No description
  • Rust 83.2%
  • Dockerfile 8.3%
  • CMake 3.7%
  • Shell 3%
  • C 1.8%
Find a file
Dušan Bratuša 0eb497e98b
All checks were successful
Build / build (push) Successful in 1m2s
Release / release (push) Successful in 1m6s
Version 0.1.0-rc.2
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-07 17:27:40 +02:00
.forgejo Build release binaries for x86-64-v3 instead of no SIMD or the host CPU 2026-10-07 17:20:07 +02:00
src Build release binaries for x86-64-v3 instead of no SIMD or the host CPU 2026-10-07 17:20:07 +02:00
.gitignore Add README and ignore downloaded models 2026-10-07 14:32:25 +02:00
Cargo.lock Version 0.1.0-rc.2 2026-10-07 17:27:40 +02:00
Cargo.toml Version 0.1.0-rc.2 2026-10-07 17:27:40 +02:00
README.md Build release binaries for x86-64-v3 instead of no SIMD or the host CPU 2026-10-07 17:20:07 +02:00

Transcorder

Live transcription of your microphone ("You") and whatever your computer plays ("Them"): calls, meetings, videos. Built for macOS, Windows and Linux, and runs locally with whisper.cpp, so no audio leaves your machine.

[00:00:04] You: Can you hear me okay?
[00:00:07] Them: Yes, loud and clear. Let's get started.

Download

Ready-made Linux and Windows (x86_64) builds are on the releases page. They need a CPU with AVX2 (Intel since 2013, AMD since 2015). On Windows, unzip the archive and keep the DLLs next to transcorder.exe. Either way you also need a model (below).

Build

You need Rust 1.88 or newer, CMake, a C++ compiler and libclang (whisper.cpp is compiled into the binary).

  • macOS 14.2+: xcode-select --install and brew install cmake
  • Windows 10+: Visual Studio Build Tools (C++ workload), CMake and LLVM
  • Linux (Debian/Ubuntu package names): sudo apt install build-essential cmake clang libclang-dev pkg-config libasound2-dev libpipewire-0.3-dev
cargo build --release      # binary in target/release/transcorder
cargo install --path .     # or put transcorder on your PATH

Whisper uses the GPU through Metal on macOS. On Windows and Linux, add --features cuda (NVIDIA, needs the CUDA toolkit) or --features vulkan (needs the Vulkan SDK); otherwise it runs on the CPU.

Get a model

curl -L --create-dirs -o models/ggml-base.en.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin

In Windows PowerShell, type curl.exe rather than curl. Other models come from the same place:

File Size Notes
ggml-base.en.bin 148 MB English only, fast; the default
ggml-small.bin 488 MB Multilingual
ggml-large-v3-turbo-q5_0.bin 574 MB Multilingual and most accurate; wants a GPU or Apple Silicon

Run

./target/release/transcorder

That records the default microphone and everything the computer plays. Other options:

transcorder --list-devices                     # what --mic and --system accept
transcorder --mic "USB" --system "Headphones"  # pick devices by part of their name
transcorder -m models/ggml-small.bin -l de     # another model and language
transcorder --no-mic                           # only what the computer plays
transcorder -o meeting.txt                     # also save the transcript

The transcript goes to stdout and status messages to stderr. -o appends to the file, so earlier sessions in it are kept, and each session starts with a Recording started 2026-10-07 14:30:05 line giving the local time that [00:00:00] stands for. Ctrl+C stops recording and still transcribes what was being said; press it twice to quit at once.

macOS permissions

The first run asks for the microphone and for recording system audio. macOS gives these permissions to the app you run Transcorder from (Terminal, iTerm, VS Code and so on). If "Them" stays silent, allow that app under System Settings → Privacy & Security → Screen & System Audio Recording → System Audio Recording Only, then restart it.

How it works

  • The microphone is recorded with cpal. System audio comes from a Core Audio process tap of the whole system mix on macOS, WASAPI loopback on Windows, and PipeWire sink capture on Linux.
  • Each stream is downmixed to mono and resampled to 16 kHz. Then earshot detects speech, and an utterance ends at a 0.7 s pause.
  • Utterances go to Whisper on a separate thread. Recording the two sides separately is what labels them "You" and "Them", with no speaker detection needed.

Limitations

  • On speakers instead of headphones, the microphone also hears the other side, so their words show up twice. Use headphones.
  • Text appears after each pause, roughly 1–3 s later, not word by word.
  • --language auto detects the language for every utterance. That is slower and can misfire on short phrases, so pass the language code if you know it.
  • Linux needs PipeWire as the sound server (the default on current Ubuntu, Fedora and Debian).
  • So far it has only been tried on macOS; the Windows and Linux paths are untested.
  • Let people know you're transcribing them; some places require their consent.

Releasing

Set the version in Cargo.toml, then push a matching tag:

git tag v0.1.0 && git push origin v0.1.0

.forgejo/workflows/release.yml builds and tests it, then publishes a release with the Linux tarball and the Windows zip. A tag with a suffix, such as v0.2.0-rc.1, becomes a prerelease.