Ogaal CTC

Ogaal CTC is a compact, developer-focused Somali automatic speech recognition system from Ogaal Labs. It combines a 300M-parameter Wav2Vec2/XLSR CTC acoustic model with the published transcript_only language-model decoder for practical local transcription.

Key capabilities

• local CPU or GPU inference
• individual file and folder transcription
• long-audio chunking with speech-region detection
• text, JSON, JSONL, and CSV outputs through the companion toolkit
• bundled decoder matching the published evaluation path

Published Somali results

The bundled transcript_only decoder produced:

• validation WER: 0.2179
• test WER: 0.2114
• test CER: 0.0997

These metrics are for the bundled language-model decoder, not raw greedy CTC decoding.

Training overview

• model family: Wav2Vec2 / XLSR-300M
• architecture: Wav2Vec2ForCTC
• training clips: 39,604
• training audio: approximately 72.1 hours
• 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.

Acoustic-model quick start

from transformers import AutoModelForCTC, AutoProcessor

model_id = "Ogaal-Labs/Ogaal-CTC"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForCTC.from_pretrained(model_id)

This loads the acoustic model for raw or greedy CTC decoding. It does not reproduce the published language-model-decoded metrics by itself.

Published decoder workflow

Use the companion GitHub toolkit for local CPU/GPU inference with the bundled transcript_only decoder:

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

python scripts/infer_ogaal_ctc.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: CTC acoustic model weights
• config.json, processor_config.json, tokenizer_config.json, vocab.json: runtime assets
• decoders/transcript_only/: published language-model decoding assets
• frozen_system.json: fixed decode configuration
• best_val_metrics.json, test_metrics.json, release_metadata.json: evaluation and packaging metadata
• 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
• research and evaluation using the released decode path

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
• confidence values are experimental and should not be treated as calibrated probabilities
• published WER and CER use the bundled transcript_only decoder, not raw greedy decoding

Links

• GitHub toolkit: https://github.com/Ogaal-Labs/Ogaal-CTC
• 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.