Everything you'd want to know about getlivebpm.com. Still stuck? Contact us.
getlivebpm.com is a real-time BPM (tempo) detector that runs entirely in your browser. Open the page, let it listen — from your microphone or a shared audio tab — and it reports the BPM within seconds. No installs, no uploads, no account.
A rolling ~12-second audio buffer is resampled to 11.025 kHz and converted to a log-mel spectrogram. A compact TempoCNN neural network (running in TensorFlow.js) predicts a 256-class softmax over BPM values from 30 to 285. Parabolic interpolation around the peak gives decimal precision.
No. All audio capture and all inference happens locally in your browser. Nothing is sent to a server, nothing is recorded, and you don't need an account.
Only for the initial page load (which includes the ~5 MB model). Once the page is loaded, detection runs fully offline. If the browser has cached the page and model, subsequent visits work without a connection too.
TempoCNN was trained on ~12-second audio windows. Shorter windows miss the rhythmic structure (kick-snare patterns, downbeats, repetition) the model relies on. At 30 BPM — the low end of the detection range — one beat takes 2 seconds, so you need at least 12 seconds to see six beats.
After the initial 12-second fill, inference runs every 1 second using the most recent 12 seconds of audio — so readings update in near-real-time.
Microphone input works in every modern browser (Chrome, Edge, Firefox, Safari). System audio capture — the "Share audio" option when sharing a tab — works best in Chrome and Edge. Firefox and Safari have more limited support for that specific feature.
When no audio source is active, you can tap the button in time with a beat you hear or imagine. The app averages the intervals between taps to compute a BPM. If you stop tapping for more than 2 seconds the history resets — so you can start over cleanly any time.
On clean music with a clear beat, TempoCNN is typically within ±0.5 BPM. Classical music, spoken word, very quiet signals, and songs with tempo changes are harder — the console log shows a confidence value; readings below ~30% confidence should be treated with care.
No audio data ever leaves your device. We run standard anonymous analytics (Google Analytics) to understand basic traffic patterns — see our privacy policy for details.
The "DeepSquare k16-3" variant of TempoCNN, a compact convolutional network (~3M parameters) from Schreiber & Müller's 2019 SMC paper on tempo and key estimation. We converted the original TensorFlow frozen graph to tfjs-graph-model format for in-browser inference.
About five seconds to detect, zero data leaving your device.