A short feature video plays with music in the editor but arrives at the website silent. The person who built the page sees a speaker button, so they assume the file carries the music. The person who exported it sees a green meter, so they assume the page simply starts muted. Both could be right about what they saw and wrong about the delivered artifact. A meter says something was active on one path; a button says a player exposes a control. Neither proves that audible program sound reached the file the visitor receives.
The quickest repair starts with a smaller question: at which handoff did the evidence change? We can answer that without turning the volume up, regenerating speech, rerendering the project, or overwriting the first export. This article is about diagnosis, not a substitute for listening to the final piece. Its example assets and result values are synthetic teaching examples.
Put a name on the version before deciding it is broken
A media workflow can produce a source take, editor timeline, preview render, master export, uploadable rendition, player-selected source, and local cached copy. People call all seven “the video.” A report that says “the video has audio” is nearly meaningless until it identifies a version. Begin by recording the expected asset, observed asset, time offset from the start, and playback context.
Suppose an editor hears an eventual music cue at 00:08 in Master 07. A colleague checks the web player at 00:03 on Mobile Cut 06. Another opens a copy downloaded from chat that was automatically altered during sharing. There is no mystery yet; they have not tested the same thing under the same conditions. Ask for an exact filename, visible release identifier, or hash when both people have the same file. Identify the loaded rendition when a website switches among formats or resolutions. A poster, title, and duration can be the same across variants.
Write down which sound is supposed to appear at which cue. A logo sting, low background atmosphere, and narration may live in different stems. “No music at 00:08” is not necessarily “no narration at 00:19.” If the expected sequence begins with intentionally quiet material, define an interval with recognizable sound before opening an analysis tool. Nobody should diagnose five minutes of program audio from a silent half-second at the beginning.
Keep the first failed output. It gives you something to compare when a revised export starts working. Save a second copy for experiments instead of fixing the only deliverable and then trying to remember the original state. Working with familiar fixtures, one intended cue and one known quiet section is a good preliminary check. Stay at a comfortable playback level, especially when the next test may unexpectedly contain normal-volume audio.
Ask four different questions
- Is there an audio stream in this asset? The container may have none, one, or multiple. A stream inventory answers this without assessing whether its samples are useful.
- Does that stream contain the intended signal where it should? A valid stream can contain actual silence, unwanted noise, a countdown, the wrong take, or very quiet samples.
- Can the listener receive the right stream and channels? A specific player may select a silent alternative; narration could be confined to one channel or cancel when combined.
- Is this playback context allowed to make sound? A user setting, browser state, app tab, muted device, distant scene mix, or missing user gesture can suppress a perfectly good file.
These are separate tests, not four causes to try fixing simultaneously. Start with the one that requires the fewest changes. If the colleague who reported the bug can hear a known-good clip in the same player, you have evidence about the player and output route. If both files appear quiet there, test the failing file in an independent local player and also verify a known-good file there. These two comparisons tell you more than trying yet another export preset.
Inspect the asset before trying to repair it
The ffprobe documentation describes how to list streams and select audio streams for inspection. On an authorized local copy, a first read-only check can look like this:
ffprobe -v error -select_streams a -show_streams -of json "candidate.mp4"
This command asks for information on audio streams, not a conversion. Read any errors along with the output. An empty stream list after successful probing is not the same thing as an unreadable file, an unrecognized input, or a typo in the path. If there are tracks, note their order, codec, declared channel configuration when available, tags, and disposition fields for troubleshooting. Treat metadata such as the title “Mix” as a label, not an assessment of the actual signal.
Run the same check on the preceding known-good artifact. If the master has audio and the downloaded final does not, the evidence points between those two nodes. Compare the export or rendition mapping; do not spend time stripping HTML controls. If both outputs have audio streams, move to signal checks. Audio presence may help explain silence but does not establish why a container was assembled with or without the desired stream.
