HoldSpeak
Local voice typing, meeting transcription, and grounded Q&A over your work.
Tested on LinuxKnown vulnerabilities
Use it
Before you install: known vulnerabilities
pickfit found 20 known high or critical vulnerabilities in the packages a fresh install brings (OSV.dev, 3 Nov 2026). Weigh this before you install it.
pypi:anyio 4.12.1AnyIO: TLSStream IDNA 2003 host name encoding enables potential TLS certificate spoofing (GHSA-82r6-8w77-94w6, critical, fixed in 4.14.2)pypi:certifi 2026.1.4certifi bundles emSign root CAs with compromised intermediate CA keys (GHSA-6q48-6c7w-9gw2, critical, fixed in 2026.10.18)
See 8 more
npm:@xmldom/xmldom 0.8.13— GHSA-27p8-2357-5qqv (high), fixed in 0.8.15npm:@xmldom/xmldom 0.8.13— GHSA-4w3w-2rp5-g8jm (high), fixed in 0.8.14npm:@xmldom/xmldom 0.8.13— GHSA-8344-3jmq-59r6 (high), fixed in 0.8.15npm:@xmldom/xmldom 0.8.13— GHSA-93r5-fhx6-vmg9 (high), fixed in 0.8.15npm:@xmldom/xmldom 0.8.13— GHSA-965w-775f-mr7g (high), fixed in 0.8.15npm:@xmldom/xmldom 0.8.13— GHSA-c7q8-3ch8-vqpv (high), fixed in 0.8.15npm:@xmldom/xmldom 0.8.13— GHSA-w2rr-34g9-rvrj (high), fixed in 0.8.14npm:@xmldom/xmldom 0.8.13— GHSA-x4fp-j954-r2f4 (high), fixed in 0.8.15
10 more findings not listed here.
Linux x86_64
You need first
- Runtimes:
Python 3.10+,Node.js 22.12+
These steps ran from a folder named work in the home folder, which already existed. Make it and run them from there:
mkdir -p "$HOME/work" && cd "$HOME/work"git clone HoldSpeak and uv venv + uv pip install -e . (candidate 1); llama-cpp-python built from source in ~3 min
cd "$HOME"/work && rm -rf HoldSpeak && git clone --depth 1 https://github.com/karolswdev/HoldSpeak.git && cd HoldSpeak && uv venv && . .venv/bin/activate && uv pip install -e .
Where the README and reality differ
- README: uv pip install -e . (Linux: uv pip install -e '.[linux]')
What worked: The plain install succeeded; doctor then reports 'Transcription backend: Install Linux backend with pip install holdspeak[linux]', so the [linux] extra is needed for a working Linux transcription backend. Not tested with the extra.
Verified in pickfit: the commands exactly as its sandbox ran them
For the record only: they name the sandbox's home, /home/sbx.
cd /home/sbx/work && rm -rf HoldSpeak && git clone --depth 1 https://github.com/karolswdev/HoldSpeak.git && cd HoldSpeak && uv venv && . .venv/bin/activate && uv pip install -e .cd /home/sbx/work/HoldSpeak && . .venv/bin/activate && which holdspeak && holdspeak --helpcd /home/sbx/work/HoldSpeak && . .venv/bin/activate && holdspeak doctorCheck that it works
cd "$HOME"/work/HoldSpeak && . .venv/bin/activate && which holdspeak && holdspeak --helpThe output should contain “Examples:”.
First thing to try
cd "$HOME"/work/HoldSpeak && . .venv/bin/activate && holdspeak doctorWhat it does
These come from the README; pickfit didn’t check them one by one.
- Hold a key to dictate text into any appNot tested
- Transcribe meetings and confirm decisionsNot tested
- Ask questions about your meetings and projectsNot tested
- Send results to GitHub, Jira, or SlackNot tested
How far pickfit got
Linux x86_64
- Installs
- Starts
- Answers requests
- Works offline
macOS: the project says it runs there; pickfit hasn’t tried it.
Good for
- dictate by hotkey to type text into any app
- record or import a meeting for transcription
- ask grounded questions about meetings and projects
- send summaries to GitHub, Jira, Slack, or email
An alternative to superwhisper, MacWhisper, VoiceInk, Wispr Flow and Aqua Voice.
How pickfit tested it
Each run starts from a clean machine and follows the steps above. How pickfit verifies
Linux x86_64
Its service answers requests
- Replayed from scratch
- Works offline
- Checked on 16 Nov 2026
- Version 0.4.0
Differences the test machine hit
- README: a C++ compiler (the local model runtime builds from source)
In the test: llama-cpp-python 0.3.37 built from source during uv pip install (about 3 minutes); the box's gcc/g++/cmake was enough. - README: holdspeak (launches web runtime)
In the test: Headless, the runtime prints 'Voice typing hotkey unavailable' and doctor marks Global hotkey FAIL; the web server still runs with --no-open. The URL with a per-run token is printed in the log. - README: run holdspeak doctor if something does not work
In the test: doctor exits 1 when it reports FAIL items; on this box those were the hotkey and microphone, which need a desktop. Exit 1 is the doctor's report, not an install failure. - README: Set up local AI downloads the speech model and a starter language model (in the browser UI)
In the test: Not done: the default Dictation model is Qwen3.5-4B-Q4_K_M.gguf, too large for the inference check; the local model is set up through the Settings > Engines UI, and I did not find a headless generation command.