Why is my recording quiet with high peaks?
ChatGPT
Claude
Google AIA quiet recording can still have high peaks. One brief sound may use most of the available headroom while the rest of the voice stays much lower. Before turning up the whole file, check the peak, the average level and the passage being measured. More gain raises all of them, including background sound.
Peak and RMS answer different questions
The sample peak is the largest absolute sample value in the excerpt. It tells you how close that sample is to digital full scale, not how consistently audible the speech is. Whole-excerpt RMS describes the root mean square of the samples across the selection. Pauses are included, so adding a long quiet section can lower the reading without changing the spoken words. Audacity's Measure RMS documentation explains why the selection matters.
Neither number is the sound pressure at your ears. If the file measures normally but sounds quiet on one device, keep the file unchanged and check the output volume and listening path. The difference between dBFS and dB SPL helps separate those problems. RMS is also not integrated LUFS, which uses a different measurement method. Audacity distinguishes the two in its loudness normalization reference.
Can one gain change meet both limits?
Imagine a mono excerpt with a sample peak of −2.8 dBFS and RMS of −24.5 dBFS. Suppose a particular job calls for RMS of at least −23 dBFS and sample peaks no higher than −3 dBFS. These are illustrative limits, not a universal voice target.
| Question | Calculation | Result |
|---|---|---|
| Gain needed for the RMS minimum | −23 − (−24.5) | +1.5 dB |
| Peak after that gain | −2.8 + 1.5 | −1.3 dBFS, above the selected ceiling |
| Gain allowed by the peak ceiling | −3 − (−2.8) | At most −0.2 dB |
No single gain change meets both limits here. Reducing the file by 0.2 dB addresses the peak but lowers RMS further. Raising it by 1.5 dB addresses the RMS minimum but exceeds the peak ceiling. Peak normalization does not remove this conflict. It applies gain to reach a chosen peak, as described in Audacity's Normalize reference.
Measure a representative excerpt
- Keep an untouched original. Export a copy containing normal speech and a louder phrase, not just silence or the loudest word. Keep its start and end points for the later comparison.
- Check the copy locally. Open WaveLocus recording review. It accepts up to three supported mono or stereo excerpts, each up to 60 seconds and 16 MiB. Audio and filenames stay on your device. The local check does not need a ChatGPT connection.
- Read each channel separately. Note the sample peak, RMS and full-scale sample count. Keep stereo channels linked unless changing their balance is part of the task.
- Listen before processing. Open “Locate and listen to a passage” in the findings and choose a peak or full-scale group. Keep device volume low, then press Play for a short, attenuated preview. Edge fades can change boundary clicks, so inspect the original in your editor for an exact check. Decide whether the sound is wanted speech, an isolated disturbance or something to investigate further. A number cannot make that listening decision.
If the concern happens during capture rather than in an exported file, use the microphone test to check the selected input. Do not raise input gain simply to force a finished-file reference onto live microphone readings.
Choose the next action
| Observed result | Useful next check |
|---|---|
| Low RMS with room below the chosen peak ceiling | Consider a bounded gain change on a copy, then listen and measure again. |
| Low RMS with little or no peak headroom | Inspect loud passages and selection boundaries before adding gain. |
| Long pauses dominate the excerpt | Also measure representative speech. Label the different selection rather than calling it an improvement. |
| Full-scale samples are detected | Inspect the original and capture chain. Turning the file down cannot reconstruct lost information. |
Compression is a possible editing decision, not an automatic consequence of low RMS. It changes dynamics rather than shifting every sample by the same gain. Audacity's Compressor manual describes that distinction. The choice depends on what you hear and what the recording is for. An intentional loud phrase does not become a defect because a reference is missed.
Compare the same passage after editing
Change one thing in a copy and export the same passage again. Keep the duration, channels and selection boundaries consistent. Otherwise a lower RMS could come from extra silence or a different take, not your edit. Write down the change, then check the revision against the original. Audacity's Amplify documentation explains how linked gain preserves relative channel levels.
For listening, compare at similar playback levels before deciding that one version sounds clearer. A louder presentation alone is not evidence of less noise or distortion. To test whether you repeatedly hear a difference between matched passages, try the ABX audio test. It cannot decide which version is artistically better.
Know when the check is not enough
No detected full-scale samples does not prove that earlier clipping, distortion or inter-sample peaks are absent. WaveLocus's excerpt check does not measure true peak, integrated LUFS, noise floor or intelligibility. Its partial narration-level reference is not current ACX certification or a requirement for every podcast, video or voice message. Use the destination's current delivery specification for a final file.
Start with a local recording check if you need the measurements. If the same file sounds different across devices or applications, use guided diagnosis for that playback-path problem instead. The calculation above is an illustrative gain example, not a measurement of a real recording or a listening-quality score.