Babel-LID (Pilot — Metadata Only)

Open-source multi-duration language identification dataset with acoustic robustness tiers.

Pilot release: 4 languages (en, fr, de, es), 339,328 metadata rows. Real pipeline applied end-to-end on Modal.

Note: This release contains only the metadata (parquets). The 84,816 augmented WAV files (4 tiers × 2 channels × ~10K source files) are stored on the Modal Volume babel-lid-v2 and will be added in a subsequent release once the upload pipeline is optimized.

Stats

Metric	Value
Languages	4 (en, fr, de, es)
Source audios	10,825 (LibriSpeech + MLS)
Verified annotations	10,604
Augmented WAV files	84,816 (stored on Modal)
Metadata rows	339,328
Train / Val / Test	236,544 / 51,936 / 50,848

Real Pipeline (all on Modal)

1. Stage 1 — Download real speech from LibriSpeech (en) + MLS (fr, de, es)
2. Stage 2 — Whisper large-v3 transcription verification (A10G GPU)
3. Stage 3 — 3-model LID ensemble: Whisper + ECAPA-TDNN + MMS-LID-256 (A100 GPU)
   • 92% high-confidence (3/3 unanimous)
   • 5.6% medium-confidence (2/3)
   • 2.2% rejected
4. Stage 3b — Linguistic cross-verification via Whisper ASR confidence
5. Stage 4 — Acoustic augmentation: 4 tiers × 2 channels = 8 variants per source
6. Stage 5 — Duration block metadata: D1/D3/D5/D10

Augmentation Tiers

Tier	Augmentation	Details
T0	Clean	Original audio, 16kHz mono
T1	Additive noise	White noise at SNR 5–15 dB
T2	Reverb	Synthetic RIR, RT60 0.3–0.8s
T3	Multi-distortion	Reverb + noise SNR 0–10 dB
C0	Wideband	Full 16kHz bandwidth
C1	PSTN telephony	300–3400 Hz bandpass → 8kHz → μ-law G.711 → 50Hz hum → 16kHz

License

CC-BY 4.0