Back to blog

Audio Annotation in Supervisely: Waveform and Spectrogram Labeling

Updated Labeling

Label speech, machine noise and other sound events as time segments on a waveform and a spectrogram, with spectrogram settings fixed for the whole project.

Audio Annotation in Supervisely: Waveform and Spectrogram Labeling

We’re introducing Audio Annotation in Supervisely — a new way to turn raw recordings into structured training data without leaving the platform.

Audio projects combine playback, a waveform, a spectrogram and labels on one synchronized timeline. You can mark the exact moment when someone speaks or a sound occurs, describe the recording as a whole, review the result and export everything for the next stage of your ML pipeline.

The audio toolbox with an unlabeled speech recording on the waveform and the spectrogram, and the Labels panel listing audio_quality, beep, speaker_1 and speaker_2The audio toolbox: playback, the waveform and the spectrogram on one timeline, with the project's labels on the right

For the full reference, see the Audio Labeling Toolbox in the Supervisely docs. It covers every control, setting and hotkey.

See what you hear

Listening alone is often not enough to place precise boundaries. A waveform shows changes in signal amplitude, while a spectrogram reveals how frequencies change over time.

Three labeled segments, speaker_1, beep and speaker_2, on the waveform and the spectrogram of one recording, with the whole-recording tag audio_quality: usable above the plotsspeaker_1, beep and speaker_2 segments on the waveform and the spectrogram, with the whole-recording tag audio_quality: usable

In the example above, the two speech segments form complex frequency patterns. The short beep is much easier to recognize as a narrow horizontal line. Because the waveform and spectrogram share the same playhead and time selection, you can use both views to find an event and label it accurately.

Spectrogram settings are saved for the project, giving the team a consistent view across all recordings. The default Linear frequency scale and Balanced detail are a good starting point, while the View tab provides more control when the task requires it.

The View tab with the Linear frequency display, Balanced detail balance, Magma colors and contrast cutoffs, beside the speaker_1, beep and speaker_2 segments on the waveform and the spectrogram and the Annotations timeline with the whole-recording tag audio_quality: usableThe View tab's spectrogram settings, Linear frequency display and Balanced detail, beside a labeled recording

Label events and entire recordings

Audio Annotation supports two complementary levels of labeling.

Time-range labels describe events inside a recording. Drag over the waveform or spectrogram, then apply a label such as speaker_1, beep or speaker_2. Segment boundaries are stored with sample-level precision. For multichannel audio, annotations can be applied to the mixed signal or to a specific channel.

Whole-recording labels describe the complete file. In this example, audio_quality: usable tells downstream users that the recording is suitable for the dataset. It is regular annotation metadata — not an automatic quality check or a review status — so teams can define values that match their own workflow, such as usable and noisy.

A speech recording with the whole-recording tag audio_quality: usable shown as a badge above the waveform and the spectrogram, and listed under Whole recording as Entire recording in the Annotations tabThe whole-recording tag audio_quality: usable, as a badge above the plots and under Whole recording in the Annotations tab

Together, these two levels cover both what happens and when and what is true about the recording as a whole.

Review and refine in the same workspace

All annotations appear in a searchable timeline below the audio. Annotators and reviewers can play a segment, filter by label, drag its edges or enter exact start and end values in seconds or samples.

Adjust a segment and save the change

This keeps the full workflow in one place: import recordings, define the label schema, annotate events, add file-level metadata and check the final result.

Move annotations into your training pipeline

Run Export to Supervisely format when the dataset is ready. The export keeps the recordings, time-range labels, whole-recording tags and project spectrogram settings.

For Python workflows, the Audio SDK tutorial shows how to read the annotations, reproduce the project spectrogram and save labeled time ranges as training crops. This makes the path from annotation to model input reproducible instead of relying on manual timestamps or separate tools. The tutorial requires Supervisely SDK 6.74.41 or newer.

Audio projects can support speech datasets, sound-event detection, bioacoustics, call analysis, and machine or vehicle noise research.

Try Audio Annotation

You can explore the complete example from this post in the sample project. Open 03_two_speakers_and_beep.wav to see three event labels and the whole-recording tag audio_quality: usable.

Audio Annotation Sample Project
Supervisely Dataset

Audio Annotation Sample Project

Practice project with audio recordings. Perfect for trying audio segment and recording tagging.

To start with your own data, create an Audio project and import WAV, FLAC, MP3, OGG or M4A files.

The Create Project page with Audio selected as the type of projectCreate a project with Audio selected

Audio projects require Supervisely 6.18.0 or newer and a license that includes the audio recordings module.

Want to evaluate Audio Annotation on a large dataset? Contact us.

Frequently asked questions

What is audio annotation?

Audio annotation is labeling sound events in a recording as time segments, such as speech turns, machine noise or a beep. In Supervisely you select a range on the waveform or the spectrogram and pick a label.

When should I use a spectrogram instead of a waveform?

The waveform shows how loud the sound is over time, which is enough to find speech turns or a knock with silence around it. The spectrogram shows which frequencies sound at each moment, so it is the view for events that differ from the background by pitch, such as a whine in an NVH test or a bird call.

How do I label a whole audio file with one label?

Click a label with no range selected, or use its whole recording button. The tag applies to all channels and shows as a badge above the plots. Tick Go to next recording after labeling the whole recording to classify many files quickly.

Are spectrogram settings the same for every annotator?

Yes. The settings belong to the project: after you click Save for project in the View tab, every recording is shown with them for every annotator. Existing labels are not changed when a setting changes, so choose the settings before labeling starts.

Which audio formats does Supervisely support?

WAV, FLAC, MP3, OGG and M4A. 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.

What do I need to start labeling audio in Supervisely?

Supervisely 6.18.0 or newer and a license that includes the audio recordings module. Create a project of the Audio type, import recordings and open one in the toolbox. From Python, use Supervisely SDK 6.74.41 or newer.

Ira Soldatova

Python Developer at Supervisely

More from the blog

All posts