mutisya/w2v-bert-adapters-14lang-e10-28_54-v9

W2V-BERT 2.0 ASR Adapters (v25 - ConvAdapter)

This repository contains per-language ConvAdapter modules for automatic speech recognition (ASR) trained on top of facebook/w2v-bert-2.0.

Model Description

• Base Model: facebook/w2v-bert-2.0 (600M parameters, frozen)
• Adapter Architecture: ConvAdapter (Conv1d + depthwise temporal conv, dim=64)
• Decoder: Lightweight transformer decoder (2 layers)
• Training: CTC loss with extended vocabulary for double vowels

Trained Adapters

Summary: 2 adapters trained

• ✅ Good (WER < 30%): 0
• ⚡ Medium (30-60%): 0
• ❌ Collapsed (≥ 90%): 2

Adapter	Language	WER	Status	Train Samples
ach_Latn	Acholi	99.63%	❌ Collapsed	4,825
eng_Latn_salt	English (SALT)	94.16%	❌ Collapsed	4,804

Architecture (v25 ConvAdapter)

The model uses:

1. Frozen w2v-bert-2.0 encoder - Extracts audio representations
2. ConvAdapters - Conv1d + depthwise temporal conv (kernel=5) for local acoustic context
3. Lightweight decoder - Transformer decoder blocks (trainable)
4. LM head - Per-language vocabulary projection (trainable)

ConvAdapter Details

• Down projection: Conv1d(1024 → 64, k=1)
• Temporal conv: DepthwiseConv1d(64, k=5)
• Up projection: Conv1d(64 → 1024, k=1)
• Activation: SiLU (Swish)
• ~131K params per adapter, ~3.2M total

Usage

Each adapter folder contains:

• adapter_weights.pt - ConvAdapter weights
• decoder_weights.pt - Decoder block weights
• lm_head_weights.pt - Language model head weights
• final_norm_weights.pt - Final layer norm weights
• vocab.json - Language-specific vocabulary
• adapter_config.json - Adapter configuration
• metrics.json - Training metrics

Training Configuration

• Epochs: 10
• Base Learning Rate: 0.0005 (adaptive based on dataset size)
• Batch Size: 48 x 1
• Extended Vocabulary: True
• Adapter Dimension: 64
• Conv Kernel Size: 5

License

Apache 2.0