Automatic Speech Recognition
Transformers
Safetensors
Shona
whisper
audio
multi-stage-finetuning
Eval Results (legacy)
Instructions to use Code-Dev/whisper-small-shona-v2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Code-Dev/whisper-small-shona-v2 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("automatic-speech-recognition", model="Code-Dev/whisper-small-shona-v2")# Load model directly from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq processor = AutoProcessor.from_pretrained("Code-Dev/whisper-small-shona-v2") model = AutoModelForSpeechSeq2Seq.from_pretrained("Code-Dev/whisper-small-shona-v2", device_map="auto") - Notebooks
- Google Colab
- Kaggle
πΎ Verdi AI OS: Whisper Small Shona ASR (Stage 1b Finished)
This repository hosts a highly specialized fine-tuned checkpoint of OpenAI's Whisper Small optimized natively for the Shona language (sn-ZW). It forms the core Speech-to-Text foundation for Verdi AI OSβan operating system built for African agriculture, marketplace tracking, and logistics metrics automation.
π Training Progression Metrics
- Stage 1a (FLEURS Baseline): Dropped baseline WER to 37.15%
- Stage 1b (Studio Bible Vocals - 6,000 Rows): Finished at 29.39% WER with an ultra-low Validation Loss of 0.234
π» Technical Configurations
- Architecture Base: Whisper Small (244M parameters)
- Batch Size: 16
- Optimization Target: Word Error Rate (WER) using greedy search decoding paths
- Next Infrastructure Phase: Stage 1c Volume Extension (Google Waxal Shona Corpus)
- Downloads last month
- -
Evaluation results
- Test WER on google/fleurs + manassehzw/shona-bible-bdsc-alignedself-reported29.390