Wav2Vec2 Base 960h

Facebook Wav2Vec2-Base-960h is an automatic speech recognition (ASR) model that transcribes 16 kHz sampled speech audio into text. It is a self-supervised model pre-trained on raw audio and then fine-tuned on 960 hours of LibriSpeech data, using a contrastive task defined over quantized latent representations. This approach achieves competitive word error rates while being conceptually simpler than prior semi-supervised methods.

Key Capabilities

• State-of-the-art performance on LibriSpeech: 3.4% WER on the clean test set and 8.6% WER on the more challenging "other" test set.
• Robust to limited labeled data: the underlying wav2vec 2.0 framework (for which this model is a fine-tuned variant) can outperform previous methods using 100× less labeled data, as demonstrated in the original paper.
• Designed for 16 kHz input audio; the model directly processes raw waveforms without feature engineering.

Benchmark Results on LibriSpeech

These numbers reflect the Base-960h variant fine-tuned on all labeled LibriSpeech data. The larger model variant (paper reference) achieves lower WER (1.8%/3.3%) but uses a different architecture.

Background

The model is part of Facebook AI’s wav2vec 2.0 series (Baevski et al., 2020, arXiv:2006.11477). It learns powerful speech representations from raw audio alone, then fine-tunes on transcribed speech. Gigarouter hosts this model as a managed, OpenAI-compatible API — no local installation or framework handling required.