You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This model was fine-tuned on non-standard Kenyan English speech data, including speech from individuals with speech-related personal circumstances. By requesting access, you agree to use this model for research, accessibility, or product development purposes only, to cite this work if used in a publication or product, and not to use it in any way that stigmatizes or misrepresents the speakers whose data contributed to this model.

Whisper-small-kenyan-english-nonstandard-v2 (Non-Standard, Noise-Robust)

Fine-tuned from openai/whisper-small on non-standard Kenyan English speech, with waveform-level noise augmentation to improve robustness under real-world Kenyan field deployment conditions (mobile networks, ambient crowd noise, device volume variation).

Developed by Alexander Sheldon Mainye.

Model Details

Property	Value
Base model	`openai/whisper-small`
Total parameters	241,734,912
Trainable parameters	241,734,912 (full fine-tuning)
Language	Kenyan English (`en-KE`)
Task	Automatic Speech Recognition
Domain	Non-standard Kenyan English
Training dataset	`cdli/kenyan_english_nonstandard_speech_v1.0`

Training Report

Performance

Evaluated on the held-out test split of cdli/kenyan_english_nonstandard_speech_v1.0 (926 examples). Clean evaluation uses unaugmented audio; noisy evaluation applies the same waveform augmentation pipeline used during training.

Split	WER	CER
Dev set (clean)	21.65%	13.47%
Test set (clean)	15.77%	9.32%
Test set (noise-augmented)	21.62%	13.61%
Robustness gap (clean → noisy)	5.85 pp	4.29 pp

The 5.85 pp robustness gap represents a 56% relative reduction compared to earlier runs trained at higher augmentation probability (prior best: 13.5 pp gap), while clean test WER remained competitive.

Training Details

Dataset

cdli/kenyan_english_nonstandard_speech_v1.0 — nnon-standard Kenyan English speech covering etiologies.

Split	Raw size	Filtered (<= 30s)
Train	4,378	4,243
Validation	542	542
Test	928	926

Augmentation Pipeline

Waveform-level augmentation was applied during training only (AUGMENT_PROB = 0.4). Evaluation always used clean audio for reproducible WER/CER comparisons.

Augmentation	Parameters	Purpose
Volume perturbation	Gain range 0.7–1.3 (always applied)	Device volume variation
Gaussian noise	Level 0.002–0.015	Ambient crowd noise
GSM codec simulation	8kHz downsample/resample	Mobile network compression
Room reverb	Disabled (`AUG_REVERB = False`)	—

SpecAugment was also applied during training (mask_time_prob = 0.05, mask_feature_prob = 0.05).

Hyperparameters

Hyperparameter	Value
Learning rate	3e-6
LR scheduler	Polynomial decay (`lr_end = 1e-8`, `power = 1`)
Warmup steps	100
Weight decay	0.01
Batch size	32
Max steps	2,000
Early stopping patience	7
Best checkpoint step	1,800
Precision	FP16

Training ran the full 2,000-step budget without early stopping triggering. The best validation WER (21.65%) was achieved at step 1,800, confirming that patience 7 and the polynomial decay schedule were appropriate for this dataset size.

Usage

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch

processor = WhisperProcessor.from_pretrained("smainye/whisper-small-kenyan-english-nonstandard")
model = WhisperForConditionalGeneration.from_pretrained("smainye/whisper-small-kenyan-english-nonstandard")

# Load your audio (must be 16kHz, mono)
# audio is a numpy array or torch tensor of shape (samples,)
inputs = processor(audio, sampling_rate=16000, return_tensors="pt")

with torch.no_grad():
    predicted_ids = model.generate(
        inputs["input_features"],
        language="en",
        task="transcribe"
    )

transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
print(transcription[0])

Notes on intended use

Optimised for non-standard Kenyan English.
Most robust to ambient crowd noise and GSM/mobile codec artifacts (as trained).
Not fine-tuned for standard American or British English; general-purpose WER on those varieties may be lower than openai/whisper-small baseline.
Evaluation was conducted on a noise-augmented test set simulating field conditions; real-world WER will vary based on recording quality and acoustic environment.

Limitations

Dataset size: the training set comprises 4,243 examples. Performance on speaker demographics, dialects, or acoustic conditions not well-represented in cdli/kenyan_english_nonstandard_speech_v1.0 may degrade.
Room reverb not covered: reverb-heavy environments (lecture halls, large rooms) were not part of the augmentation pipeline in this run. Performance in strongly reverberant conditions is untested.
Model capacity: based on whisper-small (241M parameters). Larger Whisper variants may yield lower WER on this domain.
Language: Suited for Kenyan English speakers.

Training Infrastructure

Training was conducted using Modal GPU infrastructure with the Hugging Face transformers Seq2SeqTrainer.

Citation

If you use this model in your work, please cite it as:

@misc{whisper-small-kenyan-english-nonstandard,
  author       = {Mainye, Alexander Sheldon},
  title        = {Whisper Small Fine-tuned on Non-Standard Kenyan English Speech},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/smainye/whisper-small-kenyan-english-nonstandard}
}