🔉 New SAM-Audio model for audio segmentation – great potential for data augmentation

[LimitlessGreen]

🔉 New SAM-Audio model for audio segmentation – potential for data augmentation

Hi everyone,

I wanted to share a recent finding that might be interesting and potentially useful for data augmentation in bioacoustic workflows.

Meta’s Segment Anything family has a new member: SAM-Audio — an audio segmentation model that enables prompt-based separation of sound events from mixed audio recordings.

You can try it here:
🔗 https://aidemos.meta.com/segment-anything/editor/segment-audio/?media_id=1343694437028496

For my test, I used "bird chorus" as the separation prompt on a mixed recording.

---

🎵 Example results

• Combined input audio

combined_audio.mp4

• Isolated target sound (bird chorus)

isolated_sound.mp4

• Remaining background (audio without isolated birds)

without_isolated_sound.mp4

As you can hear, the bird-only result is surprisingly clean, especially given that this is a prompt-based, model-agnostic separation approach.

---

🧠 Why this might be relevant for BirdNET

This approach could be useful as a semi-automated way to generate training data, for example:

• A: Cleaner bird vocalizations
• B: Background / environmental noise

Both could then be recombined in a controlled way to generate augmented samples with:

• known signal-to-noise ratios,
• different acoustic contexts,
• or rare species embedded into realistic soundscapes.

This is not intended as a replacement for curated field recordings, but potentially as:

• 🧪 a data augmentation tool
• 🔧 a preprocessing step
• 🌱 a way to bootstrap datasets where clean samples are scarce

---

💡 Motivation

I’m mainly sharing this as an informational note and possible inspiration for future workflows around BirdNET or related tooling.

Best regards