If you redeploy an asset after checking it, verify the actual delivered file or media segment. A local fix followed by a public path serving an old cached rendition is not a successful repair. Conversely, don't call a web copy stale solely because its byte hash differs from the source; a site may legitimately package or transcode it. Where controlled expectation includes a transformation, compare its meaningful output and identifiers, not a hash that was never expected to match.
Read the shape of the signal, not only its existence
The FFmpeg filter documentation describes silence detection based on a level threshold and duration. A quick non-writing test, using an invented sample threshold suited only to demonstrating the technique, is:
ffmpeg -hide_banner -i "candidate.mp4" -map 0:a:0 -af "silencedetect=noise=-45dB:d=0.5" -f null -
The explicit map here selects the first audio stream of the first input. It does not mean that stream is the intended track. In a multi-stream file, inspect each plausible candidate deliberately. The null output discards the processed media; the filter's logged observations are the output you want. A log should be interpreted against the actual timeline you inspected, especially if you used trims or files whose streams start at different times.
At a chosen threshold, detection can flag quiet spans that are intentional. A spoken sentence might have part of a word below threshold, atmosphere might sit below a value chosen for speech, and noise may occupy a section where intended dialogue is entirely absent. Non-silence proves there was some signal by the criterion, not that the right sound occurred, that it was comfortably intelligible, or that it meets a platform's loudness specification. Use short focused listening at the expected cues to distinguish these cases.
Consider a quiet-room narrative set in an imagined demo. The intended first signal is a door click at 00:05, then a voice at 00:16. The filter flags a quiet interval from 00:00 to 00:04 and none around the voice passage. That may be correct; the listener has to confirm the click and actual voice. Another file shows non-silent material throughout because of ambience, but the voice take was never routed to the exported mix. Replacing the whole project with a “louder” version would make the ambience louder while keeping the bug.
When the channels disagree
Audio can be carried on distinct streams and on distinct channels within a stream. A stereo stream might carry left and right; a file could carry two stereo streams for different alternatives. Do not treat those arrangements as interchangeable. A voice recorded on one channel might disappear when a downstream mapping explicitly picks the other. A stereo sound that is distinct in headphones may become weak when combined by a device that plays a mono mix. Neither symptom tells you that the source file lacks audio.
Use a monitor or editor that can inspect and isolate available channels without altering the original. Check the same cue on each relevant path. If left contains speech and right only background, decide whether the source arrangement is intentional or a routing mistake. If the guide sound appears in an alternative stream, the problem may be track selection, not mixing. If you suspect cancellation, compare stereo and the actual intended mono fold-down instead of guessing that any difference has one particular cause.
FFmpeg's main documentation describes automatic selection and explicit mapping. Selection defaults are not the same as acceptance criteria. When handing off an export, specify which program track should survive and why. If there is no deliberate alternate, a delivery artifact with one unambiguous program track is easier to test than one where several tracks compete for default playback. If alternatives are required, verify each label, language or role as well as the initial default on relevant players.
An anti-pattern is to add the same audio stream twice because a prior export was silent. That can produce two selectable alternatives, not two channels of the same mix. Another is to duplicate a mono signal into stereo before understanding why it vanished. That may paper over one player's selection and create confusing assets later. Diagnose the issue at the handoff and document the smallest justified correction.
Distinguish a webpage state from the audio content
The HTML Standard's media section defines media controls and properties such as muted state and volume. In an ordinary web player, sound can be muted even while playback proceeds. Starting muted prior to a user's explicit action can be an intentional interface behavior, not a missing export. Showing a speaker switch does not prove the switch changes the state of the playback element actually in use.
Check manual playback using the control the audience sees. Does the interaction begin playback, turn sound on, or do both? Is the displayed speaker state consistent with the media element acting as the sound source? Did a second scene transition replace the element and restore its defaults? A visible waveform generated elsewhere is not evidence. Measure state around the gesture, expected cue and transition. If a separate synchronized music element supplies the sound, inspect it independently from the picture.
