Download README.md from thelamapi/Next-Echo: direct link, hf CLI and curl.
- Browser
- Download file 3.64 kB
-
https://huggingface.co/thelamapi/Next-Echo/resolve/main/README.md
- Command line
-
hf download hf://thelamapi/Next-Echo/README.md
-
curl -L -o README.md https://huggingface.co/thelamapi/Next-Echo/resolve/main/README.md
license: cc-by-nc-3.0
library_name: pytorch
tags:
- audio
- query-by-example
- sound-event-detection
- next-lab
Next Echo
Next Echo finds every moment a given example sound occurs in a longer recording — a clap, a knock, a bark — with no fixed class list. A small CNN embeds 1-second log-mel windows; the recording is scanned against the example. It is compared with plain log-mel similarity; the final score fuses both after per-recording normalisation.
Space: https://huggingface.co/spaces/thelamapi/Next-Echo — record a 2-second example with your microphone (or pick one), choose a recording, and listen to each hit. Runs in the browser.
Results on constructed 30 s test recordings (ESC-50 fold 5; 40 recordings per cell)
Event F1 (±0.5 s) / false alarms per minute — log-mel baseline vs Next Echo (better F1 in bold):
| Classes / condition | Log-mel similarity | Next Echo (learned) |
|---|---|---|
| seen/clean | 0.676 / 2.15 | 0.619 / 3.25 |
| seen/noisy | 0.606 / 1.45 | 0.630 / 2.30 |
| seen/overlap | 0.788 / 0.80 | 0.749 / 1.05 |
| seen/absent | – / 4.00 | – / 4.95 |
| unseen/clean | 0.617 / 2.10 | 0.658 / 3.30 |
| unseen/noisy | 0.667 / 1.10 | 0.717 / 2.05 |
| unseen/overlap | 0.790 / 0.85 | 0.767 / 1.70 |
| unseen/absent | – / 3.65 | – / 5.60 |
| seen/all-present | 0.690 / 1.47 | 0.662 / 2.20 |
| unseen/all-present | 0.691 / 1.35 | 0.712 / 2.35 |
unseen = the 10 classes never used in training (clapping, door_wood_knock, rooster, church_bells, glass_breaking, keyboard_typing, siren, cow, sneezing, can_opening). absent recordings contain no target, so only false alarms count. Thresholds were chosen on validation recordings built from fold 4.
Method
- Audio: 16 kHz mono, 512-point STFT (hop 256), 64 mel bands (50–8000 Hz), log.
- Embedding: 4 conv blocks + linear → 64-d, supervised contrastive loss (τ = 0.1) on random 1-s crops of 960 clips from 40 classes, with gain, mixing and noise augmentation, 5000 steps.
- Training windows mimic the test conditions: real waveform segments mixed with another sound and brown noise (0–30 dB SNR), alongside clean segments like the user's example.
- Query: the loudest 1-s windows of the example are embedded and averaged. Recording: windows every 0.256 s.
- Score: per-recording robust normalisation (median/MAD), then Next Echo = α·z(learned) + (1−α)·z(log-mel) with α = 0.75 chosen on validation recordings (options α ∈ {1, 0.75, 0.5}). Consecutive windows above the threshold form one detection.
- Baseline: cosine between per-band mean/std of log-mel windows, raw or normalised — whichever validated better (raw).
- The browser computes the same log-mel (own FFT) and CNN with
nlrt.js; Node parity on 3 real snippets: max abs difference 7.105427357601002e-15.
Usage
python inference.py --query clap.wav --recording scene.wav # 16 kHz mono 16-bit WAV
python inference.py --selftest
Limitations and license
- Recordings are mixtures of real clips, not naturally recorded scenes; real rooms add reverberation and far more varied backgrounds.
- Very short or very quiet examples give weak queries. Sounds that resemble each other (knock vs footsteps) cause false alarms.
- Non-commercial: ESC-50 is CC BY-NC 3.0, so this model, its demo audio and the Space are for non-commercial use only.
- Trained on NVIDIA L4; profile full. Failure cases are listed in
evaluation.json.

