You don't have to listen.
Forward a voice message. Read it a few seconds later.
-
No headphones. Read it in a meeting, on the train, in bed.
-
Stays private. Transcribed on a private server. Audio is never stored.
-
About 3 seconds. For a whole minute of voice.
Try it now
Send the bot a voice message. Free, no sign-up.
tldl.1337 · Telegram: @too_long_didnt_listen_bot · WhatsApp: invite-only
Your voice messages stay yours.
Audio is deleted right away. It's transcribed and then thrown away, along with any copy the messenger bridge made.
No transcripts kept. The bot replies and forgets. It only remembers whom it has already welcomed.
No cloud services. The models run on the server's own CPU. No audio goes to an AI company.
No account needed. The bot only sees what the messenger shows it: an ID, never your contacts.
The messenger itself still delivers the message, like any other chat.
Use the messenger you already have.
Telegram Create a bot with @BotFather and paste the token. No public URL needed.
Signal Register a number for the bot in signal-cli-rest-api. It doesn't need a phone after that.
WhatsApp Through WAHA, linked to a spare phone that opens WhatsApp at least every 14 days. Invite-only: WhatsApp doesn't allow public bots like this.
Fast first, then better.
| Engine | Languages | 1 min | 5 min |
|---|---|---|---|
| NVIDIA Parakeet v3 | 25, detected automatically | 3.5 s | 12.8 s |
| Whisper large-v3-turbo | 99, detected automatically | 13.7 s | 48.8 s |
Measured on an Intel i5-12400 without a GPU. You get Parakeet's transcript right away. Whisper then double-checks it, and the message updates with its version, which is better on colloquial speech.
Running in a few minutes.
git clone https://github.com/then4p/tldl.git && cd tldl cp config.example.yaml config.yaml cp .env.example .env docker compose up -d
Needs Docker. Fill in your bot token and allowed senders, then start it. It uses about 0.8 GB of RAM while a model is loaded and about 0.1 GB when idle.