A useful contrast is local player audibility versus manual website playback versus muted autoplay. If the fail occurs only before an explicit interaction and behavior fits design, clarify the control instead of forcing autoplay. If clicking a speaker icon never changes the actual player state, repair the wiring. If the earlier file passes manual web playback and the new one doesn't, identify relevant format differences or asset mismatches. There may be no need to change player code until the cause has been narrowed.
Browser policies and device behavior matter; do not assume one device's allowed playback or output route matches another. Record enough context to let a colleague reproduce: browser and device, source selected, whether the page had received a user gesture, what the controls displayed, and which offset should be audible. If there is a mute button in both your editor and website, specify which one was used.
Follow one silent export through its handoff
In an invented workflow, Production Master 07 and Web Render 07 have the same visible footage and length. A later upload produces a second device rendition. The bug report says the sound effect behind a title is missing. The editor places that effect at 00:12 in the project and hears it, though the recording only matters on the exported artifact. The team makes a three-row evidence register.
- Master 07: audio stream present, the effect audible at 00:12 in independent local playback.
- Web Render 07: audio stream present, but sample inspection finds the same cue silent and the other voice cue audible at 00:22.
- Device rendition: sound check inconclusive until manually unmuted; after interaction, voice cue audible and effect absent.
This does not support “mobile audio broken.” It supports “effect absent by the first Web Render 07 inspection while voice was preserved.” The rendering configuration now earns a closer look. Maybe the project used a group bus included in the master but omitted from the web export. Maybe the sound effect was disabled in the particular sequence selected for web rendering. Maybe the earlier mastering session made an intended last-minute change that wasn't copied to the web sequence. From the three rows alone, those are hypotheses, not confirmed causes.
The appropriate next test is a fresh export to a new local name with exactly one identified routing difference checked. If it restores the effect while preserving the other cue, then test its device rendition after the authorized handoff. Block the repair if it silently swaps soundtrack versions or changes timing that the editor intended. Releasing acceptable audio matters more than getting an arbitrary inspector meter to move.
Put the handoff claim on a receipt
One useful delivery receipt says which artifact includes what, not which application was used. For an invented review export it might specify “first audio stream is the general program mix; alternate audio absent by design; room tone audible at 00:04; voice starts around 00:19; a 1.4-second gap precedes the closing card.” Intentional silence should get a cue too when its absence might look like a failure. If you can name the expected behavior, a reviewer can look for it without the editor open.
Choose whether the check evaluates media as requested or media as downloaded. An internal editor may judge an exact named file. A public reviewer may see an adaptive delivery with a separate audio rendition. Conflicting results are not a reason to ignore either receipt; they tell you where reproductions diverged. Make the issue owner follow the divergence rather than requesting a screen recording that could have its own muted microphone or mixed output.
Keep observations and repair guesses in separate columns. A receipt can say “local master showed two candidate tracks, Web Render 07 only one, cue missing around 00:12.” A hypothesis can say “second track may have been omitted by the web preset.” Report that distinction in the handoff so the next person can check the preset without treating an untested root cause as settled. If you made two changes at once, say so. An audible output proves the outcome improved, not which change supplied the fix.
Decide whether hearing, signal, and track lists are telling the same story
A troubleshooting call often begins “the video has audio,” but the phrase may mean the interface displays a speaker icon, somebody heard a previous version, or a container inspector found a stream. Those are three evidence types. A stream can hold audio samples whose identifiable content is outside the section played, in a different channel than expected or quiet enough to be hard to hear. A file may be valid while the web player does not start audible playback in the attempted state. Put each assertion into a row with a named artifact, playback location and observable result rather than merging them into one diagnosis.
Use the expected role of the clip as your target. For a video with an introduction followed by deliberate silence, testing an arbitrary first second may be the wrong acceptance criterion. A clip edited to use voice on one segment and a sound bed later needs checks covering both. To design a synthetic handoff, put a safe, expected label on each region: brief spoken line, deliberate pause, second spoken line and an ending without sound. The issue is not whether the media inspector can find nonzero samples somewhere. It is whether a listener gets the intended material at useful places in the exact delivered version.
