DNS-Noise / README.md
ChristianYang's picture
Add mic_ir split: MADIR + MicIRP + CTF 2020 Tiny IRs microphone impulse responses (24k mono FLAC)
47b9e46 verified
|
Raw
History Blame Contribute Delete
2.11 kB
---
license: cc-by-4.0
task_categories:
- audio-classification
tags:
- audio
- noise
- rir
- augmentation
- dns-challenge
- microphone
- impulse-response
pretty_name: DNS Noise + RIR (24k mono)
configs:
- config_name: default
data_files:
- split: noise
path: noise/*.parquet
- split: rir
path: rir/*.parquet
- split: mic_ir
path: mic_ir/*.parquet
---
# DNS-Noise
DNS-Challenge `noise_fullband` and `impulse_responses` republished as 24 kHz mono
FLAC for on-demand streaming augmentation, plus a microphone-characteristics
split. Three splits:
- `noise` — environmental noise (long files segmented into chunks)
- `rir` — room impulse responses (one row per file)
- `mic_ir` — microphone impulse responses (one row per file)
```python
from datasets import load_dataset
noise = load_dataset("ChristianYang/DNS-Noise", split="noise", streaming=True)
rir = load_dataset("ChristianYang/DNS-Noise", split="rir", streaming=True)
mic_ir = load_dataset("ChristianYang/DNS-Noise", split="mic_ir", streaming=True)
```
## mic_ir split
Microphone impulse responses for simulating microphone frequency colouration,
aggregated from three sources. `source_file` is prefixed with the source
dataset:
| prefix | source | content |
|---|---|---|
| `madir/` | [Zenodo 4633508](https://zenodo.org/records/4633508) — Multi-Angle, Multi-Distance Microphone Impulse Response Dataset (CC-BY-4.0) | studio mics, `Raw_IRs` only, one bit-depth variant per IR (24-bit preferred); folder names encode mic model, polar pattern, distance and angle |
| `micirp/` | [MicIRP](http://micirp.blogspot.com) via [audb](https://github.com/audeering/audb) v1.0.0 (CC BY-SA 4.0) | vintage/classic microphones |
| `ctf_2020_tiny_irs/` | Collected Transients "Tiny IRs" (2020) | consumer-device microphone captures (cellphone/tablet/laptop/camera mics, incl. mic-through-speaker chains); pure loudspeaker IRs are excluded |
Processing: mixed down to mono, resampled to 24 kHz (soxr HQ),
peak-normalised to 0.97, encoded as 16-bit FLAC. Absolute level is not
meaningful — normalise IR energy before convolution.