Works with the AI Agents you already use
WhatsApp Voice Note Transcribe
Transcribe and translate voice notes 100% on your own machine with faster-whisper. Built for Arabic, including Egyptian dialect.
Free
WhatsApp Voice Note Transcribe
Example session with this skill installed
Transcribe this WhatsApp voice note 'feedback.ogg'. It's in Arabic. Check if the customer mentions any issues with the new subscription flow or app names.
- Read your context and instructions
- Compiled the whatsapp voice note
I transcribed feedback.ogg (21 seconds) on this machine with faster-whisper large-v3 on your GPU: 4 s to load, 3 s to transcribe. No audio or text left your computer.
Short answer
Subscription issue: yes. The customer says the subscription won't renew from their iPhone, and reopening the app didn't help.
App names: none clearly said. One word at 00:15–00:17 is missing from the transcript and may be an app name. Please listen to that part.
Not said: whether an error message appeared, or when the problem started.
Transcript and translation
| Time | Arabic (as transcribed) | English |
|---|---|---|
| 00:00–00:04 | السلام عليكم، كنت بحاول ادخل على البروفايل بتاعي | Hi, I was trying to get into my profile |
| 00:04–00:11 | بس الصبسكريبشن مش راضي يتجدد خالص من الايفون | But the subscription [?] won't renew at all from the iPhone |
| 00:11–00:17 | جربت افتح الابلكيشن كذا مرة ومفيش فايدة، حتى نزلت | I tried opening the app several times, no use. I even downloaded [?] (a word is missing here) |
| 00:17–00:21 | بس برضو نفس المشكلة | but it's still the same problem. |
Key quote: «الصبسكريبشن مش راضي يتجدد خالص من الايفون», "the subscription won't renew at all from the iPhone".
Reliability
| Time | Transcribed | Likely meaning |
|---|---|---|
| 00:04 | الصبسكريبشن | "subscription", an English word written phonetically |
| 00:15–00:17 | حتى نزلت … | A word is missing after "I even downloaded", possibly an app name. Worth listening to. |
Files
feedback/transcript_ar.txt: the raw transcript, unchangedfeedback/transcript_en.txt: the translation, with[?]marksfeedback/confidence_log.txt: per-segment confidence
Next steps
- Listen to 00:15–00:17 to catch the missing word.
- If it's a bug report, ask the customer for their iOS version and any error message. They didn't mention either.
Connects securely to your tools. The creator never sees your data.
About this skill
Stop sending sensitive audio to cloud APIs. This skill transcribes WhatsApp voice notes and other audio entirely on your own machine with faster-whisper (Whisper large-v3). The audio and transcripts never leave your computer.
It's built for Arabic voice notes, including Egyptian and other dialects with English words mixed in, and works with any language Whisper supports.
What it does
- Local transcription: turns .ogg/.opus, .m4a, .mp3, .wav and anything else ffmpeg can read into a timestamped transcript, with no cloud calls.
- Honest translation: Claude writes an English translation and marks every uncertain word with
[?]instead of guessing. - Answers your questions: give it questions (for example, survey questions) and it quotes the key sentences in the original language with a translation, and says plainly when something wasn't answered.
- Reliability report: lists the least reliable segments and what they probably say. Whisper often drops brand names and English words in dialect speech, so you know what to listen to again.
- GPU or CPU: uses an NVIDIA GPU if you have one (about 17 s for a 2.5-minute note) and falls back to the CPU (about real time).
Privacy
- Offline by default. The model downloads once, and only after you approve it.
- The original audio is never modified. Outputs go in a new folder next to it.
Requirements
- Python 3.10+ and ffmpeg
- About 3 GB of disk for the Whisper large-v3 model (a one-time download)
- An agent that can run shell commands on your machine. Tested with Claude Code.
Source and updates
Open source (MIT): https://github.com/hesham-abboud/whatsapp-voice-note-transcribe
- Dialect transcripts need a human check: Whisper often drops or garbles brand and app names and English words mixed into Arabic. The skill flags these segments, but can't recover words the model never wrote.
- Requires ffmpeg, Python 3.10+ and an agent with shell access. Tested only with Claude Code on Windows 11.
- One-time model download: about 3 GB for large-v3 (or 1.5 GB for medium). Network proxies may block it, in which case the skill stops and reports the blocked host.
- CPU transcription runs at about real time (a 2.5-minute note takes about 2 minutes). An NVIDIA GPU with CUDA 12 is about 6x faster.
- GPU and CPU output can differ slightly, and GPU runs occasionally add a short made-up phrase on the last second of audio. Check the final line.
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 any payment.
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
4 uses
Times used to date
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 1 month ago
- Free with an Agensi account
Requires
Permissions
Allowed Hosts
File Scopes
Audio files are only read, never modified. Outputs (transcript_<lang>.txt, transcript_en.txt, confidence_log.txt) are written to a new folder next to the audio file, named after it. The CLI is installed into ~/.whatsapp-voice-note-transcribe, and the Whisper model is cached in ~/.cache/huggingface. Network use: github.com (installing the CLI, only if the bundled source isn't available) and huggingface.co (one-time model download of about 3 GB). Both need the user's approval first. Transcription itself is fully offline. No audio or transcripts are ever uploaded.