Record what reproduction actually means in the report. “Asset Q in local Player A at 0:04–0:07; no apparent sound while visually playing” is an excellent observation. “The whole export lost the audio” is stronger and may be impossible to establish from that test. Likewise a screen recording can prove which user interface was visible without preserving the audio route and state you heard. If permitted, include a short written replay recipe that allows others to inspect the exact issue independently. Don't assume a video of a muted player itself resolves whether the downloadable file carries the story soundtrack.
To keep the work accessible, collect conditions needed for the real audience rather than reporting “I heard it” without context. Output device, channel placement and controls may matter, and a brief ordinary clip should be verified at a comfortable listening level. Do not compensate for uncertainty by turning up headphones or overriding device safety settings. Tools that describe samples or simply playing a benign test at normal levels can establish useful comparisons without exposing someone to a potentially loud passage. The first task is to identify the expectation and the observed location of the discrepancy.
Trace stable properties along a small handoff ledger
A useful ledger can separate source, edited working copy, encoded deliverable, version uploaded and version served. Define each by an unambiguous local reference or a legitimate provided identifier. If two files have the same name but one is an older proxy, a support question that says “play final.mp4” can lead different viewers to different conclusions. A hash on exact local files can distinguish their bytes when such use is appropriate, but two legitimate repackagings may also have different hashes. The checksum is a precise answer to a narrow question, not evidence that the soundtrack is the same after all processing.
Write expected audio alongside each step. The source may have several streams for different roles. The editor's preview may show selected tracks on a timeline, while the output intentionally includes one mixed stream. An online service may then provide its own rendition. A count of streams from a step can help identify where the question begins but may not explain the selection policy. The FFmpeg stream-selection documentation is relevant when FFmpeg and its choice rules actually apply; it does not identify what an unrelated editor or hosting service chose. Ask which version, tool and settings were active when obtaining the relevant evidence.
For a toy problem, imagine a file at each stage with one voice track until the hosted page supplies two selectable audio choices, only one of which carries the expected greeting. Opening the original export and hearing the greeting helps localize the report, but does not determine why the hosted page presents another choice or which version of it the application played. Make a separate row for the web selection and the access route under test. If you can compare the intended public deliverable with the local export through authorized means, record the exact comparison. If you can't establish which version was served, leave the corresponding conclusion open rather than treating the version the studio currently opens as definitive.
A bug can also reside in the test set itself. If only one file at an unknown version silently fails, it may not be a suitable fixture for diagnosing all player states. Keep a small known-good synthetic media sample with expected behavior, a separate true-no-stream example and a sample that intentionally contains quiet sections. A successful check on the known-good clip in the same conditions helps distinguish a general inability to play from an asset-specific difference. It cannot prove that every codec, device or browser in the target audience behaves identically. Record the target matrix needed for the project instead of implying that one local success settles deployment.
Leave the samples alone while checking whether timing can explain the symptom
A clip with a present stream can produce the symptom if the useful passage is outside the short section examined. Start by checking actual duration and intended positions against the exact asset. The ffprobe manual documents representations for streams and related metadata; read the tool output in the context of the file and command that produced it. A listed duration is not itself evidence that an intended sentence was preserved. And a passage at a time in the original source need not occur at the same place in the edited version. A clear test can use a small set of expected segments rather than rely on one general assumption about the timeline.
Suppose a fictional language lesson has a deliberately quiet two-second title card, then a spoken greeting and two later phrases. The editor tests only the title card and thinks the audio is gone. A reviewer scrubs far enough to hear the greeting on a normal device, but never tries the last phrase that prompted the complaint. Both have partial observations, and neither was wrong about the moment they examined. A better worksheet lists several time-bounded checks, what should happen at each, which actually played, and what was skipped. It does not need a full transcript or language interpretation in order to identify the section that is silent unexpectedly.
