rendercheck

The worst bugs in generated media don't throw.

"the audio often cuts off the final sentence […] though the API returns success without error signals"

— a developer on the OpenAI forum, April 2026, describing production output

If you generate speech or video with a model — TTS, voice agents, podcasts, avatars, AI video — your tests catch the exception that never happens. They do not catch the narration that reads at 300 words per minute, the voice track sitting 18 dB below the footage it's cut against, the clip that rendered at 42% length and got cached as a success, the captions that describe the audio three seconds before it happens, or the file whose audio track is missing entirely.

The 2026 state of the art for catching these is a person listening to the output. That works, and it costs more than everything else in your pipeline combined.

rendercheck makes them throw.

from rendercheck import assert_pace, assert_loudness, looks_ok

assert_pace("episode-12.mp3", "episode-12.vtt")
assert_loudness("episode-12.mp3")
looks_ok("slide-14.png", ["the title fits on one line"])

Plain assert functions. No framework, no runner, no service. They raise AssertionError, so they already work in pytest, in CI, or in a five-line script. Eighteen of the nineteen checks have no dependencies and make no network calls — if you have ffmpeg, you're ready.

Quickstart

You need ffmpeg on your PATH (brew install ffmpeg, apt-get install ffmpeg, or winget install ffmpeg). Then:

pip install rendercheck
rendercheck demo

Or drop a file into the playground — same checks, running on ffmpeg compiled to WebAssembly, nothing uploaded.

demo synthesises ten defective files and runs the real checks against them, so you can see it fire without owning a broken render. Verbatim, first two of eight:

Narration too fast
  A voice picked to match a presenter's face read English at machine-gun speed. Valid audio, correct timing, perfectly in sync.

  $ rendercheck check machine-gun.wav --script narration.vtt

  FAIL  pace        narration pace 300 WPM exceeds 245 (300 words in 60.0s) -- this reads as machine-gun delivery and listeners cannot follow it: machine-gun.wav
  PASS  loudness    -16.1 LUFS
  PASS  dead air    0.0 s silence
  PASS  truncation  8.9 dB of fall-off at the end
  PASS  clipping    0 samples at 0 dBFS

Levels that don't match
  Synthesised narration landed 18 dB under the footage it was cut against. Nobody noticed until viewers rode the volume knob.

  $ rendercheck check too-quiet.wav

  SKIP  pace        no --script given
  FAIL  loudness    -34.2 LUFS is 18.2 dB quieter than the -16 target -- it will sound inaudible next to correctly-levelled audio cut alongside it: too-quiet.wav
  PASS  dead air    0.0 s silence
  PASS  truncation  8.8 dB of fall-off at the end
  PASS  clipping    0 samples at 0 dBFS

…and one of the two added in 0.3.0:

Captions against the wrong clock
  A concatenation added three seconds of pre-roll after the captions were written. Both files are perfectly valid on their own.

  $ rendercheck check late-captions.wav

  SKIP  pace        no --script given
  PASS  loudness    -16.0 LUFS
  PASS  dead air    0.0 s silence
  PASS  truncation  74.3 dB of fall-off at the end
  PASS  clipping    0 samples at 0 dBFS
  FAIL  captions    late-captions.vtt runs 3.0s late against late-captions.wav, past the 0.75s limit -- every line arrives at the wrong moment, and both files are individually valid so nothing else catches it

Then point it at your own output:

rendercheck check episode-12.mp3 --script episode-12.vtt --preset podcast

Exit code is 1 if anything failed — or if nothing could be measured, because a run that looked at nothing is not a clean one. A path you typo'd exits 2. --json gives you the same report for pipelines in any language, and --strict rejects partial runs too.

Where is this file going?

