Acoustic Tribe War

0:0Round 1 · 0:00
50% 70% line · hold 3 s to win 50%

Sound check

Before the fight, teach the system who is who. Record the room, then each player alone. It learns a voiceprint for each of you, tries to pull your voices apart from a mixture, and scores how cleanly it can.

Input

Demo players are synthetic voices (Ember ≈ 112 Hz, Frost ≈ 228 Hz) so you can see the whole pipeline. Switch to Microphone to enroll real people.

01Room tone

Everyone stays quiet for 3 s. This becomes the ambient model and the noise floor.

02

Player A speaks or sings alone for 5 s, at match volume.

03

Player B speaks or sings alone for 5 s, at match volume.

04Separation test

Mixes both solo takes with the room tone, separates the mixture with your voiceprints, and compares each recovered voice with the original take. References are your solo takes, so they carry some room noise; treat scores as relative.

MetricEmberFrostWhat it tells you
Record the three takes above; the test runs automatically.

05Who-is-who drill

15 s of prompts: A, B, silence, B, A, silence. The live tracker labels every 32 ms frame; the table shows what it heard against what was prompted.

Not run yet
Prompt ↓ / Heard →EmberFrostRoomOutsider
No data yet

06Voiceprints & verdict

The verdict appears after the separation test.

PESQ (ITU-T P.862) is not computed in the browser. Export the test kit: it contains the WAVs and a Python script that runs PESQ, STOI and BSS-Eval with the reference libraries.

← Wenjie Lin