WakeCore
Open hotword detection for private voice systems. Lokale Hotword-Erkennung für private KI-Systeme.
WakeCore detects a hotword ("wake word") such as Hey Computer in a live audio
stream, entirely on the device: no cloud calls, no API keys, no telemetry. You
load a .wake file, feed 16 kHz audio frames and get a detection event when the
phrase is spoken.
Status (October 2026): SDK 0.1.1. The native engine is available for Linux x86_64 (glibc 2.29 or newer, e.g. Ubuntu 20.04+, Debian 11+). Other platforms are not available yet. See CHANGELOG.md.
Packages
| Language | Path | Install |
|---|---|---|
| Python | wakecore/ |
pip install wakecore (PyPI) |
| Node.js | node/ |
npm install wakecore (npm) |
| C | include/wakecore.h |
header, link libwakecore_engine |
| C++ | cpp/ (header-only) |
header, link libwakecore_engine |
| Rust | rust/ |
git dependency (not on crates.io) |
| C# / .NET | csharp/ |
project reference (not on NuGet) |
All bindings call the same native engine (libwakecore_engine.so), which is
downloaded separately.
1. Install the engine
pip install "wakecore[audio]" # Python 3.10+; [audio] adds microphone input
wakecore install-engine # downloads the engine to ~/.wakecore/0.1.0/
wakecore check hotwords/hey_computer.wake
install-engine fetches the engine from
https://download.wakecore.de/engine/linux/x86_64/0.1.0/, checks it against
SHA256SUMS and puts the license notices next to it. The engine version
(wakecore.ENGINE_VERSION) is independent of the SDK version, so SDK updates keep
using an installed engine.
Without Python, download the same files by hand:
mkdir -p ~/.wakecore/0.1.0 && cd ~/.wakecore/0.1.0
for f in libwakecore_engine.so SHA256SUMS NOTICE LICENSE-Apache-2.0.txt; do
curl -fsSO "https://download.wakecore.de/engine/linux/x86_64/0.1.0/$f"
done
sha256sum -c SHA256SUMS
The SDKs look for the engine in WAKECORE_NATIVE_DIR, then in
~/.wakecore/0.1.0/.
2. Detect a hotword
Python
wakecore listen hotwords/hey_computer.wake # microphone, Ctrl+C to stop
from wakecore import Runtime, AudioCapture
with Runtime.load("hotwords/hey_computer.wake") as rt, \
AudioCapture(rt.sample_rate, rt.frame_length) as mic:
for frame in mic: # 512 samples, 16 kHz, int16
if rt.process(frame):
print("wake!")
Runtime.load(path, sensitivity=0.6) overrides the sensitivity stored in the
file; rt.set_sensitivity() changes it while running. After a detection, further
detections are suppressed for cooldown seconds of audio (default 0.4).
Node.js
npm install wakecore # uses koffi (installed as optional dependency)
const { Runtime } = require("wakecore");
const rt = Runtime.load("hotwords/hey_computer.wake", { sensitivity: 0.5 });
// feed Buffers of rt.bytesPerFrame bytes (512 samples, 16-bit LE, mono, 16 kHz)
const hit = rt.process(frame);
if (hit) console.log("wake!", hit.frameIndex);
rt.close();
C / C++
#include "wakecore.h"
/* wc_engine_open(body, len, &engine) with the bytes after the 8-byte .wake header,
then wc_engine_feed(engine, frame, &detected) per 512-sample frame. */
#include "wakecore.hpp"
auto rt = wakecore::Runtime::load("hotwords/hey_computer.wake");
if (rt.process(frame)) { /* wake */ } // frame: rt.frame_length() int16 samples
g++ -std=c++17 -Iinclude -Icpp/include app.cpp \
-L"$HOME/.wakecore/0.1.0" -lwakecore_engine -Wl,-rpath,"$HOME/.wakecore/0.1.0"
Rust
[dependencies]
wakecore = { git = "https://github.com/Keyvanhardani/wakecore", tag = "v0.1.1" }
let mut rt = wakecore::Runtime::load("hotwords/hey_computer.wake")?;
if rt.process(&frame)? { println!("wake!"); } // frame: [i16; 512]
build.rs links against the engine in WAKECORE_NATIVE_DIR or
~/.wakecore/0.1.0/; at run time set LD_LIBRARY_PATH=~/.wakecore/0.1.0.
Try it with cargo run --example detect_wav -- ../hotwords/hey_computer.wake audio.wav.
