Podcast transcriber to SRT that run on my Iphone
| V1 gave 70% accuracy and had been improved to 90% on V3 |
I used to need my MacBook for every podcast episode. I recorded on my iPhone, moved the file to the Mac, ran Whisper medium in Terminal to get an SRT, pasted it into Claude to fix the Singlish, then copied the result to my blog. It worked, but I had to carry the Mac everywhere.
So I used Claude Pro to help me build a small web app. Now the whole job happens on my phone.
What it does
- Records my voice on the iPhone and also keeps a full-quality .m4a for the podcast upload.
- Sends the audio to a hosted Whisper large-v3 model and gets back an SRT in about a minute.
- Cleans up common mistakes automatically, then copies the SRT plus a ready-made prompt, so I can paste it straight into Claude.
- Costs about US$0.06 for a 30-minute episode, or nothing within the free tier.
How I built it
- Asked Claude to propose three options (web app, phone app, or a simple bookmark), and picked the web app.
- Signed up for a free Groq account to get an API key for Whisper.
- Put the app on Cloudflare Workers, with the key stored as a secret and a passcode so only I can use it.
- Installed it on my iPhone Home Screen from Safari.
- Tested it with a real recording and fixed what went wrong.
What went wrong, and how I fixed it
My first test was only about 70% accurate. The hint I gave Whisper was echoed back into the text during silence. Forcing Chinese turned my English sentences into Chinese. Names like Claude and Groq were misheard as "Claw" and "Grodd". I changed the hint, set the language to auto-detect, added a filter for text over silence, and added a list of automatic corrections.
It is not perfect, and I still let Claude fix the Singlish and sentence structure. But I no longer need the MacBook to turn a recording into a blog post.
How the pieces fit
iPhone Safari app, then a Cloudflare Worker (holds the secrets), then Groq (runs Whisper large-v3), then the SRT comes back to the phone, then I paste it into Claude.
Tools and links
Tool | What I use it for | Link |
|---|---|---|
Groq console | Whisper large-v3 hosting, API keys, free tier | |
Groq Whisper model page | Model details and limits | |
Cloudflare dashboard | The Worker, secrets, the srt.garynbc.com domain | |
Cloudflare Workers limits | Free plan limits (100,000 requests a day) | |
Wrangler docs | The Terminal tool that deploys the app | |
Node.js | Needed to run Wrangler on the Mac | |
OpenAI Whisper | The speech model behind it, also used locally on my Mac | |
Claude | Building the app and fixing the Singlish | |
QuickTime Player (Mac) | Exporting audio from a video: File, Export As, Audio Only | Built into macOS |
YouTube Studio | Downloading my own channel's videos |
Comments
Post a Comment