MosesJoshuaCoker/SpeechtoText

from huggingface_hub import ModelCard

card = ModelCard("""--- language: - en - kri tags: - whisper - speech-recognition - krio - sierra-leone - automatic-speech-recognition license: mit datasets: - MosesJoshuaCoker/30_hours_krio_voice metrics: - wer - cer base_model: openai/whisper-small pipeline_tag: automatic-speech-recognition

Whisper Small — Krio Speech Recognition

Fine-tuned openai/whisper-small for Krio (Sierra Leone Creole) automatic speech recognition.

Performance

Metric	Score
WER	5.29%
CER	2.29%

Trained on 30 hours of Krio speech across 6,126 samples.

Usage

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import librosa, torch

processor = WhisperProcessor.from_pretrained("MosesJoshuaCoker/SpeechtoText")
model = WhisperForConditionalGeneration.from_pretrained("MosesJoshuaCoker/SpeechtoText")

def transcribe(audio_path):
    audio, sr = librosa.load(audio_path, sr=16000)
    inputs = processor.feature_extractor(
        audio, sampling_rate=16000, return_tensors="pt"
    ).input_features
    with torch.no_grad():
        predicted_ids = model.generate(
            inputs, language="english", task="transcribe", num_beams=4
        )
    return processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]

print(transcribe("your_krio_audio.wav"))

Training Details

Setting	Value
Base model	openai/whisper-small
Training samples	6,126
Eval samples	681
Epochs	5
Batch size	8
Learning rate	1e-5
GPU	NVIDIA RTX PRO 6000 Blackwell

Dataset

MosesJoshuaCoker/30_hours_krio_voice

About Krio

Krio is an English-based creole language spoken by approximately 4 million people in Sierra Leone. It is the country's lingua franca. This is believed to be one of the first publicly available Krio ASR models. """)

card.push_to_hub(repo_name) print("✓ Model card uploaded — no more YAML warning")