Audio
Label audio recordings with time segments on a waveform and a spectrogram, with spectrogram settings fixed per project so every label can be reproduced.
The Audio Labeling Toolbox is a browser-based interface for labeling recordings with time segments: acoustic events, speech, machine noise, anything that has a start and an end. It is built for tasks where the evidence is in the frequency content, such as NVH and acoustic quality inspection, bioacoustics, and sound event detection.
Waveform and spectrogram of the same recording on one playhead
Segments labeled with tags, down to a single sample, on all channels or on one channel
Whole-recording tags for classifying a file as a whole
Label where you draw: pick a label right on the plot, or from the Labels tab with one click or a hotkey
Spectrogram settings set once per project, so every annotator sees the same analysis and the analysis can be reproduced later for training
Playback of the whole recording or of the selection, looped, at 0.25× to 2×
🔗 Sample project: try the toolbox on the Audio Annotation Sample Project from the Ecosystem.
Create an Audio project
Open Projects and click + New → New project wizard.
Choose Audio as the type of project. The type cannot be changed later.
Name the project and create it.
Audio projects need a Supervisely license that includes the audio recordings module.
Import recordings
Supported formats: WAV, FLAC, MP3, OGG, M4A.
Plain recordings in any folder structure are uploaded to one dataset, without labels. Files with other extensions are skipped. See Audio recordings.
Supervisely format keeps datasets, segments, whole-recording tags and the project's spectrogram settings. See Supervisely audio format.
Import with the import wizard on the project, or from Team Files or Cloud Storage. Audio is always uploaded; it cannot be added by link.
From Python, install the SDK with the audio extra, pip install "supervisely[audio]", and follow Audio in the Python SDK.
Overview
The toolbox opens when you click a recording in an Audio project.
Top bar — previous and next recording, undo and redo, theme, the hotkeys reference and More.
Toolbar — the Label range tool.
Audio panel — the time ruler, the waveform and the spectrogram, the overview strip and the transport bar.
Annotations and Recordings tabs.
Labels and View tabs.
Top bar and layout
Undo / Redo step through annotation edits (Ctrl+Z / Ctrl+Shift+Z).
Theme switches between the dark and the light theme.
The hotkeys button opens the full list of hotkeys.
Tabs can be dragged to another column or stacked together. More → Restore Default Layout puts them back.
Label a segment
Drag across the waveform or the spectrogram to select a range. A plain click only moves the playhead; Shift-click selects a single sample.
Click Choose a label next to the range, and pick a label. Type to search the list.
If the label has a value, enter it. For a one-of value, keys 1–9 pick an option. Enter saves, Esc cancels.
The segment is saved at once. It covers the selected samples, both ends included, and it is about the channel you are viewing: Mix labels all channels, Ch N labels channel N only.
You can also label from the Labels tab. The line at the top shows the selected range and its channel. Each label row has two buttons: selected range and whole recording. Clicking a label, or pressing its hotkey, labels the selected range, or the whole recording when no range is selected.
A label's scope, set in the project's tag settings, decides where it can go: on ranges, on the whole recording, or both. Labels that apply to objects only are not listed.
Click + in the Labels tab to create a label without leaving the toolbox, or the gear icon to manage labels in the project settings.
Whole-recording tags
Use the whole recording button on a label row, or click a label when no range is selected. These tags apply to all channels and show as badges above the plots. Click a badge to select it and edit its value. To change its label, use the Annotations tab.
Tick Go to next recording after labeling the whole recording to move on as soon as the tag is added.
Edit and delete
On the plot. Click a segment to select it.
Drag its edges to resize it; Alt-drag moves it. Changes stay a draft until you click the check mark (Save changes). Esc or the cross cancels them.
Change label swaps the label and keeps the range and the channel.
In the Annotations tab. Every segment and whole-recording tag of the recording is listed, with a small timeline.
Search, or filter by label and by channel.
Each row can be hidden, edited or deleted, and a segment can be played.
