Loading page
Command-line tools for scripting and automation.
AudioLens doesn't ship a dedicated CLI binary — instead, a handful of
shell and Python scripts under scripts/ cover the operations you'd
otherwise script by hand: starting/stopping the stack, checking queue
health, and importing datasets.
| Script | Purpose | Needs |
|---|---|---|
download_ravdess.py | Fetches the 144-clip RAVDESS subset from Zenodo (~208 MB) into Backend/data/ravdess_subset/. --zip <file> reuses a zip you already have. | Nothing |
download_saa.py | Fetches the 150-clip Speech Accent Archive subset from accent.gmu.edu into Backend/data/SAA_dataset/ | Nothing |
download_librispeech_1000.py | Downloads the ~1000-row LibriSpeech-1000 sample from the HF datasets-server | Nothing |
prepare_l2arctic.py | Builds the 150-clip L2-ARCTIC subset and its phone-error annotations from a copy you downloaded: --source <extracted folder> | Free registration, see below |
prepare_common_voice.py | Builds the 100-clip Common Voice cv-valid-dev subset: --kaggle (Kaggle CLI) or --source <folder> | Free Kaggle account, see below |
prepare_savee_subset.py | Prepares a SAVEE subset (up to 100 clips) — see the callout below | A copy you already own |
fetch_datasets.py | Runs the no-account ones (RAVDESS, SAA, LibriSpeech-1000) together | Nothing |
import_librispeech_custom.py | Imports a LibriSpeech-style dataset you already have on disk | Local files |
run_layer_probe_analysis.py | Runs layer-probe analysis from the command line, outside the UI | Fetched datasets |
Run them from the repository root, for example
python scripts/download_ravdess.py. All the dataset scripts accept
--dry-run (show the plan, change nothing) and --out <dir> (default:
Backend/data/<dataset folder>), print the dataset's license and citation
first, and skip clips that are already present, so re-running is safe.
Open the L2-ARCTIC page,
fill in the form (name, email, affiliation), agree to CC BY-NC 4.0 and
complete the reCAPTCHA. An email titled "Access to L2-ARCTIC corpus"
arrives with a Google Drive link; download and extract it, then run
python scripts/prepare_l2arctic.py --source /path/to/extracted.
Citing Zhao et al. (2018) is required.
Create a free Kaggle account, accept the mozillaorg/common-voice
dataset's terms on its page, make an API token (Settings → API) and save it
as ~/.kaggle/kaggle.json, run pip install kaggle, then
python scripts/prepare_common_voice.py --kaggle. Already have the
download? Use --source /path/to/common-voice instead. Newer Common Voice
releases (23.0 onward) are distributed through the Mozilla Data Collective
and use different file names, so they don't match this subset.
No account. Just run the script.
docker compose commandsScale worker replicas:
docker compose up -d --scale worker-cpu=2 --scale worker-model-local=2Tear the stack down, optionally wiping volumes (Redis data, local storage):
docker compose down
docker compose down -vPre-warm model weights in a one-off container instead of waiting for the first real job to trigger the download:
docker compose run --rm worker-model-local python3 -c \
"from transformers import pipeline; pipeline('automatic-speech-recognition', model='openai/whisper-base')"See Self-hosting for the full boot sequence these commands fit into.