denizaybey/lid-32-mms1b

Language Identification Model for 32 Most-Spoken Languages

This repository contains resources for 32-language audio language identification (LID) based on Meta AI's MMS (Massively Multilingual Speech) models.

🎯 Quick Start — Use the SOTA Model Directly

The best-performing model for this task is facebook/mms-lid-126, which achieves 97.2% accuracy on FLEURS across 126 languages and covers all 32 target languages.

from transformers import pipeline

classifier = pipeline("audio-classification", model="facebook/mms-lid-126")
result = classifier("/path/to/audio.wav")
print(result[0]["label"])  # ISO 639-3 code, e.g. "eng"

Or use the standalone inference script in this repo:

python inference.py --audio sample.wav

📊 Supported Languages (ISO 639-3)

Code	Language	Code	Language
eng	English	kor	Korean
cmn	Mandarin Chinese	ita	Italian
hin	Hindi	tha	Thai
spa	Spanish	guj	Gujarati
fra	French	fas	Persian (Farsi)
ara	Arabic	pol	Polish
ben	Bengali	ukr	Ukrainian
por	Portuguese	mal	Malayalam
rus	Russian	kan	Kannada
urd	Urdu	ory	Oriya
ind	Indonesian	mya	Burmese
deu	German	pan	Punjabi
jpn	Japanese	nld	Dutch
mar	Marathi	pus	Pashto
tel	Telugu
tur	Turkish
tam	Tamil
vie	Vietnamese

🎮 Live Demo

Try it instantly: denizaybey/lid-32-demo

🏋️ Fine-Tuned Model

A fine-tuned version (denizaybey/lid-32-mms1b) is also available, trained specifically on the 32 target languages from the google/fleurs dataset using:

• Base model: facebook/mms-1b (wav2vec2, 1B params)
• Dataset: FLEURS (clean, professionally recorded audio)
• Training: AdamW, lr=3e-5, effective batch=32, 5 epochs with early stopping

📁 Files

• inference.py — Standalone CLI for inference
• train.py — Full training script for reproduction
• README.md — This file

🔬 Technical Details

Attribute	Value
Architecture	wav2vec2 (MMS-1B)
Parameters	~1B
Sampling Rate	16 kHz
Max Audio Length	30 seconds
Output Format	ISO 639-3 language codes
Training Data	google/fleurs (32 language configs)

📚 References

• Pratap et al., "Scaling Speech Technology to 1,000+ Languages", 2023 — arXiv:2305.13516
• Conneau et al., "FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech", 2022 — arXiv:2205.12446