Rendercheck README
The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Rendercheck listing page.
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.
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:
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:
…and one of the two added in 0.3.0:
Then point it at your own output:
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:
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:
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:
"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:
| The measurement | Already available from |
|---|---|
| Integrated loudness, true peak | pyloudnorm, ffmpeg's loudnorm |
| Silence detection | pydub.silence, ffmpeg's silencedetect |
| Duration, stream layout, frame rate | ffprobe |
| Black frames, freezes | ffmpeg's blackdetect, freezedetect |
| Caption↔audio offset | ffsubsync — which corrects it |
| Container and codec conformance | MediaConch — policy-driven, pass/fail, from the CLI |
| Speaker identity | resemblyzer |
| Video quality metrics | VMAF, ffmpeg-quality-metrics |
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:
Node, Remotion, anything that renders in a build step:
Docker, if you would rather not have a Python toolchain at all — ffmpeg is already in the image:
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:
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:
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.
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.
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.
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.
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.
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.
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:
| Check | The defect |
|---|---|
assert_no_truncation | Speech that stops mid-sentence while the API returns success. The single most-reported defect in generated audio; measured against the file's own average, so it holds for quiet and loud content alike. |
assert_has_sound | A clip that comes back silent — an upscale step drops the audio track, a mux points at the wrong stream, a synthesis writes zeroes. |
assert_no_clipping | A gain stage pushed the waveform past full scale. Crackles on consonants, and turning it down afterwards does not undo it. |
assert_true_peak | Measures clean locally, distorts after upload. Loudness and peak are different problems: a lossy codec reconstructs the waveform between samples, and clips where it goes over. |
assert_no_black_frames | Generated video truncating to black instead of erroring: right length, valid container, nothing in the last third. |
assert_not_frozen | The picture stops moving. Every frame present, every frame the same frame. |
assert_captions_aligned | Captions written against one clock, audio rendered against another. Every line arrives at the wrong moment, and both files are individually perfect. |
assert_streams_aligned | Sound and picture that do not cover the same stretch of time — a mux that ran out of one input, or a concatenation that mistimed its first segment. |
assert_format | A render that quietly fell back to 720p, came out at the wrong frame rate, or is variable-rate where the pipeline downstream assumes constant. |
assert_not_blank | An image generator that failed and returned an empty canvas — reported against DALL·E, Stable Diffusion, Qwen, Gemini and Krita, always the same way: no error, no warning, correct dimensions, nothing on it. Catches any flat canvas, not only a black one; blackdetect sees none of the others. |
assert_loudness_range | A file with no single workable volume setting: turned up for the quiet passages, the loud ones startle. Different question from assert_loudness, which only asks where the middle sits. |
assert_audio_format | Mono delivered where stereo was specified, or 44.1 kHz where the spec says 48 — resampled downstream by whichever converter happens to be in the chain. |
assert_lip_sync | A talking head whose mouth is out of time with its voice. The container is correct, the streams are valid, the durations agree, and assert_streams_aligned passes — because that reads the container's timing and this reads the content's. Off by default, and it declines to answer more often than it answers: see the note under What it does not check. |
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:
| result | |
|---|---|
| A course known to be good | clean — 162 WPM, −14.1 LUFS, no dead air |
| Content re-rendered after a loudness fix landed | passes at −14.2 LUFS |
| Four episodes rendered before that fix | fails at −19.4 to −21.3 LUFS |
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:
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
--scriptpath was read as narration — one word — and produced a confident, wrong verdict about the audio:1 WPM is below 110. --presenterwithout--known-namesdefaulted the roster to the assigned presenter, which made the speaker check structurally incapable of firing. It printedPASSon a script naming somebody else.looks_okblamed 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_synccorrelates 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_okonly 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.