WavLM Base Plus

Microsoft WavLM-Base-Plus is a self-supervised speech embedding model that generates general-purpose representations from raw audio, optimized for 16 kHz sampled speech. Built on the HuBERT framework, it incorporates gated relative position bias in the Transformer layers to improve sequence modeling and an utterance mixing training strategy that overlays unsupervised speech segments to enhance speaker discrimination.

The model was pre-trained on 94,000 hours of English audio drawn from three large-scale corpora: Libri-Light (60,000 hours of audiobooks), GigaSpeech (10,000 hours of multi-domain transcribed speech), and VoxPopuli (24,000 hours of unlabeled speech from European Parliament proceedings). This diverse training setup enables the model to learn robust, multi-faceted representations that capture spoken content, speaker identity, and paralinguistic cues.

The WavLM architecture achieved state-of-the-art results on the SUPERB benchmark (with the Large variant), and the base-plus variant is well-suited for tasks such as speaker recognition, audio classification, and speech separation when fine-tuned or used as a feature extractor. As a hosted API through gigarouter, the model accepts 16 kHz audio input and returns high-dimensional embeddings without requiring local installation or GPU resources.

Refer to the WavLM paper and the official repository for further architectural details and evaluation protocols.