Delivery specs¶

A delivery spec is the rule set a deliverable has to meet: an integrated loudness target and tolerance, a true-peak ceiling, and optionally short-term and momentary maxima, a loudness-range ceiling, and dialogue gating. Audel treats specs as data. Grading reads only the spec object, so moving a deliverable from one platform to another never needs a code change.

Built-in specs¶

audel specs

Name	Integrated	Tolerance	True peak	Extra	Miss is
`ebu_r128`	-23 LUFS	±0.5 LU	-1 dBTP		rejected
`ebu_r128_s1`	-23 LUFS	±0.5 LU	-1 dBTP	short-term ≤ -18 LUFS	rejected
`atsc_a85`	-24 LKFS	±2 LU	-2 dBTP		rejected
`op59`	-24 LKFS	±1 LU	-2 dBTP		rejected
`arib_tr_b32`	-24 LKFS	±1 LU	-1 dBTP		rejected
`netflix`	-27 LKFS	±2 LU	-2 dBTP	speech-gated	rejected
`spotify`	-14 LUFS	±2 LU	-1 dBTP	music	advisory
`youtube`	-14 LUFS	±2 LU	-1 dBTP		advisory
`apple_music`	-16 LUFS	±2 LU	-1 dBTP	music	advisory
`apple_podcasts`	-16 LUFS	±1 LU	-1 dBTP	speech	advisory
`podcast`	-16 LUFS	±2 LU	-1 dBTP	speech	advisory

A rejected miss is graded error, so the verdict is fail. Broadcasters reject deliveries that miss their loudness spec. An advisory miss is graded warning, because streaming platforms normalize playback loudness themselves.

The values follow the published recommendations. Contracts often tighten them. When yours does, write a custom spec rather than relying on a built-in.

Use a spec¶

CLIPythonEnvironmentREST

audel check promo.wav --spec ebu_r128_s1
audel qc master.mxf --spec atsc_a85 --html qc.html --json qc.json

from audel import check

report = await check("promo.wav", spec="ebu_r128_s1")
print(report.spec, report.verdict, [t.integrated_lufs for t in report.tracks])

export AUDEL_DELIVERY_SPEC=ebu_r128        # default for every check
export AUDEL_CONTENT_TYPE=speech           # enables the speech noise-floor check

curl -F file=@promo.wav -F spec=ebu_r128_s1 http://127.0.0.1:8000/check
curl http://127.0.0.1:8000/specs

With no spec set, Audel grades exactly as 0.1 did. It uses the AUDEL_LOUDNESS_TARGET preset with a ±2 LU advisory tolerance, and the report shows spec: "preset.podcast".

Custom specs¶

An operator spec is a JSON file. Unknown keys are rejected and every value is range-checked.

{
  "name": "acme_tv",
  "description": "ACME network delivery, 2026 contract",
  "integrated_lufs": -24,
  "tolerance_lu": 1,
  "max_true_peak_dbtp": -2,
  "max_short_term_lufs": -20,
  "max_lra_lu": 15,
  "gating": "program",
  "strictness": "reject",
  "content": "mixed"
}

audel check ep101_mix.wav --spec /etc/audel/acme_tv.json

Spec files are operator configuration

The CLI, the Python API and AUDEL_DELIVERY_SPEC accept a spec file path. The REST service and the MCP server accept built-in names only, so a remote caller or an agent can never point Audel at a file on the host.

Dialogue gating¶

Some specs measure loudness over speech only. Audel approximates dialogue gating in three steps. It finds speech with the Silero voice-activity detector bundled in faster-whisper, listening to every channel except LFE on its own, from one streamed decode high-passed at 100 Hz. A downmix would miss dialogue in some immersive channels, or in channels mixed in antiphase. A detected passage counts only where its channel is audible in the mix, within 20 dB of the loudest channel at that moment, so a voice buried under music elsewhere isn't taken for dialogue. It then computes BS.1770 integrated loudness from the main pass's own 400 ms loudness blocks, counting only blocks that lie entirely inside speech, with the standard absolute and relative gates. Then it grades that value against the target.

The finding records basis: "speech-gated" and dialogue_coverage, the fraction of the programme that is speech. Confidence is medium. If no dialogue is detected (less than 0.4 s), or the measurement fails, the grade says so (dialogue_unmeasured) and uses programme loudness. Without the [asr] extra, Audel falls back to programme loudness and marks the finding low confidence.

This is an approximation

Speech-gated loudness is not Dolby Dialogue Intelligence. When a contract requires a licensed dialogue-intelligence meter, use Audel as the pre-flight check and the licensed meter for sign-off.