Dama12/whisper-small-moore

Whisper Small Mooré (Mooré Speech-to-Text)

This is a fine-tuned version of openai/whisper-small for Mooré (Moore) automatic speech recognition (ASR), a Gur language spoken in Burkina Faso by over 8 million people.

Model Details

• Developed by: Soumana Dama — Full Stack Developer & AI Engineer — Founder & Lead AI Engineer at Scoinvestigator AI
• Model type: Whisper (Speech-to-Text / Automatic Speech Recognition)
• Language(s) (NLP): Mooré (ISO 639-3: mos)
• Finetuned from model: openai/whisper-small
• License: Apache 2.0
• Developer Contacts:
  • LinkedIn: Soumana Dama on LinkedIn
  • Scoinvestigator AI: scoinvestigator.com
  • GitHub: Damasoumana1
  • Portfolio: soumanadama.netlify.app
  • Email: soumanadama93@gmail.com

---

Uses

Direct Use

This model is designed to transcribe spoken Mooré audio into written Mooré text. It can be integrated into educational tools (like multilingual AI tutors), voice assistants, and accessibility software.

Out-of-Scope Use

This model is not designed for translating audio directly into other languages (except when paired with translation pipelines like NLLB-200), nor is it suitable for high-risk clinical or legal transcriptions without human supervision.

---

Training Details

Training Data

The model was fine-tuned on a curated dataset of 23,437 audio/text pairs in the Mooré language, carefully processed and structured.

Training Procedure & Parameters

• Hardware: 2x Tesla T4 GPUs (Kaggle Compute)
• Inference/Training configuration:
  • Lazy loading of spectrograms via custom PyTorch Dataset loader to prevent system RAM bottlenecks.
  • Training completed in multiple runs utilizing checkpoints (resumed from checkpoint-3000 to the final state).
• Optimization settings:
  • FP16 precision enabled.
  • Intermediate evaluation loss tracked to select the best checkpoint without triggering GPU Out-Of-Memory crashes.

---

How to Get Started with the Model

Use the code below to load and run the model in Python:

import torch
from transformers import AutoProcessor, AutoModelForConditionalGeneration

# Load model and processor
processor = AutoProcessor.from_pretrained("Dama12/whisper-small-moore")
model = AutoModelForConditionalGeneration.from_pretrained("Dama12/whisper-small-moore")

# Use a pipeline for simple inference
from transformers import pipeline
asr_pipeline = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    device="cuda" if torch.cuda.is_available() else "cpu"
)

# Transcribe an audio file
# Note: Ensure the audio file is sampled at 16kHz
transcription = asr_pipeline("path_to_your_audio_file.wav")
print(transcription["text"])