C# / .NET
Reference csharp/WakeCore.csproj (net8.0):
using var rt = WakeCore.Runtime.Load("hotwords/hey_computer.wake");
if (rt.Process(frame)) Console.WriteLine("wake!"); // frame: short[512]
Sample hotwords
hotwords/ contains eight ready-made files (also on
Hugging Face). They use .wake format
v2 and need the engine published on 28 Oct 2026 (Linux x86_64); the format v1
files shipped with SDK 0.1.0 do not work with it.
| File | Say | Default sensitivity | Detected | False alarms in speech | Fires on other samples |
|---|---|---|---|---|---|
computer_an.wake |
Computer an | 0.4 | 34/36 | 8 (all with "Computer") | 9/12 Hey Computer |
guten_morgen.wake |
Guten Morgen | 0.7 | 24/36 | 1 | – |
hallo_welt.wake |
Hallo Welt | 0.7 | 33/36 | – | – |
hey_assistant.wake |
Hey Assistant | 0.4 | 35/36 | 2 ("the assistant …") | – |
hey_computer.wake |
Hey Computer | 0.4 | 36/36 | 6 (all with "Computer") | 11/12 Computer an |
hey_keyvan.wake |
Hey Keyvan | 0.5 | 36/36 | – | – |
okay_deutsch.wake |
Okay Deutsch | 0.5 | 36/36 | – | – |
wake_up.wake |
Wake up | 0.5 | 35/36 | 1 ("woke up") | – |
How this was measured (28 Oct 2026, each file at its default sensitivity): Detected = 12 recordings of the phrase (three synthetic voices, alone and inside a sentence) under three conditions: normal, 20 dB quieter, and with room reverb. False alarms in speech = recordings with at least one detection in 9.3 minutes of German and English speech without the hotword (48 sentences and paragraphs × three voices; 21 of them contain similar words such as "Computer", "Morgen", "assistant", "woke up"). Fires on other samples = how many of the 12 recordings of another sample phrase trigger this file. Synthetic voices are not real rooms and microphones, so treat the numbers as a rough guide.
What follows from it:
- Computer an and Hey Computer react to the word "Computer" on its own and to each other. Use only one of them, and not where "Computer" is said often.
- Guten Morgen is the weakest sample, especially with reverb (4 of 12), and of course fires whenever someone actually says "Guten Morgen".
- Hey Assistant and Wake up occasionally fire on "the assistant …" or "woke up".
Sensitivity runs from 0 to 1: higher detects more and fires falsely more often. Test with your own microphone and room before relying on a value. Custom hotwords for your own phrases: see https://wakecore.de.
Audio format
16 kHz, mono, 16-bit signed little-endian PCM, 512 samples (32 ms) per frame.
License
The SDK in this repository is MIT licensed, see LICENSE.
The engine binary and the .wake files contain a third-party component under
the Apache License 2.0. Its notice and license text are installed next to the
engine (NOTICE, LICENSE-Apache-2.0.txt) and are also in
hotwords/. Keep them with the engine and the .wake files when
you pass them on.
Metadata
Release files for wakecore 0.1.1
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| wakecore-0.1.1.tar.gz | 26.8 kB | With provenance | Details for wakecore-0.1.1.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| wakecore-0.1.1-py3-none-any.whl | Python 3 | none | any | Details for wakecore-0.1.1-py3-none-any.whl |
Total release size: 49.3 kB
Release files / wakecore-0.1.1.tar.gz
| Download URL | wakecore-0.1.1.tar.gz |
|---|---|
| Size | 26.8 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
3d2f9fed582e71bc0df5c3ac32e732de55feb36ca05892a91c5124f005bc816d
|
|
BLAKE2b-256 checksum How to use checksums |
c4540551d53031104d6da8f28a208fbc96a9c451016bccdd2277fed207c491fd
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 28, 2026.
Transparency logRelease files / wakecore-0.1.1-py3-none-any.whl
| Download URL | wakecore-0.1.1-py3-none-any.whl |
|---|---|
| Size | 22.5 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
c6cb434dfac1df2e389e3da65a8723c36a1a533d9f7480c87286d1af812f3b1e
|
|
BLAKE2b-256 checksum How to use checksums |
e800a6b754259fe34f961ca26a896615b4d728a76083ed3fea7c0fff7a996bb5
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 28, 2026.
Transparency logSecurity for wakecore 0.1.1
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for wakecore-0.1.1.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 28, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.