noisekit 0.1.5

Generate degraded speech datasets for noise-robust ASR benchmarking.

Takes a clean HuggingFace speech dataset, applies real-world degradation presets via audiomentations, and scores each output with PESQ, SNR, and NISQA, producing a JSONL manifest ready for noise-robustness benchmarking.

Six atomic degradation scenarios are built in: telephony (G.711 + low-bitrate codec), wideband codec compression, ambient noise, clipping distortion, and far-field reverb. Atomic presets compose into compound multi-condition scenarios.

[!NOTE] Degradations are programmatically simulated. Scores may not generalize to genuine production recordings; validate final benchmarks on annotated real-world data.

How it works

flowchart LR
    A[("HuggingFace\nDataset")] --> B["noisekit generate"]
    B --> C["7 atomic presets\ncodec · noise · reverb\ndropout · clipping"]
    B --> D["3 compound presets\nmulti-condition chains"]
    C & D --> E[("WAVs + metadata.jsonl\nPESQ · SNR · NISQA")]

Install

No installation needed. Run directly with uvx:

uvx noisekit --help

Or install for development:

git clone https://github.com/Karamouche/noisekit.git
cd noisekit
uv sync
uv run noisekit --help

Usage

Generate a degraded dataset

uvx noisekit generate \
  --dataset google/fleurs \
  --config en_us \
  --split test \
  --samples 300 \
  --preset telecom \
  --preset low_bitrate \
  --output ./benchmark_dataset \
  --seed 42

--preset is repeatable: pass it once per preset.

If your dataset stores transcripts under a different column name (e.g. utterance, raw_text, translation), use --transcript-column:

uvx noisekit generate \
  --dataset my-org/my-dataset --split test \
  --samples 100 --preset telecom \
  --transcript-column utterance \
  --output ./out

By default, noisekit tries these columns in order: text, sentence, transcription, normalized_text. An error is raised if none are found and --transcript-column is not set.

For noise, you can supply your own background-noise WAVs with --noise-dir (e.g. MUSAN, DEMAND, or FSD50K):

uvx noisekit generate \
  --dataset google/fleurs --config en_us --split test \
  --samples 300 --preset noise \
  --noise-dir ~/datasets/musan/noise \
  --output ./benchmark_dataset --seed 42

Output:

benchmark_dataset/
├── metadata.jsonl          # one entry per generated file (AudioFolder format)
└── audio/
    ├── sample_0000_telecom.wav
    ├── sample_0001_low_bitrate.wav
    └── ...

The output is directly loadable as a HuggingFace dataset:

from datasets import load_dataset
ds = load_dataset("audiofolder", data_dir="./benchmark_dataset")

Each metadata.jsonl entry:

{
  "file_name": "audio/sample_0042_telecom.wav",
  "source": "common_voice_en_23136613.mp3",
  "dataset": "google/fleurs",
  "language": "en-US",
  "preset": "telecom",
  "transcript": "the cat sat on the mat",
  "snr_db": 5.2,
  "pesq_mos": 2.78,
  "nisqa_mos": 2.14,
  "nisqa_noisiness": 1.93,
  "nisqa_discontinuity": 2.41,
  "nisqa_coloration": 1.87,
  "nisqa_loudness": 2.3
}

Score an existing audio folder

# File stats only (duration, RMS, peak)
uvx noisekit score ./audio_folder --output scores.json

# With PESQ + SNR (requires matching reference files)
uvx noisekit score ./audio_folder --reference-dir ./clean_audio --output scores.json

# Skip NISQA (faster, no model download)
uvx noisekit score ./audio_folder --no-nisqa --output scores.json

List available presets

uvx noisekit list-presets
uvx noisekit list-presets --verbose   # show full transform stack

Presets

Nine built-in presets: six atomic scenarios, three compound multi-condition presets, and a clean reference control. None use synthetic white noise; codec artifacts, real ambient recordings, and room simulation produce the degradation instead.

Atomic presets

telecom is scored with PESQ narrowband at 8 kHz (before the final upsample); all other presets are scored wideband at 16 kHz.

All dependencies, including pyroomacoustics (used by reverb), are bundled with no extra install needed.

noise accepts a --noise-dir pointing at a directory of background-noise WAVs (e.g. MUSAN, DEMAND, FSD50K). If omitted, noisekit auto-downloads a small MUSAN noise-only subset (~20 files, ~120 MB) to ~/.cache/noisekit/noise/musan_ambient/ on first use.

Compound presets

Compound presets chain two atomic presets together. Noise is applied first (acoustic environment), then codec or dropout (digital processing on the already-degraded signal).

You can also define your own compound preset with a chain: key in a YAML file:

name: my_compound
description: "Noisy environment then telephony codec"
chain:
  - noise
  - telecom

Custom presets

Pass your own YAML file with --preset-file:

uvx noisekit generate \
  --dataset google/fleurs \
  --samples 100 \
  --preset-file ./my_preset.yaml \
  --output ./output

Preset format:

name: my_preset
description: "Custom telephony simulation"
transforms:
  - type: Resample
    parameters:
      min_sample_rate: 8000
      max_sample_rate: 8000
    p: 1.0
  - type: Mp3Compression
    parameters:
      min_bitrate: 16
      max_bitrate: 32
      backend: lameenc
    p: 1.0
  - type: Resample
    parameters:
      min_sample_rate: 16000
      max_sample_rate: 16000
    p: 1.0

Any transform from audiomentations is supported. Use ${NOISE_DIR} as a placeholder for --noise-dir inside your preset YAML. Use chain: instead of transforms: to compose built-in atomic presets sequentially.

Requirements

• Python ≥ 3.10
• uv for uvx usage
• No system dependencies: MP3 encoding uses pure-Python lameenc, no ffmpeg needed