mutisya/w2v-bert-asr-african-languages

W2V-BERT-2.0 ASR Adapters - African Languages

Per-language bottleneck adapters for multilingual ASR using w2v-bert-2.0.

Model Architecture

• Base: facebook/w2v-bert-2.0 (frozen, 600M params)
• Adapters: MMS-style bottleneck (dim=64, ~2M params each)
• Decoder: Lightweight transformer (1 layer, ~6M params)
• Per-language: Vocabulary and LM head

Available Adapters

Adapter ID	Language	WER	Train Samples
Training in progress...

Training Progress

• Completed: 0/5
• Samples per adapter: 15,000
• Epochs: 5

Usage

from huggingface_hub import snapshot_download

# Download model
model_dir = snapshot_download("mutisya/w2v-bert-asr-african-languages")

# Load specific adapter
adapter_path = f"{model_dir}/adapters/swh_Latn_v1"

License

Apache 2.0