The pencil opens an inline editor under the list: Start, End (inclusive) in seconds or in samples, Channel, and the value. Save or Cancel it there. Change label is there too, for segments and whole-recording tags.
Tick several rows to show, hide or delete them together. Deleting always asks to confirm.
Channels
The headphones menu in the transport bar chooses what you see and hear: Mix all channels, or a single Channel N. The button shows Mix or Ch N.
When one channel is selected, segments on other channels are hidden. If the selected segment is on another channel, Show ch N switches to it.
Navigation and playback
Scroll to zoom around the pointer, Shift-scroll to pan, or drag the time ruler to scrub.
Zoom out, Fit recording and Zoom in buttons are in the transport bar. Fit recording (Alt+F) shows the whole file.
The overview strip shows the whole recording: drag the window to pan, drag its edges to zoom, click to jump.
Previous interval and Next interval in the transport bar jump to the neighbouring segment. Shift+P / Shift+N select it and play it.
Play selection and Loop selection (Shift+L) play just the selected range.
Playback speed: 0.25×, 0.5×, 0.75×, 1×, 1.25×, 1.5× or 2×.
The time field shows the playhead position and accepts a time to jump to. Its tooltip shows the sample rate, channel count and length in samples.
Hotkeys
Play selection or recording / pause
Space
Play from playhead / pause
Shift+Space
Toggle selection loop
Shift+L
Select and play the previous / next interval
Shift+P / Shift+N
Fit recording
Alt+F
Undo / redo
Ctrl+Z / Ctrl+Shift+Z
Apply a label
the label's own hotkey
These work when the plot is focused (click it first):
Move playhead by one sample / one second
← → / Shift+← Shift+→
Go to start / end
Home / End
Cancel a drag or a draft
Esc
Spectrogram settings
A spectrogram is one of many possible pictures of the same audio, and the settings decide what is visible: a small FFT window separates events in time but blurs close frequencies together, a narrow contrast range hides quiet events. Labels are only comparable if they were drawn on the same picture. That is why the settings belong to the project: every recording is shown with them, for everyone.
They are in the View tab. A banner at the top says whether they are set:
Not set for this project — recordings use the default analysis. Click Save for project, or change any setting, to fix it for the project.
Set for this project — changes apply to every recording in the project, for everyone. They never edit existing labels.
A change is saved to the project as soon as you make it.
Frequency display
Linear (general purpose), Logarithmic (lower pitches), Mel (speech and hearing)
Detail balance
Timing, Balanced, Pitch — FFT windows of 512, 2048 and 8192 samples. Any other window shows as Custom.
Colors
Magma, Viridis, Grayscale
Display style
Sharp, Smooth
Quiet cutoff / Bright cutoff
the dBFS range the colors span; the quiet cutoff must be lower
Advanced analysis
FFT window (32 to 32768 samples), Step (samples), Window function (Hann, Hamming, Blackman), Mel bands (2 to 512, Mel only)
The defaults are Linear, a 2048-sample window with a 512-sample step, Hann, 128 mel bands, −100 to 0 dBFS, Magma, Sharp.
The settings are saved in the project and exported with it, so the same spectrogram can be computed outside Supervisely — for example as training input in PyTorch or TensorFlow.
Recordings tab
Lists the recordings of the dataset with their length. Click one to open it. A recording that cannot be decoded is marked cannot be displayed.
The row menu downloads the original file, and the metadata panel shows the recording's ID, dates and metadata.
Export and SDK
Export a project or a dataset with Export to Supervisely format. The result keeps segments, whole-recording tags and the spectrogram settings. See Supervisely audio format.
Audio in the Python SDK covers uploading, labeling, rendering spectrograms and exporting training crops.
Browser support
The views are drawn on the GPU with WebGPU, or with WebGL2 where WebGPU is not available.
WAV, FLAC and MP3 play in any supported browser. OGG and M4A are decoded by the browser, which needs the instance to be opened over HTTPS. If the browser cannot decode a recording, the toolbox shows Recording cannot be displayed; re-encode it to WAV or FLAC, or use a browser that supports the codec.
Last updated
Was this helpful?