Count every sound hiding in the mix.

Drop in a single-channel recording and EchoCount estimates how many sources overlap, names each sound, and separates sparse speech. Processed on the server with automatic 30-minute cleanup.

Podcast panels Café chatter Street scenes Choirs Open-plan offices Classrooms Rainforests Rush-hour traffic Orchestras

Built around one question

Counting is a different job from separating. EchoCount answers how many first, and it shows how sure it is.

Hear it, then see it

Your upload becomes a waveform you can play and scrub. The part the model analysed is marked, so every result can be checked by ear.

One prediction per clip

The whole clip matters

The model reads one complete eight-second spectrogram and estimates one source count with MC-dropout uncertainty.

Measured, not promised

Exact count accuracy across held-out test clips (n=1,285): 43.8% exact, 64.3% within ±1.

Unseen speakers (n=352): 13.1% exact · MAE 2.05

What it can hear

0sound classes (527 AudioSet + 32 domain)
0sound groups
0active mixtures
2–31sources covered
Hammer open-vocabulary Phone vibration open-vocabulary Dog bark Cat meow Cello Scissors Telephone ring Human voice open-vocabulary Door knock open-vocabulary Keyboard typing Water open-vocabulary Car horn open-vocabulary

Fast and checked

Every separated stem passes 3 checks: self-consistency, correlation (< 0.9), and energy (±5%).

$ 8s clip · Windows CPU (AMD64, PyTorch 2.15.0+cpu)
decode 11.7 ms · feature 37.5 ms
count 11.6 ms · tag 1.336 s
separate 1.065 s · zip 20.7 ms
snapshot 2026-11-15

The model is ready.

A trained checkpoint was found, so predictions are available now.

Validation results for this checkpoint: MAE 1.60 sources · macro MAE 1.60 · exact count 45.0% · within one source 65.9%.

0sample rate
0analysis window
0mel bands
0maximum count

From file to count

  1. Load

    Convert to mono and resample to 16 kHz.

  2. Crop or pad

    Fit the clip to one 8-second window.

  3. Convert

    Turn the full clip into a 64-band log-mel spectrogram.

  4. Classify

    A small CNN estimates one source count for the full clip.

Questions people ask

Does it separate the sources into files?

For 2-speaker speech mixtures, yes — the installed SepFormer model separates two speakers into WAV stems plus a residual and ZIP bundle. General non-speech separation and 3+ speaker separation are disabled unless a trained separator checkpoint is loaded.

Why does it say the model is not trained?

The app needs a checkpoint at checkpoints/source_count_clip.pt. Run python scripts/download_models.py or train with python scripts/train.py all --full. The Docs page lists the exact commands.

What happens to my audio?

Uploaded audio is processed on the server in an isolated job folder and deleted automatically after 30 minutes and on server startup. Your run history stays in this browser only.

What about recordings longer than 8 seconds?

Only the middle 8 seconds are analysed, and the waveform marks which part. Shorter recordings are padded with silence.

Can I trust the confidence number?

It shows how decisive the network was, not whether it was right. Check accuracy on a held-out test set before relying on it.

Drop a file. Get a count.

No account required. Try a synthetic mixture if you do not have audio handy.

Open the analyzer