Turingram

1. What it is

Every meeting transcription service that works well requires a bot account, a cloud upload, and a subscription. Turingram takes a different position.

It runs as a desktop app on your machine. You hit record before the call starts. Turingram captures your microphone and your system audio on separate channels. Whisper transcribes them locally. Speaker diarization runs automatically. When the call ends, you export and do whatever you want with the result.

Nothing joins the call. Nothing leaves the machine. $19, one time.

2. How capture works

Turingram records two tracks simultaneously: your microphone and your system audio output. Keeping them separate means the transcription can assign your voice to one speaker and the far end to another before diarization even runs.

On Linux, PipeWire echo cancellation keeps voices clean. On macOS and Windows, the app uses the platform's native audio routing. No virtual cable required, no extra driver setup.

You start recording before the call, stop when it ends. The app works regardless of which conference platform you're on. Zoom, Teams, Meet, or a browser tab, the capture layer doesn't care.

3. Transcription and speaker diarization

Transcription runs via OpenAI's Whisper, downloaded and run on your own hardware. Four model sizes are available: base.en, small.en, medium.en, and large-v3. You pick the tradeoff between speed and accuracy in settings. Models download once; after that the app works offline.

Speaker diarization is automatic. Turingram uses ECAPA embeddings and spectral clustering to detect how many distinct speakers were present and separate their segments. You assign names after the recording. You can also hint the expected count before you start if you already know how many people are in the call.

4. agentHook output

Most transcription tools export text. Turingram also exports agentHook JSON: a structured format with timestamped segments, speaker labels, and meeting metadata. It feeds directly into AI agent workflows without a parsing step.

Drop the JSON file into a Claude Code session, pipe it to an n8n workflow, or upload it to a Claude Project. The format is designed to be machine-readable first.

Four export options in total: Markdown for notes, SRT for video sync, generic JSON, and agentHook. Same recording, different shapes for different tools.

5. Privacy

There is no server component. Whisper runs locally. The audio files stay on your machine. No account is required to use the app or access any of its features.

Data privacy here is a consequence of the architecture, not a policy claim. The data can't go anywhere because the design doesn't send it anywhere.