By Lucas, founder of FixMyVideos · Updated
Video Plays but Has No Audio: Where the Sound Went and How to Get It Back
The picture is there and the sound is not. That one symptom has several unrelated causes: the file may hold no audio track, hold one your player refuses, or hold one that ends long before the picture. Each of these has happened on files that went through our own pipeline, with every automated check calling the file fine. Below is what was inside them and what we changed, followed by the recordings that never had sound and the files where nothing is left to recover.
First, ask the file what it holds
You need ffprobe, which comes with FFmpeg. Run this on the file that plays silent:
ffprobe -v error -show_entries stream=index,codec_type,codec_name,sample_rate,channels,duration -of compact your_file.mp4
Then count the audio packets, which tells you how much sound is really there regardless of what the header claims:
ffprobe -v error -select_streams a -count_packets -show_entries stream=codec_name,nb_read_packets -of compact your_file.mp4
Read the result like this:
- No
codec_type=audioline. There is no soundtrack in this file. If this is the output of a repair tool, the tool left it behind; the original broken file may still carry it. - An audio stream with
codec_name=mp3, or raw PCM such aspcm_s16le, inside an .mp4. The sound is probably there and Apple's players are refusing it. See the next section. (PCM inside a .mov is QuickTime's own format and plays normally.) - An audio stream whose packet count is tiny compared with the video's length. The track is a stub, or it stops early. To turn packets into seconds, divide by about 47: an AAC packet holds 1,024 samples, so at 48 kHz roughly 47 of them make one second.
When the track exists but the player stays silent
The clearest case we have measured is MP3 inside an MP4. The file is legal: the audio sample entry is mp4a and its descriptor names object type 0x6B, which the MP4 specification registers for MPEG-1 audio, MP3 included. ffprobe names it, FFmpeg decodes it, and our own playability check scored it clean with nothing to report. QuickTime on a Mac played the picture with no sound. Our founder heard the silence on a test file our pipeline had produced, before any of it reached a customer. Every check on our Linux servers had passed that file, and none of them runs Apple's playback stack.
Raw 16-bit PCM behaves similarly. FFmpeg will write it into an MP4 without a warning (it ends up tagged ipcm), and neither browsers nor QuickTime play it.
What we changed: when our pipeline rewraps a recording from another container into MP4, it now stream-copies the audio track only when it is AAC, the one codec we have heard play in that container on the players our customers use. Everything else is re-encoded to AAC, after a one-second rehearsal mux, and the finished file is probed again to confirm the audio track survived.
You can do the same thing yourself, and on an intact file it is usually all you need:
ffmpeg -i your_file.mp4 -c:v copy -c:a aac -b:a 192k fixed.mp4
The picture is copied untouched, only the sound is re-encoded. If the result plays with sound in QuickTime, you were in this case and you are done.
When the audio was there and the repair threw it away
Here the sound existed and a tool decided it did not. It helps to know how a repair tool finds audio in a file whose index is damaged.
An MP4 or MOV keeps its media in one big block (mdat) and describes it in an index (moov): where each video frame and each audio chunk starts and how long it is. When that index is gone, a carving tool walks the video frames by their own structure, and whatever sits between two video frames becomes a candidate audio chunk. That guess is only as good as the camera's layout, and cameras put other things in those gaps.
A GoPro file with its index intact. An 847 MB HEVC recording from a GoPro had a valid moov naming 3,628 AAC audio samples. Our scanner had also found 9 gaps between video frames: telemetry, padding, and one 2 MiB scan artifact. The scanner's 9 guesses were used and the index's 3,628 real entries were never read. None of the 9 sizes was a whole number of 24-bit stereo samples, so a later filter dropped them all, and the file was delivered with no sound. The rule now is that when the video table comes from a real index, the audio table and the audio format come from that same index, always. Re-run on the same file, the audio came back byte-identical to the camera's AAC, with an A/V offset of 0.0 ms measured at three points.
An Atomos ProRes recording. A 13 GB, 22-minute ProRes recording from an Atomos recorder carries 24-bit PCM. The recorder writes whole 512-byte sectors, so an audio chunk of 149,760 bytes is padded out to 150,016, and the extra 256 bytes repeat the start of the next chunk. 150,016 is not divisible by 6 (the size of one 24-bit stereo sample), so our filter judged all 2,506 real audio chunks misaligned and dropped them. Worse, the delivery then carried a 2.15-second "audio track" made of ProRes frame bytes, so the file reported audio as present. The engine now finds the padding by matching the end of each chunk against the start of the next, applies the trim only when the overlap is proven on enough pairs, and flags any delivery whose audio covers less than half the picture.
