in speaker labels; diarization requires a separate tool (the skill explains the merge).
- Word timestamps drift on sung vocals and very noisy audio; the QC pass is a required step, not an optional extra.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 24 days ago