How to Transcribe Meetings Without Uploading to the Cloud

Every major meeting transcription service — Otter, Fireflies, tl;dv, Fathom — works by uploading your audio to their cloud infrastructure. That’s how they run speech recognition at scale, and it’s a reasonable trade-off for many people.

But it’s not the only option. Here’s how to transcribe your meetings without the audio leaving your machine.

Why avoid the cloud?

There are a few reasons you might want local transcription:

Privacy. Your meetings contain sensitive information — client details, business strategy, HR discussions. You may prefer not to have that audio sitting on a vendor’s servers.

Compliance. NDA clauses and internal security policies can create constraints around sharing meeting content with third parties. Local transcription eliminates this concern for the audio processing stage.

Security. Cloud services can be breached. If your meeting audio is never uploaded, it can’t be accessed in a breach of the transcription vendor.

Cost. Cloud transcription services charge per minute or per seat. Local transcription using open-source models is free beyond the compute cost on your hardware.

Offline use. Local transcription works without an internet connection. Useful if you’re recording in environments with restricted connectivity.

Option 1: Use whisper.cpp directly

whisper.cpp is a C++ implementation of OpenAI’s Whisper model. It runs locally, supports GPU acceleration, and produces high-quality transcriptions.

Setup:

# Clone and build
git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp
make

# Download a model
./models/download-ggml-model.sh large-v3

# Transcribe a WAV file
./main -m models/ggml-large-v3.bin -f meeting.wav -otxt

This works well if you’re comfortable with the command line. The friction is that you need to manage recording yourself, convert audio to WAV format, and process files manually after each meeting.

Option 2: Combine recording + whisper.cpp with a script

You can automate the pipeline with a shell script:

#!/bin/bash
# Record system audio (macOS example using BlackHole)
ffmpeg -f avfoundation -i "BlackHole 2ch" -ar 16000 meeting_$(date +%Y%m%d_%H%M).wav &
RECORD_PID=$!

echo "Recording... Press Enter to stop."
read

kill $RECORD_PID
./whisper.cpp/main -m ./models/ggml-large-v3.bin -f meeting_*.wav -otxt

On Linux, you’d use PulseAudio or PipeWire for capture. This is more involved but gives you full control.

Option 3: Use a dedicated app

If you want local transcription without managing the infrastructure yourself, Harknotes handles the full pipeline:

Summaries run through any OpenAI-compatible endpoint — a local model via Ollama keeps the whole pipeline offline, or you can bring your own API key for a cloud provider (which sends the text transcript, never audio). You can skip this step if you want the raw transcript only.

Comparing the approaches

Approach	Setup effort	Offline	Speaker labels	Structured notes
whisper.cpp directly	High	Yes	No	No
Custom script	Medium	Yes	No	No
Harknotes	Low	Yes (transcription)	Yes	Yes

What about quality?

Whisper’s large-v3 model is genuinely excellent. For clear meeting audio with two to eight speakers, accuracy is typically over 95% for English. Other languages vary but are generally competitive with cloud alternatives.

The main limitation is speed on CPU. On a modern Apple Silicon Mac or a machine with a dedicated GPU, transcription is fast — often faster than real-time. On older hardware or CPU-only machines, you may wait a few minutes for a long meeting to process.

Getting started with Harknotes

Harknotes is free and open source (AGPL-3.0). Transcription and summaries are unlimited — you bring the AI, either a local model or your own API key.

Download Harknotes →