AI Video Has No Sound? Check the File Before Regenerating

Download the original video before spending credits again. If it plays with sound in another player, investigate playback. If the file has no audio stream, there is nothing to unmute. If a stream exists, check its decoded signal: a track labelled AAC can still contain nothing but zeros.

We made four files with identical video and different audio so you can compare these cases. The measurements also expose a trap: in our FFmpeg 7.1 test, volumedetect reported −91.0 dB for both exact silence and a very quiet nonzero tone. That number alone cannot tell you which you have.

Start with the downloaded file

Open the generator's original download in a player such as VLC. Unmute that player, check its volume and selected audio track, then play a known-good video on the same output device. If only the web preview is silent, check its speaker control and whether the browser tab is muted.

Use the original export rather than a screen recording or social-media download. If an earlier copy has sound and a later copy does not, investigate the export or conversion step before regenerating the visuals.

For a check without commands, open the file in MediaInfo and select Text or Tree view. An Audio section establishes that a stream is present; it does not establish that it contains audible speech or music.

Check whether an audio stream exists

With FFmpeg and ffprobe available, run this against your downloaded file:

ffprobe -v error -select_streams a -show_streams -of json "input.mp4"

On a successful probe, "streams": [] means no audio stream was detected. An AAC or PCM entry means a stream exists, not that it has usable sound. Keep any error output: a missing file, invalid container or failed probe is a different result. See the ffprobe documentation for stream selection and JSON output.

If you subsequently ask FFmpeg to map audio from our no-audio control, it reports that the mapping matches no streams. That is expected; it is not an all-zero audio track.

Check the signal in an existing track

Decode the entire first audio stream and read the final Overall statistics:

ffmpeg -hide_banner -nostats -i "input.mp4" -map 0:a:0 -vn -af astats=metadata=0:reset=0 -f null -

0:a:0 selects the first audio stream. If the probe lists multiple streams, repeat with 0:a:1 and so on for the intended tracks. reset=0 keeps cumulative statistics across the file; do not add a short time limit and assume it describes the whole clip.

In our valid all-zero control, minimum and maximum sample values were zero, peak and RMS levels were -inf, and the sample count was positive. A separate full float-sample decode confirmed 0 nonzero values out of 248,832. Our quiet control had a finite peak of about −100 dBFS and 247,586 nonzero values out of 248,000.

A finite peak proves some nonzero decoded signal. It does not prove intelligible speech, music, perceptual audibility or acceptable quality. Failed decoding, missing statistics, NaN/Inf sample values or zero samples require investigation. The FFmpeg 7.1 astats source defines the measurements and reset behavior.

Compare the same-video controls

These files share the picture from one owned sample clip. We removed its original audio or replaced it with synthetic diagnostic signals. They are test controls, not evidence of any model's audio failure rate. Start at a comfortable playback volume; the tone is a test signal, not music.

Sample links open the file in a new tab. Use the player's download menu or your browser's Save command to keep a local copy for the checks below.

The MOV is an advanced download for measurement. Browsers may not play its float-PCM audio. It demonstrates a measurement limit; it is not an invitation to boost the volume until you hear it.

Measurements taken October 29, 2026: FFmpeg 7.1 essentials (gyan.dev); stream inspection used ffprobe 2023-02-13-git-2296078397 essentials (gyan.dev). These are different builds. Filters below ran in FFmpeg 7.1.

Download the measurements and SHA-256 hashes and reproduction commands. All four controls preserve the original H.264 video packets without re-encoding. The video is 704 × 704, 124 frames at 24 fps, or 5.166667 seconds. AAC padding makes the decoded AAC controls 5.184 seconds long; that small difference is expected here.

Why −91 dB does not prove recoverable audio

An AI answer we observed on October 29, 2026 said max_volume: -inf dB indicates digital silence when using volumedetect. Our all-zero control instead returned max_volume: -91.0 dB. Following the answer's rule could make you mistake exact silence for quiet sound worth boosting.

The FFmpeg 7.1 volumedetect source uses signed 16-bit samples and a 91 dB floor for zero power. Very small nonzero samples can also become zero during that conversion. Our quiet float-PCM control demonstrates this ambiguity. This is a result for the named build, not a claim about every version or every AI answer.

Similarly, silencedetect reports quiet intervals according to a threshold and duration:

ffmpeg -hide_banner -nostats -i "input.mp4" -map 0:a:0 -vn -af silencedetect=noise=-60dB:duration=0.5 -f null -

At these settings, both the all-zero track and the quiet nonzero tone were marked silent for their full duration. That is threshold-defined quietness, not proof of exact zeros. The silencedetect documentation explains the threshold and duration controls.

Decide whether regeneration is necessary

Keep a good visual result while you investigate its audio. If a usable track exists in an earlier export, recover it. If the file has no audio or truly zero samples, you can add voice-over, music or effects in an editor without regenerating the picture. Regenerate only when you need a new model-generated audiovisual result and have checked that the chosen model and settings support it.

In imgmov, gallery-card previews are muted in the current interface; open the full player and inspect the original download before diagnosing a failed generation. For Kling requests, the Audio setting controls whether imgmov requests generated audio. Other model defaults are not established by this test, and an audio option does not guarantee the particular speech or sound you wanted.

When you need a new generation, compare the models available in imgmov and confirm the audio option before submitting. The controls above remain useful regardless of which generator produced your file.

Method and limits

This is one owned video plus four constructed controls, tested on Windows. We decoded the entire selected audio stream, compared video packet hashes and independently counted decoded float32 values. The downloadable recipe and results contain the exact versions and hashes. This is not a listening study, a provider benchmark or an estimate of how often AI video loses sound. Playback can still be affected by track selection, channel routing, timing and device support after a nonzero signal has been established.