"How loud should this be?" has no single answer — it depends entirely on where the file ends up, and every platform publishes a different number. --preset turns that table into something a build can enforce:

$ rendercheck presets

  preset    target     tol     peak  source
  youtube     -14L   1.0dB   -1.0TP  YouTube normalises playback to -14 LUFS
  spotify     -14L   1.0dB   -1.0TP  Spotify, including podcasts, at -14 LUFS
  tiktok      -14L   1.5dB   -1.0TP  TikTok and Instagram, measured rather than published
  podcast     -16L   1.0dB   -1.0TP  AES71 / Apple Podcasts: -16 LUFS stereo, -19 mono
  apple       -16L   1.0dB   -1.0TP  Apple Music Sound Check, -16 LUFS
  web         -16L   2.0dB       --  spoken-word web video -- rendercheck's own defaults
  ebu         -23L   1.0dB   -1.0TP  EBU R128, European broadcast
  atsc        -24L   2.0dB   -2.0TP  ATSC A/85, North American broadcast
  netflix     -27L   2.0dB   -2.0TP  Netflix delivery, dialog-gated

None of those numbers are ours. The contribution is that --preset ebu is a decision a reviewer can read, where --target-lufs -23 is a magic number the next person will not dare touch. A preset that states a ceiling also switches on the true-peak check, which catches a master measuring clean locally and distorting after upload. web exists only to name the built-in defaults, so it states none and behaves exactly like passing no preset at all.

Project-wide settings go in rendercheck.toml (or [tool.rendercheck] in pyproject.toml) so a CI step is not eight flags on one line:

preset = "podcast"
max_silence = 5.0

Flags you type still beat the file, and the file beats the built-in defaults.

In pytest they're just asserts — no plugin, no fixtures:

@pytest.mark.parametrize("episode", EPISODES)
def test_episode_is_shippable(episode):
    assert_pace(episode.audio, episode.vtt)
    assert_loudness(episode.audio)
    assert_no_dead_air(episode.audio)

"Isn't this forty lines of pyloudnorm?"

For one of the nineteen checks, roughly yes. None of these measurements are novel, and it would be dishonest to imply otherwise:

Most of those hand a number to a researcher. The two that are already gates gate a different thing: MediaConch checks that a file conforms to a container policy, which is a preservation question, not a perceptual one — a file can pass every MediaConch rule and still be narrated at 300 WPM. ffsubsync will happily realign captions that were never wrong, because it has no opinion about whether they needed it.

What is actually missing, and what this is:

• A threshold that came from a defect, not from a paper. 245 WPM because a real voice narrated at 280 and shipped. −16 LUFS because narration landed at −34 against footage at −13.
• A message that says what a person would notice. "−34.0 LUFS" is a reading. "18.0 dB quieter than the −16 target — it will sound inaudible next to correctly-levelled audio cut alongside it" is a bug report.
• Fail-open on infrastructure, fail-closed on a defect, so it can sit in CI without becoming the thing that breaks the build for its own reasons.
• Exit codes and one command over a directory, rather than a notebook.

Against the LLM-eval tools the difference is structural rather than a matter of coverage. promptfoo, DeepEval and RAGAS are excellent and none of them can do this: their test case is a string. There is no assertion to add, because there is nowhere to put the file. Use them for the script; use this for what the script turned into.

And if you already run broadcast QC — Interra BATON, Telestream Vidchecker, QCTools — you have had most of this for twenty years. It just isn't in your git hooks.

In your pipeline

GitHub Actions — installs ffmpeg and fails the build on a defect:

- uses: rogermsc/rendercheck@v0
  with:
    files: out/
    preset: podcast
    strict: "true"

Node, Remotion, anything that renders in a build step:

npx rendercheck check out/

Docker, if you would rather not have a Python toolchain at all — ffmpeg is already in the image:

docker run --rm -v "$PWD:/work" ghcr.io/rogermsc/rendercheck check /work/out.mp4

Coding agents, via MCP. An agent that just wrote a render pipeline and ran it has no way to tell whether the file that came back is any good; the other media MCP servers cut and transcode, which hands it more media rather than an answer:

claude mcp add rendercheck -- rendercheck mcp

It is listed in the MCP registry as io.github.rogermsc/rendercheck, so clients that read the registry can install it without being told where it lives. No key is needed — every check except looks ok is deterministic.

check_media returns one verdict per check with the measured value, so the model can act on "−34 LUFS, 18 dB under target" rather than on a file it cannot hear.

promptfoo — its assertions are all string-shaped, so an eval can confirm the narration script and tell you nothing about the audio. examples/promptfoo/ closes that half in thirty lines:

assert:
  - type: python
    value: file://rendercheck_assert.py:get_assert

Anything else — --json on stdout, one object per file, plus exit codes.

