Show HN: Voice gender classifier for European voice AI (1MB, ONNX, 4ms) (huggingface.co)

A new lightweight voice gender classifier has been introduced, specifically tailored for real-time voice AI pipelines in European languages. This model, which operates on CPU within 5 ms and boasts a compact size of 0.64 MB, enables voice AI assistants to accurately detect a caller's gender based on their voice. This capability is crucial in languages like Polish, German, and Spanish, where gender influences grammatical structure and politeness forms. The classifier utilizes a 2-layer Bidirectional LSTM architecture to process 3-second audio clips and is compatible with ONNX, allowing for efficient inference without reliance on PyTorch. The significance of this development lies in its optimization for the specific linguistic needs of European markets, where gender-specific language use is essential. Evaluated across diverse datasets, the model achieved impressive accuracy rates, particularly on standard-accented English and other targeted languages. Though it performs well, it also has limitations, such as reduced accuracy with strongly accented speech and a focus only on binary gender classification. The implementation of this classifier can enhance user experiences for voice AI applications, making them more responsive and culturally aware.