BA-Lang - Spoken Language Identification checkpoints

Checkpoints for spoken language identification (LID): a self-supervised speech backbone, pooled with MHFA (Multi-Head Factorized Attention) and scored with either a standard cosine-softmax classifier or BA-LR (Binary Attribute Likelihood Ratio), a Bernoulli classifier over a learned binary attribute embedding.

• Training code / recipe: [https://github.com/AMIAD-Research/BA-Lang]
• BA-LR / BA-Lang model: Y. Jelassi et al., "Reconnaissance de la langue parlée explicable avec BA-LR", JEP 2026 — isca-archive.org/jep_2026/jelassi26_jep.html

Architecture

raw audio (16 kHz)
      │
      ▼
 SSL backbone (MMS-1B or wav2vec2-XLS-R)
      │
      ▼
 Multi-Head Factorized Attentive pooling (MHFA)
      │
      ▼
 classification head
   ├─ Linear, or
   └─ BA-Lang: binary encoder → per-class Bernoulli log-likelihood
      │
      ▼
 language logits

Checkpoints

File	Backbone	Corpus	Languages	Head	Step	Val. accuracy	Size
MMS-Fleurs-step-010360-val_classif_acc-0.9168118238449097.ckpt	MMS-1B	FLEURS	102	Linear	10360	91.68%	3.86 GB
MMS-Fleurs-e2e-balang-stepstep-008856-val_classif_acc-0.9167537689208984.ckpt	MMS-1B	FLEURS	102	BA-Lang (e2e)	8856	91.68%	3.87 GB
MMS-Fleurs-e2e-balang-MIR-rho-0.95-step-032472-val_classif_acc-0.90935212373733...ckpt	MMS-1B	FLEURS	102	BA-Lang (e2e) + Matryoshka-inspired regularization, ρ=0.95	32472	90.94%	3.87 GB
MMS-Fleurs-e2e-balang-ND-rho-0.857-step-047232-val_classif_acc-0.89306861162185...ckpt	MMS-1B	FLEURS	102	BA-Lang (e2e) + Nested Dropout, ρ=0.857	47232	89.31%	3.87 GB
MMS-Tidylang-step-008533-val_classif_acc-0.9778640866279602.ckpt	MMS-1B	Tidylang	35	Linear	8533	97.79%	3.86 GB
MMS-Tidylang-e2e-balang-step-005565-val_classif_acc-0.9929152131080627.ckpt	MMS-1B	Tidylang	35	BA-Lang (e2e)	5565	99.29%	3.87 GB
XLSR-Fleurs-step-005904-val_classif_acc-0.8094159960746765.ckpt	wav2vec2-XLS-R	FLEURS	102	Linear	5904	80.94%	1.27 GB
XLSR-Fleurs-e2e-balang-step-047232-val_classif_acc-0.8120864033699036.ckpt	wav2vec2-XLS-R	FLEURS	102	BA-Lang (e2e)	47232	81.21%	1.27 GB
XLSR-Tidylang-step-008904-val_classif_acc-0.9662737846374512.ckpt	wav2vec2-XLS-R	Tidylang	35	Linear	8904	96.63%	1.27 GB
XLSR-Tidylang-e2e-balang-step-032648-val_classif_acc-0.9860262870788574...ckpt	wav2vec2-XLS-R	Tidylang	35	BA-Lang (e2e)	32648	98.60%	1.27 GB

Datasets

Training is done separately on two corpora — pick the checkpoint matching the languages you need.

• FLEURS (102 languages, ISO 639-3): afr, amh, ara, asm, ast, aze, bel, ben, bos, bul, cat, ceb, ces, ckb, cmn, cym, dan, deu, ell, eng, est, fas, fil, fin, fra, ful, gle, glg, guj, hau, heb, hin, hrv, hun, hye, ibo, ind, isl, ita, jav, jpn, kam, kan, kat, kaz, kea, khm, kir, kor, lao, lav, lin, lit, ltz, lug, luo, mal, mar, mkd, mlt, mon, mri, msa, mya, nep, nld, nob, nso, nya, oci, ori, orm, pan, pol, por, pus, ron, rus, slk, slv, sna, snd, som, spa, srp, swa, swe, tam, tel, tgk, tha, tur, ukr, umb, urd, uzb, vie, wol, xho, yor, yue, zul.
• Tidylang (35 languages): TidyVoiceX_ASV (train/dev) and TidyVoiceX2_ASV (eval), from Mozilla Data Collective (train/dev, eval): ara, bak, bel, ben, bul, cat, chv, cym, deu, div, ell, eng, fas, fra, hin, hye, jpn, kat, lit, lug, mal, mar, nld, ori, pol, por, rus, tam, tha, tuk, tur, uig, uzb, yue, zho.

import yaml
lang_list = yaml.safe_load(open("configs/data/fleurs.yaml"))["language_list"]
# or configs/data/tidylang.yaml for the Tidylang checkpoints

Quickstart — inference on your own audio

git clone https://github.com/AMIAD-Research/BA-Lang
cd BA-Lang
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

pip install huggingface_hub
python -c "
from huggingface_hub import hf_hub_download
hf_hub_download(
    repo_id='yjelassi/BA-Lang',
    filename='MMS-Fleurs-e2e-balang-stepstep-008856-val_classif_acc-0.9167537689208984.ckpt',
    local_dir='.',
)
"

Then, on a single audio file (no manifest needed — this loads the checkpoint directly and runs one forward pass):

import torch
import torchaudio
import yaml

from balr_lid.models.lid_model import LanguageIdentificationModel

CHECKPOINT = "MMS-Fleurs-e2e-balang-stepstep-008856-val_classif_acc-0.9167537689208984.ckpt"
LANGUAGES = yaml.safe_load(open("configs/data/fleurs.yaml"))["language_list"]

model = LanguageIdentificationModel.load_from_checkpoint(
    CHECKPOINT, map_location="cpu", weights_only=False
)
model.eval()

waveform, sr = torchaudio.load("my_audio.wav")
if waveform.shape[0] > 1:  
    waveform = waveform.mean(dim=0, keepdim=True)
if sr != 16000:
    waveform = torchaudio.functional.resample(waveform, sr, 16000)

with torch.no_grad():
    out = model(waveform) 
    probs = out["logits"].softmax(dim=-1)[0]

top5 = probs.topk(5)
for score, idx in zip(top5.values, top5.indices):
    print(f"{LANGUAGES[idx]}: {score.item():.3f}")

Batch inference over a dataset

To evaluate on many files at once (accuracy, confusion matrix, embedding dump), build a manifest CSV (see docs/data_manifest.md) and use the repo's batch script instead:

python -m balr_lid.inference \
    --checkpoint MMS-Fleurs-e2e-balang-stepstep-008856-val_classif_acc-0.9167537689208984.ckpt \
    --manifest data/manifests/fleurs-test.csv \
    --audio-root /path/to/fleurs/audio \
    --language-list afr amh ara ... \
    --output predictions.csv

--language-list must match the checkpoint's training corpus, in order (see Datasets above).