AkashPrasadMishra/fma-small-segments-150k

FMA-Small Segmented Music Dataset

Overview

This dataset is derived from FMA (Free Music Archive) — small subset, a collection of full-length music tracks released for music analysis/MIR research. The original FMA-small clips are fixed at ~30 seconds each, but to make the data directly usable for training fixed-length audio classification models, every source track has been sliced into uniform 1.5-second segments, using the same extraction method as the FSD50K segmented noise dataset.

This makes the dataset suitable for tasks such as:

• Keyword spotting (KWS) — music as a hard negative / distractor class, alongside noise
• Music / speech / noise discrimination — a 3-way classification setup (music vs. speech vs. environmental noise)
• Audio scene / background classification
• Voice activity detection (VAD) — music segments as a non-speech negative class, distinct from environmental noise negatives
• Audio augmentation — background music mixing for robustness training
• General-purpose audio pretraining where fixed-length windows are required

---

Segmentation Method

Each source track is split into consecutive, non-overlapping 1.5-second windows, starting from 0.0s:

Segment 1:  0.0s  → 1.5s
Segment 2:  1.5s  → 3.0s
Segment 3:  3.0s  → 4.5s
...

For example, a 30-second source track yields 20 segments (0.0–1.5, 1.5–3.0, ..., 28.5–30.0) — identical logic to the FSD50K noise dataset, just applied to music tracks instead of environmental sound clips.

Notes on edge cases:

• If a source track's duration is not an exact multiple of 1.5 seconds, the final remaining chunk is shorter than 1.5s. (Confirm/update: dropped, zero-padded to 1.5s, or kept at its natural shorter length.)
• This windowing is applied identically and independently to every track — no overlap between consecutive segments, and no mixing of audio content across tracks.

---

Dataset Structure

fma_small_seg_150k/
├── <original_track_id>_000.wav   # segment 0.0–1.5s
├── <original_track_id>_001.wav   # segment 1.5–3.0s
├── <original_track_id>_002.wav   # segment 3.0–4.5s
├── ...

(Adjust the naming convention above to match your actual output filenames if different.)

---

Audio Format

Property	Value
Segment length	1.5 seconds (fixed)
Sample rate	(fill in, e.g. 16 kHz)
Channels	(fill in, e.g. mono)
Bit depth	(fill in, e.g. 16-bit PCM)
File format	.wav

---

Dataset Statistics

Metric	Value
Source tracks (from FMA-small)	(fill in, FMA-small = 8,000 tracks)
Total extracted 1.5s segments	~150,000
Total duration	(fill in)

---

Intended Use Cases

• Keyword spotting negatives — music segments, alongside environmental noise, teach a KWS model to avoid false triggers on music playback (radio, TV, background tracks).
• Music / speech / noise 3-way classification — combined with the FSD50K noise dataset and speech/keyword data, this enables training classifiers that distinguish these three broad audio categories.
• Data augmentation — mixing these clips into clean speech or keyword audio at varying SNR levels to build music-robust models (e.g. voice assistants that must work with background music playing).
• Negative sampling for any binary/multi-class audio classifier that needs realistic "non-target, non-noise" audio.
• General audio scene classification where music presence/absence is a relevant signal.

---

Source Dataset & Licensing

This dataset is derived from FMA-small, itself a curated subset of the Free Music Archive. FMA tracks are released under a variety of Creative Commons licenses (CC-BY, CC-BY-SA, CC-BY-NC, CC0, etc.) — license terms vary per individual track, not for the dataset as a whole. If you redistribute or publish this derived/segmented version:

• Retain a mapping from each segment back to its original FMA track ID, artist, and license.
• Respect the individual track's license terms (attribution requirements, non-commercial restrictions, share-alike, etc.) for any derived segment.
• Refer to the official FMA metadata for the authoritative per-track license information.

This README does not itself grant any license beyond what the original FMA tracks carry.

---

Known Limitations

• Fixed 1.5s windowing means genre/style-level musical context is lost within a single segment — this dataset is not suited for tasks requiring full-track context (e.g. genre classification, structure analysis).
• No content-aware/event-centered cropping was applied — segmentation is purely time-based, not tied to musical phrase or beat boundaries.
• Segment-level labels (if provided) are inherited from track-level FMA metadata (e.g. genre) and are not independently re-verified per segment — a given segment may not audibly represent its parent track's overall genre tag (e.g. a quiet intro segment tagged as "Rock").