Local GUI¶
audiobench --gui launches a small local Gradio app that wraps the same Python
APIs the CLI uses. It is optional and aimed at one thing: making it easier to
structure testing sessions (multi-suite, multi-model matrices) and to
browse results organised by session. Everything in the GUI can be done from
the terminal as well — the GUI is not a replacement for the CLI.
Install and launch¶
The app binds to http://127.0.0.1:7861 by default and opens your browser.
Flags:
--gui-port 7862— pick a different port--gui-host 0.0.0.0— expose to your LAN (default is localhost only)--no-open— don't auto-open a browser
Test Builder¶
The Test Builder composes a matrix.yaml — the same format consumed by
audiobench run-matrix --matrix matrix.yaml.
- Cells table — one row per
(suite, model)cell. Columns:name,suite,model,seed,limit,conditions,pack,profile. Use the suite/model dropdowns + Add cell to append rows; the table itself is editable for quick tweaks. - Gate thresholds (optional) — pass/fail caps that get applied to each
cell after it runs, the same checks
audiobench gateperforms. They turn a run into a CI-style green/red decision: e.g. fail ifsound_id.min_weighted_recall < 0.6orasr_robust.max_weighted_mean_wer > 30. Leave every field blank to skip gating; fill any field and the YAML preview gains agate:section. - matrix.yaml preview — live YAML on the right. Save writes it to disk; Load rehydrates the builder from an existing YAML or JSON file.
- Run matrix — executes every cell in-process (same code path as
audiobench run-matrix), writes per-cell run JSONs tooutput_dir, and drops asummary.jsonalongside them. The Results tab will pick it up.
Results¶
A "testing session" is any directory under results/ that contains a
summary.json (the file run-matrix writes). The Results tab lists those,
plus ad-hoc top-level results/*.json runs.
- Click a session row to see its per-cell summary table (the same one
audiobench run-matrixprints). - Click a cell row to load that run's headline summary
(
audiobench inspect-style detail can be opened from the Inspect tab).
You can point the results root input at any directory if your runs live
somewhere other than ./results.
Secondary tabs¶
These are thin forms over the existing CLI commands and each tab shows the equivalent terminal command so you can copy it if you prefer the shell:
- Run —
audiobench run <suite> --model <id> ... - Compare —
audiobench compare A.json B.json - Inspect —
audiobench inspect run.json --clip N/--mixture N - Gate —
audiobench gate run.json --thresholds gate.yaml - Push —
audiobench push run.json [--repo ...] [--dry-run]
Nothing in the GUI is privileged: any artifact it writes (matrix.yaml,
per-cell run JSONs, summary.json) is interchangeable with the CLI's output.