An iPhone file with no index. iPhones write per-frame metadata records next to the audio, 144 bytes each, from a Core Media metadata track. In a carved file those records land inside the audio gaps (15 of them, exactly 2,160 bytes per gap in our test file). The first thing in the way was a size cap: every iPhone audio gap was larger than the classifier allowed for AAC, so the gaps were rejected before the parser ever saw them. With the cap widened, the parser still insisted on explaining every byte, hit the metadata, gave up, and the video shipped alone. The engine now accepts gaps of that size, recognises those records and steps over them. On an iPhone 13 Pro test recording with its index stripped, it recovered 48 kHz audio that decoded cleanly and lined up within 0.045 seconds of the picture.
A screen recording that lost its first megabyte. A 129 MB, 42-minute screen recording (H.264 at 20 fps, AAC audio written by FFmpeg's encoder) had lost its first megabyte, index included. Every preview came back without sound. FFmpeg's AAC encoder opens some frames with an identification element instead of the audio data our parser expected, and it codes silence as tiny 6-byte frames that a size filter discarded. After the fix the carve held all 108,135 audio frames, and the decoded sound was identical to a reference built from the file's own surviving tables.
When the sound stops before the picture
A track can be real and short. One Windows screen recording, 3.35 GB and 81 minutes long, had audio for its first 178 seconds only: the recorder had lost its audio device and kept filming.
On a file with no index, audio is often the only clock a repair has, and this file broke our rebuild in two separate places. Dividing all 81 minutes of frames by 178 seconds of sound gave about 820 fps, a rate no camera records, so the engine threw that evidence away and would have fallen back to a 23.976 fps default, 25% too slow. Separately, the muxer timed each frame against the audio written next to it. Once the audio stopped, that clock stopped too, and our rebuild squeezed everything after it into a single audio frame, roughly 3 minutes of video at 48,000 fps that still decoded cleanly. The engine now measures the frame rate only over the stretch of picture the audio actually covers, and runs the rest at that rate with no sound. A correct repair of a file like this keeps the recorded sound, then silence, with the picture timed correctly to the end.
When there was never any sound
Timelapse and some high-frame-rate modes record no audio, and a screen recorder can capture with the microphone off. Record a few seconds with the same settings and see whether that file has an audio stream.
We had to teach the engine this too. A carved file with no audio still has bytes between its video frames, and on a few recordings those bytes were repeating filler, about 2 bits of information per byte across hundreds of megabytes. A decoder will happily turn a fixed pattern into "silence", so the gaps looked like a quiet soundtrack. The engine now measures that filler and reports no audio, rather than delivering padding as sound and promising a soundtrack that never existed.
When nothing can come back
Some files carry no recording at all, and we say so instead of selling a repair. A file whose every byte is the same value, a file that holds a container header and nothing behind it, and a file built from the same few sectors repeated end to end (one copy we examined was 20 MB long with 12 KB of distinct data) contain no picture and no sound to find. If your file is one of these, the original on the card or drive is the only hope, and copying it again through a different reader or cable is worth trying.
Getting it checked
Upload the original broken file, not the silent output of another tool: if another repair dropped the audio, the sound is only still present in the original. The analysis and a short preview are free, and you can listen to the preview before deciding anything. Recovering the full file costs €6 under 2 GB, €18 from 2 to 20 GB, €35 from 20 to 50 GB and €59 above that, VAT included. Files up to 150 GB are accepted.
FAQ
My repaired file has sound for a few minutes, then silence. Is that a bug? Compare the audio duration from ffprobe with the moment the sound stops. If the original recording lost its audio input partway, the sound genuinely ends there and no tool can add what was never written. If the original's audio should run to the end, the repair cut it short, and the original file is worth a second attempt.
ffprobe shows an audio stream but VLC and QuickTime are both silent. What now? Count the packets with the second command above. A handful of packets on a long video means the track is a stub, possibly made of bytes that were never audio, which is what our Atomos case looked like before the fix. If the count looks complete and VLC is silent too, the codec is unlikely to be the problem, because VLC plays MP3 and PCM inside an MP4. The audio data itself is suspect, and the original file is the one to check.
Why does it play with sound in VLC but not in QuickTime or on my iPhone?
VLC and FFmpeg accept more codecs inside an MP4 than Apple's players do. MP3 in an MP4 is the case we have heard play silent in QuickTime, and raw PCM written into an MP4 (tagged ipcm) is not played by QuickTime or by browsers. Re-encoding the audio to AAC, the codec Apple's players accept in an MP4, leaves the picture untouched.
Will the preview tell me whether the sound came back? The preview includes the recovered audio when there is any, so you can listen before paying. It is a few seconds taken from one point in the file, so it cannot prove the sound runs to the end; the delivered file is checked for how much of the picture its audio covers.
Is my footage safe? Files go over HTTPS to servers in the EU, are deleted automatically after 48 hours, and are never shared or used to train anything. Privacy policy.
Related
- Audio out of sync after repair
- Moov atom not found
- Atomos Ninja recording stopped
- GoPro Hero 12 truncated file
- iPhone 15 Pro MOV won't open
- Video plays at the wrong speed
Free preview · No commitment
Ready to recover your video?
Upload your file and get a free 5-second preview in about a minute. Pay only if the preview looks right.
Upload your file →