---
library_name: transformers
license: apache-2.0
pipeline_tag: feature-extraction
tags:
- audio
- wav2vec2
---

Fairseq wav2vec2-base pretraining checkpoint (checkpoint_best, ~85000 updates),
converted to HuggingFace format with the official transformers converter and
verified (weight-level spot check + forward-pass comparison against the
fairseq model; see conversion log). Pretrained on MAGICDATA (Mandarin Chinese).

Raw fairseq checkpoints: techsword/wav2vec2-base-mandarin-magicdata-checkpoints
Used in: tone-probe experiments, https://github.com/techsword/tone-encoding-in-speech-model