Why Does a Separately Delivered Narration File Slowly Lose Alignment After Audio Sample-Rate Conversion?

Check duration before investigating frame rate

Audio duration is related to the number of samples and the rate at which they are played. For a simple continuous PCM recording, duration in seconds equals sample count divided by samples per second.

For example, 480,000 samples at 48,000 samples per second represent ten seconds. Interpreting those same samples at 44,100 samples per second would produce approximately 10.884 seconds. That is a rate-interpretation error, not a normal duration-preserving resample.

A proper resampling operation changes the number of samples to represent approximately the same duration at the new rate, subject to rounding and processing details.

Start by comparing the original and converted narration's actual duration. A discrepancy that accumulates toward the end may indicate a clock or interpretation problem, but it does not establish one specific cause.

Separate rate conversion from timestamp correction

FFmpeg documents aresample as a resampling filter. It can also stretch or squeeze audio to match timestamps, or insert or remove samples, depending on the selected options.

Those are distinct operations. Converting a nominal sample rate is not identical to compensating for timing discontinuities or mismatched timestamps.

Record the original file's sample rate, sample count where available, start time and duration. Record the same properties for the converted file and the settings used to produce it.

If the new file has the intended sample rate and nearly the same duration, compare the actual audible events before assuming sample-rate conversion caused the drift. If its duration changes in proportion to a rate mismatch, investigate whether the samples were relabeled or reinterpreted rather than resampled.

If the original recording's audio clock differs from another device's video clock, a technically correct nominal-rate conversion alone may not align them.

Hypothetical ten-second narration fixture

Hypothetical illustration—not an executed conversion: a narration recording contains 480,000 PCM samples at 48 kHz, giving a nominal duration of ten seconds.

Version A is a proper conversion to 44.1 kHz, expected to contain approximately 441,000 samples for the same ten-second duration.

Version B retains 480,000 samples but is interpreted at 44.1 kHz. Its calculated duration is approximately 10.884 seconds.

If Version B is aligned at its beginning against a ten-second video, its end will be late by approximately 0.884 seconds. This is mathematical illustration, not a measured drift from a recording.

Verify the source and converted clocks

Compare identifiable spoken events near the beginning, middle and end of both audio files. Use a consistent definition of signed offset and keep the video source unchanged.

If the audio conversion is the first stage with changed duration, repair that conversion from the untouched source. If both audio versions have the same duration but drift against a separate camera recording, investigate capture-clock differences and the video timing independently.

Do not change playback speed merely to make one endpoint line up without checking pronunciation and sentence duration.

The existing phone-VFR guide locates timing faults across source, editor and export. This guide isolates audio sample representation and resampling, which can fail even when the video has constant frame timing.

Record your own observations

The filled rows are illustrative examples. Use the blank row for your own files and observations.

Download the editable CSV worksheet

What the references establish

• FFmpeg Filters Documentation — aresample: Documents audio resampling and the separate ability to compensate for timestamps by stretching, squeezing, inserting or removing samples.
• FFmpeg Resampler Documentation: Documents resampler configuration and sample-rate conversion parameters.