sam4rano/luhya-asr

--- language: - luy license: apache-2.0 tags: - asr - automatic-speech-recognition - luhya - w2v-bert metrics: - wer - cer model-index: - name: luhya-asr-w2v-BERT results: - task: name: Automatic Speech Recognition type: automatic-speech-recognition metrics: - name: Validation WER type: wer value: 43.45 - name: Validation CER type: cer value: 11.36 --- # luhya-asr-w2v-BERT Fine-tuned Facebook Wav2Vec2-BERT 2.0 (580M parameters) for Automatic Speech Recognition in Luhya, a Bantu language spoken in Kenya. > **Historical result:** the metrics below describe the earlier 20-hour, > validation-only Wav2Vec2-BERT run. They must not be compared with the aligned > 40-hour Whisper result. Use the comparison notebook below to generate the new > matched validation/test report. Model on HuggingFace Hub: Sam4rano/luhya-asr-w2v-BERT ## Dataset | | Train | Validation | Test | |---|---|---|---| | **Source** | `Digital-Divide-Data/Luhya-ASR-Data-subset-50h` (HF Hub; formerly `DDD-Kenya/...`, now deprecated) | same | N/A | | **Samples** | 6,947 | 400 | — | | **Duration** | 20.00 hours | ~1.1 hours | — | | **Notes** | Randomly subsampled from 23,198 samples (66.52 h) with 20-hour cap; 1 corrupt sample removed | Used as-is; no hour cap | No independent test split | ## Model | Detail | Value | |---|---| | **Base model** | `facebook/w2v-bert-2.0` | | **Parameters** | 580M (24 Conformer layers, 1024 hidden size, 16 attention heads) | | **Feature extractor** | SeamlessM4T (80-dim mel spectrograms at 16 kHz) | | **Decoding** | CTC (Connectionist Temporal Classification) | | **Adapter** | TDNN final layer (kernel=3, stride=2) | | **Feature encoder** | Frozen (pretrained weights preserved) | | **Vocabulary** | 48 tokens (a-z, 0-9, basic punctuation, word delimiter) | ## Training Setup | Detail | Value | |---|---| | **Batch size** | 4 × gradient accumulation 8 = effective batch 32 | | **Learning rate** | 5e-5 with 10% warmup | | **Epochs** | 2.0 (early stop at 18,400 steps; 25 planned …