A smaller timing error can come from how the comparison was set up. If the user begins playing from a persisted position near the ending, the next couple of seconds may legitimately have no voice. If a loop fixture omits the first second, an intended short cue might be elsewhere. If different versions include different trims, the reported time may not index the same picture segment. In a local exercise, put an ordinary visible event beside the expected sample, such as a synthetic static caption indicating intended sound at that moment. That gives reviewers a check more useful than assuming that two timelines are equivalent at a shared digit.
Quiet detection output also needs a documented configuration. The FFmpeg silence-detection documentation describes threshold-based behavior. An output interval reports behavior under the relevant threshold, duration and input conditions; it is not a universal statement about speech, quality or what a given listener heard. A faint but intentional passage could satisfy one selected threshold. A loud unintended signal can fail a silence test and still be unusable for the listener. Preserve exact source, analysis settings and passage context if passing the report to another person. Do not normalize or alter the only original to make a test easier to interpret.
Test one axis at a time before blaming a player or a channel
A clear diagnosis often comes from a two-by-two table. Compare a documented synthetic known-good file and the affected file in the same local viewing conditions, then compare each through the authorized playback route of interest where applicable. Do not substitute a different output device, browser state, and version simultaneously and then credit any one of them for the change. A test matrix is a compact record of what combinations were actually tried. Unknown should be an available cell; filling a grid can make comparisons look unnecessarily broad.
A channel discrepancy can benefit from the same approach. For instance, if program material appears separately on two channels but the listening route combines or selects them differently, record evidence at the relevant stages before editing. One familiar case is a synthetic left and right test signal that can be heard separately but behaves differently under a documented combined route. That can illustrate why an audio track that is not silent under inspection may not create the expected listening result. It does not establish what a particular app does by default or that combining all audio into one channel is a good fix. If the release depends on a certain program layout, preserve that requirement and seek the appropriate procedure for changing it.
The browser case has a related division. The WHATWG media specification defines several states and controls; it does not promise that every app embeds the same player UI or that a browser will start audible playback under every condition. Document what happened before and after normal intended user interaction. A muted initial presentation, a locked volume control at some upstream stage, or an asset delivered with silent samples may all look alike to a casual screen recording. Verify the intended control path without working around security settings or inadvertently creating unsolicited audio behavior.
For an invented kiosk test, the sample has a clear spoken line and a proper page-level play control. A reviewer observes silent initial movement before interaction, followed by audible playback on the expected press. That result can satisfy the documented experience if silent initial movement is intended. If the complaint was lack of audio even after explicitly starting playback, the initial-state result doesn't answer it. Keep the sequence of actions and responses next to the expected behavior so users, frontend staff and media editors can agree on which condition is unfinished.
Hand over competing explanations rather than a guessed repair
A short evidence receipt should present what is established, what was ruled out in the tested context, and the next narrow comparison. For example, “local deliverable shows one stream; the defined speech passage is audible locally; the authorized playback test did not produce expected audio after ordinary interaction; served rendition not yet confirmed.” Such a handoff does not say the underlying export is missing its stream, nor does it seize authority to edit or republish on someone else's behalf. It identifies which comparison the owner of the next step would need. This is especially useful when several teams interact with the same name and not the same file.
A proposed change can go in a distinct field with a before-and-after acceptance test. If a different track should be selected during the authorized export, define which program roles the new output should preserve. If the playback experience needs revision, define how a user can start or control audio safely in the intended context. If the re-export is a separate action, do not call it complete until the artifact was created and correctly verified. A meeting decision and an observed successful test are different checkpoints. Keep older artifacts identified so a later branch of the conversation doesn't accidentally repeat the first failing test on stale media.
For a final tabletop drill, hand a teammate three fictional receipts with the same symptom at the top. One has no track in a precisely identified export. Another has an audio stream and a quiet time range before the expected passage. The third names an audible local passage and a still-unresolved delivery context. Ask the teammate to propose the next check for each without editing any source. If their first step is always to amplify, re-export or unmute, the point of the record hasn't travelled. The work is more useful when it teaches the person opening it how to make the question smaller before doing anything that could erase the evidence.
