Lightweight voice gender classifier for European languages released: 0.64 MB ONNX model runs in ~4 ms on CPU

A Bi-LSTM gender classifier model (166K parameters, 0.64 MB) designed for real-time voice AI pipelines supporting English, German, French, Spanish, and Italian. Exported to ONNX format; runs on CPU in ~4 ms single-threaded inference.

Achieves 94.4% accuracy on LibriSpeech test-clean, 94.3% on FLEURS multilingual test (EN/DE/FR/ES/IT), and 90.9% on LibriSpeech test-other. Performance degrades on strongly accented speech (75.6% on Edinburgh International Accents), and model is not benchmarked under noise or telephony compression.

Trained on LibriSpeech train-clean-100 and FLEURS training split, balanced 50/50 male/female. Built for voice pipelines in languages with grammatical gender, where correct gender classification enables proper inflection of adjectives, verb forms, and honorifics.