Six checks, six incidents

Each default is a threshold set by a defect that actually shipped, not a number chosen for symmetry. All output below is verbatim. Every threshold is an argument — see the reference for tuning.

Narration too fast. A voice picked to match a presenter's face narrated English at ~280 WPM. The audio was valid, correctly timed, perfectly in sync. It just sounded like a machine gun.

assert_pace("episode-12.mp3", "episode-12.vtt", max_wpm=245)

narration pace 300 WPM exceeds 245 (300 words in 60.0s) -- this reads as
machine-gun delivery and listeners cannot follow it: episode-12.mp3

Levels that don't match. Synthesised narration landed at −34 LUFS and was concatenated with footage at −13. Same file, a 20 dB step in the middle. Nobody noticed until viewers spent 45 minutes riding the volume knob.

assert_loudness("episode-12.mp3", target_lufs=-16, tol=2.0)

-34.0 LUFS is 18.0 dB quieter than the -16 target -- it will sound inaudible
next to correctly-levelled audio cut alongside it: episode-12.mp3

Truncated renders cached as successes. Encode failures produced clips a fraction of their intended length, which the pipeline cached as succeeded. Retries only re-ran the ones that had errored — and these hadn't.

assert_duration("segment-07.mp4", expected_seconds=24.0)

segment-07.mp4 is 10.0s -- 42% of the expected 24.0s. A render this short is a
silent encode failure, not a short take; re-render rather than retry

Holes in the middle. Compositing failed transiently and silently, leaving dead stretches mid-file. Right length, right average loudness. The hole only existed in the middle.

assert_no_dead_air("episode-12.mp3", max_silence=3.0)

6.2s of silence starting at 0:41 exceeds the 3s limit -- a gap this long
mid-file is a dropped segment, not a pause (2 found in total): episode-12.mp3

The wrong person speaking. A script said "I'm Jordan" while the system had assigned Alex. A whole module rendered with the wrong face and the wrong voice. Every other gate passed.

assert_speaker(script, expected="Alex", known_names=["Alex", "Jordan", "Sam"])

the script introduces the presenter as "Jordan" but Alex is assigned -- the
rendered avatar would introduce itself with someone else's name. Fix whichever
is wrong: the assigned presenter, or the name in the script

The known_names roster is required, and it's the whole trick: without it, a character in a scenario saying "I'm Rosa, a nurse" trips the check on every script that tells a story. Only a name belonging to someone who could actually have been cast counts as a claim about the speaker.

Things you can only see. Overflowing titles, colliding logos, half-empty canvases, figures cropped mid-caption. All rendered without error.

looks_ok("slide-14.png", ["the title fits on one line", "no text is clipped"])

[major] slide-14.png: the title wraps to three lines and overlaps the logo in
the top-right corner -- failed rubric item: 'the title fits on one line'

This is the only check that needs a key: pip install "rendercheck[vision]".

Thirteen more, for defects other people keep reporting

The six above came out of one pipeline. These came from reading other people's bug reports — the same complaint, filed against every provider in turn:

The blank check is the one that needs no key. Until it existed, a still had to go through the vision tier before this tool would say anything about it at all, so rendercheck check slide.png measured nothing and exited non-zero. It reads the spread between the bottom and top of the luma distribution rather than minimum to maximum, which is what makes it hold up: a blank frame carrying one stray artifact spans the full range on min/max and still reads as blank here.

The caption check is the one with no equivalent anywhere. ffsubsync and friends correct drift; the online validators lint the .srt on its own — overlapping cues, reading speed, empty rows. Neither asks whether this caption file matches this audio, which is the only question a build needs answered. It works by matching the shape of the talking against the shape of the cues, and it reports a constant offset and a drift separately: an offset is one shift from correct, and drift is not fixable by any single shift.

Every one of these skips rather than passes when it cannot measure — no video stream, no silence structure to align against, no per-stream duration in the container. blackdetect on a .wav reports nothing, and nothing would otherwise read as "looked, all clean". That is the same trap as the regression below, and it is the reason a run where everything skipped exits non-zero.

