Ogaal-Labs/Ogaal-ASR

Ogaal ASR

Ogaal ASR is a Somali automatic speech recognition model released by Ogaal Labs. It is a fine-tuned openai/whisper-large-v3 checkpoint trained on roughly 72.1 hours of Somali speech for Somali transcription workflows.

Why This Release Exists

Ogaal Labs builds local datasets and practical AI tools for Somali and African communities. Ogaal ASR was intentionally trained for Somali speech recognition so the training objective stayed focused on Somali data, Somali transcription quality, and Somali product use cases.

Training Data Overview

This release was trained on roughly 72.1 hours of Somali speech.

A private Ogaal Labs collection pipeline contributed a core part of that effort through roughly 5,000 curated prompts recorded by 19 speakers across varied genders, accents, and speaking styles.

English was not part of the training objective for this release.

Core Metrics

Held-out Somali validation:

• WER: 0.2166
• CER: 0.1054

Held-out Somali test:

• WER: 0.2278
• CER: 0.1186

Model Details

• organization: Ogaal Labs
• model family: Whisper
• base model: openai/whisper-large-v3
• architecture: WhisperForConditionalGeneration
• primary language: Somali (so)
• task: speech-to-text transcription

Training Summary

• train rows: 39604
• validation rows: 429
• test rows: 428
• train hours: 72.1
• epochs: 6.0
• learning rate: 0.0001

Quick Start

from transformers import pipeline

pipe = pipeline("automatic-speech-recognition", model="Ogaal-Labs/Ogaal-ASR")

result = pipe("test.wav")
print(result["text"])

Repository Links

• GitHub repo: https://github.com/Ogaal-Labs/Ogaal-ASR
• Ogaal Labs: https://ogaallabs.com/
• technical book: TECHNICAL_BOOK.md
• model scope note: MODEL_SCOPE.md

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 evaluation and benchmarking
• application integration for Somali ASR workflows

Limitations

• this release was trained for Somali speech recognition, not general multilingual transcription
• English was not part of the training objective
• performance may degrade on domains, accents, or recording conditions that differ from the released evaluation setup

About Ogaal Labs

This release comes from Ogaal Labs, a Somali AI lab focused on local datasets and practical AI tools.

• website: https://ogaallabs.com/
• base model: openai/whisper-large-v3

License

The Hugging Face metadata for this release is currently set to other.