Room & Noise Simulator

Hear how a clean speech recording would sound in a real room, with real background noise, through a laptop microphone or a Bluetooth headset. Use it to test a voice assistant, a speech recogniser or a voice activity detector on the same sentence under controlled conditions. The rooms are measured impulse responses, the noises are real recordings, and everything runs in your browser: your audio never leaves your machine.

1. Your recording

Choose a clean speech file (WAV, FLAC, MP3, OGG; up to 2 minutes)

2. Room

RT60 = time for the sound to decay by 60 dB. The higher, the more echoey.

3. Background noise

Signal-to-noise ratio: 10 dB(lower = noisier; 20 dB is quiet, 0 dB is as loud as the voice)

4. Microphone path

5. Output sample rate

Testing a voice assistant with it

Record a few sentences close to the microphone in a quiet room. That clean take is your reference.
Generate the same sentences across rooms and SNRs (for example 20, 10 and 5 dB), and measure word error rate on each file. Only the condition changes, so differences come from the room and the noise.
Feed the noise-only files to your voice activity detector and count detections. Background noise mistaken for speech shows up here.
Compare the built-in microphone with the Bluetooth paths. Headsets switch to a narrow voice band when their microphone is on, and recognition often degrades.

What it does, and its limits

Chain: your recording is convolved with the room's impulse response, the noise is added so that the active speech level sits the chosen number of dB above it, then the whole signal goes through the microphone path and is resampled to the output rate. Silence around the speech does not count in the SNR. The SNR is set at the microphone, before the Bluetooth path: a headset cuts part of the noise along with part of the voice, so the SNR you measure on a Bluetooth file can differ from the one you chose, as it would with a real headset.

The rooms (except the large hall) were measured from the mouth of a dummy head to a phone lying on a table in hands-free position: close to a voice assistant or a laptop on a desk.
The noise is recorded in other places than the rooms, and is added without reverberation.
The Bluetooth paths reproduce the bandwidth of the HFP voice codecs, not their coding artefacts or packet losses.
No living room, bedroom or car response yet: we found none measured with a licence that allows reuse.

Sources and licences

Room impulse responses: Aachen Impulse Response Database v1.4, RWTH Aachen University (MIT licence). Noises: DEMAND database, Thiemann, Ito & Vincent (CC BY-SA 3.0; the trimmed files served here are CC BY-SA 3.0 too), and two CC0 recordings from Freesound (BigKahuna360, DCPoke). Full details: LICENSES.txt.