Ogaal ASR

Ogaal ASR is a Somali automatic speech recognition model from Ogaal Labs. It is a fine-tuned openai/whisper-large-v3 checkpoint trained on approximately 72.1 hours of Somali speech.

Key capabilities

• local CPU or GPU inference
• individual file and folder transcription through the companion toolkit
• browser demo for recording or uploading audio
• developer-ready Transformers integration
• Somali-focused speech-to-text output

Published Somali results

• validation WER: 0.2166
• validation CER: 0.1054
• test WER: 0.2278
• test CER: 0.1186

Training overview

• base model: openai/whisper-large-v3
• architecture: WhisperForConditionalGeneration
• training rows: 39,604
• validation rows: 429
• test rows: 428
• training audio: approximately 72.1 hours
• epochs: 6
• learning rate: 0.0001
• primary language: Somali (so)

A private Ogaal Labs collection contributed roughly 5,000 curated prompts recorded by 19 speakers across varied genders, accents, and speaking styles. English was not part of the training objective.

Transformers quick start

from transformers import pipeline

pipe = pipeline(
    "automatic-speech-recognition",
    model="Ogaal-Labs/Ogaal-ASR",
)
result = pipe("test.wav")
print(result["text"])

Local toolkit

Use the companion GitHub repository for local CPU/GPU inference and the browser demo:

git clone https://github.com/Ogaal-Labs/Ogaal-ASR.git
cd Ogaal-ASR
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
git clone https://huggingface.co/Ogaal-Labs/Ogaal-ASR model

python scripts/infer_somali_asr.py \
  --audio-path /path/to/audio.wav \
  --model-dir model

Requirements: Python 3.10 or newer and ffmpeg on the system path.

Files included

• model.safetensors: fine-tuned Whisper weights
• config.json, generation_config.json, processor_config.json: runtime configuration
• tokenizer.json, tokenizer_config.json: tokenizer assets
• best_val_metrics.json, test_metrics.json: held-out evaluation summaries
• TECHNICAL_BOOK.md, MODEL_SCOPE.md: supporting release documentation

Intended use

• Somali speech transcription
• local and privacy-sensitive transcription workflows
• developer integration for Somali voice products
• evaluation and benchmarking on Somali audio

Limitations

• designed for Somali, not general multilingual transcription
• English was not part of the training objective
• accuracy may vary with accents, noise, microphones, domains, and speaking styles
• users should independently evaluate the model before high-stakes deployment

Links

• GitHub toolkit: https://github.com/Ogaal-Labs/Ogaal-ASR
• Base model: https://huggingface.co/openai/whisper-large-v3
• Ogaal Labs: https://ogaallabs.com/
• Technical book: TECHNICAL_BOOK.md
• Model scope: MODEL_SCOPE.md

License

This model repository is released under the Apache License 2.0. The base openai/whisper-large-v3 model is also distributed under Apache-2.0.