Found on real files

Not a synthetic benchmark. Pointed at the output of a production pipeline that renders narrated video at scale:

It drew the line exactly where the fix landed, on files it was never told anything about, agreeing with a conclusion humans had reached months earlier.

That table is an anecdote, and it is labelled as one. The media is private, so nobody — including its author — can rerun it. The numbers you can check are in docs/calibration.md, produced by bench/bench.py:

python bench/bench.py sync        # no network, no corpus, about a minute

That script injects defects of known size into media with a known answer and counts what each check actually said, with abstentions in the denominator and a rule-of-three bound printed next to every zero. Where a threshold in this library has a measured provenance, that is where it came from.

Two promises

It fails open on infrastructure. No ffmpeg, no key, no network, no measurement → it warns and passes. A gate that blocks your pipeline because of its own breakage gets deleted within a week, and then it protects nothing. A defect fails closed; the checker fails open. (A missing file is your typo, not infrastructure — that still raises.)

Silence is never mistaken for success. A check that couldn't run says so, as a rendercheck.Skipped warning and in the CLI output. An empty run never reads as a clean one.

We had to earn the second one, twice.

The first cut of this library returned PASS for a file with no audio track at all — silencedetect reports nothing when there is nothing to analyse, and that read as "no silence found". It is now the loudest failure in the suite, with a regression test named after it, and the line it taught is the rule everything else follows: if we measured and it is wrong, fail closed; if we could not measure, fail open.

Then, before releasing under this name, we audited the tool against its own premise and found seven more. Every one of them reported success without having established it:

• An all-skipped run exited 0. No ffmpeg on the runner meant a green build.
• A typo'd file path exited 0, contradicting the promise two paragraphs up.
• A typo'd --script path was read as narration — one word — and produced a confident, wrong verdict about the audio: 1 WPM is below 110.
• --presenter without --known-names defaulted the roster to the assigned presenter, which made the speaker check structurally incapable of firing. It printed PASS on a script naming somebody else.
• looks_ok blamed a missing API key for every exception, so an SDK mismatch passed forever and Bedrock users were sent chasing the wrong thing.
• ffprobe failures leaked a raw Python list into the message.
• Without ffmpeg the test suite crashed on collection instead of skipping.

All seven are fixed, each with a test that fails without the fix, and the exit codes are now a contract: 0 measured and clean, 1 a defect or nothing measured, 2 a path that isn't there. Details in the changelog.

A tool that catches silent failures is worth exactly as much as its own honesty about them.

What it does not check

Being explicit, because a QA tool that implies more coverage than it has is worse than none:

• Lip sync, to the tolerance that matters. assert_lip_sync correlates picture motion against the speech envelope and catches a mouth most of half a second out of time with its voice, with no face detection and no model. Two things it cannot do. It cannot resolve the errors that actually matter perceptually — EBU R37 puts detectability at +40 ms of audio lead and −60 ms of lag, and this catches under a third of 200 ms errors. And it declines to answer on about two files in three, because on those the correlation peak does not stand clear enough to read a number off. It finds a broken pipeline; it does not certify a good one, and that still needs a model (SyncNet and friends).
• Perceptual video quality. No PSNR, SSIM, or VMAF — those need a reference encode to compare against, which generated media does not have.
• Whether the narration is correct — only how fast it's read, and whether the captions line up in time. Nothing here transcribes the audio, so a voice reading the wrong script at a reasonable pace passes everything. That check wants Whisper, which is a dependency out of proportion to the rest of this. Groundedness and factual accuracy are a different problem again, well covered by the LLM-eval tools.
• Speech intelligibility. Loudness is not clarity; a correctly-levelled track can still be mumbled.
• Music, mixing, or anything non-speech. The defaults assume spoken word.
• The rubric you didn't write. looks_ok only checks what you ask it to.

More

• Symptoms — "the TTS cut off the last sentence", "the clip came back silent", and which check catches each
• Reference and tuning — every threshold, and when to turn a check off
• Contributing — the bar for a new check
• Changelog

MIT.