By Team Hypomnema · Updated October 19, 2026
Use time-frequency selections, a spectral brush, restore and removed-sound auditioning to edit audio locally.
Read the frequency threads and spectrogram
Threads is the default view: stacked line traces show measured energy in frequency bands over time. These lines are not separated instruments. Switch to Heatmap in View for color-based inspection. Time runs from left to right; frequency increases toward the top. In Heatmap, brighter colors indicate stronger energy. Horizontal trails often represent sustained tones and harmonics; vertical bursts often represent transients. The display measures the actual signal, without identifying instruments. Try the generated 12-second sound sample to explore hum, a whistle, tones and clicks before opening your own recording.
Select, brush and restore
Open mono or stereo audio up to 100 MB and 10 minutes. Select a rectangle or paint with Brush, choose the channel, and adjust reduction and soft edges. Reduction and soft edges apply across the current edit stack. Restore brings the original signal back in painted regions. Keep selection retains selected frequency regions and silences the rest in affected channels; Restore can bring original sound back outside the kept region too. Use exact time and frequency fields for keyboard editing. Touch gestures paint only in the expanded editor, so ordinary page scrolling stays available inline. Undo and redo keep up to 30 snapshots. At most 40 edits and 512 region marks are allowed; start another stroke after the 64-mark stroke limit.
Listen before exporting
The editor rebuilds the preview in a worker after edits settle. Original, Edited and Removed each have a measured spectrogram. Removed is the source minus the edited signal before independent peak protection, allowing you to hear what the edit discarded. Loop the most recent selected time range while auditioning. Avoid broad, hard-edged selections unless you intend to affect all overlapping sounds. Frequency ranges for 50/60 Hz hum, low rumble and high hiss are starting points, not automatic noise or instrument recognition.
Understand precision and export
Processing uses a 4096-sample FFT at 44.1 kHz with overlapping Hann windows and normalized overlap-add reconstruction. Tracks up to 10 minutes and 100 MB are processed in bounded chunks with contextual overlap, keeping the FFT grid aligned across boundaries. A measurement pass chooses one output gain for the whole track. MP3 encoding keeps one continuous encoder across chunks. The browser still decodes the source into memory, so device memory can limit long files. Threads and Heatmap use the same measured frequency data; time zoom reaches 256× for precise editing of long recordings. Frequency bins are approximately 10.8 Hz apart and analysis windows span about 93 ms. Time boundaries and narrow frequencies cannot be separated perfectly; hard masks may cause ringing and transient smearing. Soft edges smooth frequency transitions and cap time feathering at 25 ms. Instruments share frequencies, so removing a piano or drum region can affect vocals and other instruments. Source peak protection may first lower the decoded recording to provide headroom, with a visible notice. Output peak protection may then lower each signal independently to avoid clipping. The WAV download is the same 44.1 kHz 16-bit PCM file used for auditioning. MP3 is encoded locally at 128, 192, 256 or 320 kbps, is lossy and may include frame padding. Source metadata is not preserved. All editing is temporary and stays on your device; download the output before leaving.
Try the workflow
Keep your original documents and review the output. You can also browse all guides or contact Team Hypomnema.