| language: | |
| - sn | |
| tags: | |
| - automatic-speech-recognition | |
| - whisper | |
| - audio | |
| - multi-stage-finetuning | |
| model-index: | |
| - name: Whisper Small Shona - Verdi AI OS Stage 1c | |
| results: | |
| - task: | |
| name: Automatic Speech Recognition | |
| type: automatic-speech-recognition | |
| dataset: | |
| name: google/fleurs + manassehzw/shona-bible-bdsc-aligned (Chunk 2) | |
| type: google/fleurs | |
| args: sn_zw | |
| metrics: | |
| - name: Test WER | |
| type: wer | |
| value: 26.42 | |
| pipeline_tag: automatic-speech-recognition | |
| library_name: transformers | |
| # 🌾 Verdi AI OS: Whisper Small Shona ASR (Stage 1c Completed) | |
| This repository hosts a completely standalone, fully compiled model optimized natively for the Shona language (`sn-ZW`). It forms the core Speech-to-Text foundation for **Verdi AI OS**—an agricultural operating system built for logistics metrics automation and voice marketplace tracking. | |
| ## 📊 Training Progression Metrics | |
| - **Stage 1a (FLEURS Baseline):** Dropped baseline WER to **37.15%** | |
| - **Stage 1b (Bible Chunk 1):** Dropped WER to **29.39%** | |
| - **Stage 1c (Bible Chunk 2 - 6,000 Rows):** Finished at **26.42%** WER with an un-normalized training loss of **0.079** | |