A release checklist can finish with one beneficiary-facing question. Did a listener using the intended ordinary flow hear the expected segment at a comfortable playback level, under the conditions that were actually tested? If the only result is transport-level existence of streamed data, report only that. If the release has verified constraints, report them rather than claiming every viewer is covered. That lets the next support person know which parts of the experience have been seen and what remains uncertain, which is often the fastest way to solve the actual problem.
Preserve a low-risk check for where the symptom starts
An efficient first run may only need two observations. Play one benign, known-good fixture along the expected listening route at a normal level. Then listen to a specific anticipated moment of the suspect artifact, under the same authorized conditions. This isn't a definitive test of a complex playback system, but it can tell you whether the next comparison should be a local artifact inspection, a playback condition check, or a version match. Write down which of those cases you actually reached. A sequence of sound checks can seem much less conclusive once you notice that each file used a different device or different start position.
A common failed handoff gives someone only a screen recording where a sound control is visible. The control can confirm an interaction took place and still fail to say which selected output the listener heard, or what the exported samples contained. Keep the sequence evidence beside a narrowly described audible outcome. If the project needs stronger verification, assign a colleague an approved setup and recognizable synthetic cue rather than asking them to reconstruct the first listener's machine. A useful outcome is a reproducible case in which both people know exact version, timestamp and requested operation, even when the diagnosis remains unknown.
Sometimes a passage has sound, but not the expected sound. A short incidental noise does not satisfy a speech proof simply by crossing a sample threshold. In a synthetic training sample, put a normal gentle cue at one position and a known phrase elsewhere. Have the reviewer name which one was audible. If they write “audio present,” ask whether both intended moments were preserved in the relevant rendition. A brief label can convey the expected outcome without including unrelated recordings in tickets or forcing someone to quote private material. The distinction between a valid signal and a correct program is precisely where several repair proposals can diverge.
Finally, decide whether a repair test can be made on a separate comparison artifact. An authorized quick local trial can keep the original available, carry forward the known intended track layout, and mark the result as a test rather than a replacement. The receipt can then compare the original and test directly, retaining failures that matter rather than losing them in a filename called final. If no safe equivalent is available, retain the unresolved evidence and describe what specific conditions need a controlled test. The strongest diagnostic step is often the one that improves the question without changing anything else about the work.
Test the edges that matter to the listener
A one-cue reproduction is ideal for finding the cause and not sufficient for final acceptance. Write a tiny cue sheet with representative events: intended onset, longest quiet interval, key speech entry, transition out of the scene, and ending tail. Audio can be cut short or shifted after the opening passes. For a silent-by-design clip, verify that no accidental track plays and that the UI doesn't present a sound switch which promises something unavailable.
Check at least one device or output type appropriate to the audience when hands-on testing is possible. If a headphone-only test is all you did, say so. Test manual pause, resume and reset if the project exposes those controls; sound can start correctly and become wrong after a seek or scene change. A timing chart from the file alone does not establish how a reactive player behaves while moving among scenes.
Distinguish “audio now present” from “final mix approved.” The first verifies that a track exists and the intended event is audible. The second may involve relative levels, tonal balance, perceived clarity, duplication, looping, accessibility alternatives, or use rights outside the technical check. Preserve the receipt with a named file, offset and playback context. If you share a bug exhibit, use synthetic or authorized content, not an unrelated customer's original media.
Start with the earliest known-good version and the first known-bad one. If they aren't genuinely adjacent, shrink the gap. The feature clip doesn't need a bigger mix until you know it has the right mix. Your best result is a short, reproducible explanation: this track disappeared at this handoff, these samples survived, or this player muted the correct file until this action.
AI disclosure: This draft was produced with AI assistance; examples are illustrative, not records of actual work.