This model is a fine-tuned version of openai/whisper-medium on the MiZonal v3.0 dataset. Note: ~1 hour of conversational speech was added to this dataset version.

It achieves the following results on the evaluation set:

Wer: 21.7728
Cer: 7.3593
Real Time Factor: 0.0520

Quick Inference

python

import torch
import librosa
from transformers import WhisperProcessor, WhisperForConditionalGeneration

device = "cuda" if torch.cuda.is_available() else "cpu"

processor = WhisperProcessor.from_pretrained("andrewbawitlung/whisper-medium-mizonal3-E2-lus-v2026.06")
model = WhisperForConditionalGeneration.from_pretrained("andrewbawitlung/whisper-medium-mizonal3-E2-lus-v2026.06").to(device)

audio, sr = librosa.load("your_audio.wav", sr=16000)
input_features = processor(audio, sampling_rate=16000, return_tensors="pt").input_features.to(device)

with torch.no_grad():
    predicted_ids = model.generate(input_features, max_new_tokens=256)

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

Model description

Experiment Configurations

This repository is part of a series of experiments. The different configurations are:

E1 (Baseline): Standard training configuration.
E2 (Noise): Training with background noise augmentation.
E3 (Speed): Training with speed perturbation augmentation.
E4 (SpecAug): Training with SpecAugment (time and frequency masking).
E5 (Combined): Training with a combination of all augmentations.

All Models in this Family

Experiment	Hugging Face Repository
E1 (Baseline)	andrewbawitlung/whisper-medium-mizonal3-E1-lus-v2026.06
E2 (Noise)	andrewbawitlung/whisper-medium-mizonal3-E2-lus-v2026.06
E3 (Speed)	andrewbawitlung/whisper-medium-mizonal3-E3-lus-v2026.06
E4 (SpecAug)	andrewbawitlung/whisper-medium-mizonal3-E4-lus-v2026.06
E5 (Combined)	andrewbawitlung/whisper-medium-mizonal3-E5-lus-v2026.06

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0003
train_batch_size: 8
eval_batch_size: 8
seed: 42
optimizer: OptimizerNames.ADAMW_TORCH_FUSED
lr_scheduler_type: SchedulerType.LINEAR
num_epochs: 8

Training results

step	epoch	train_loss	eval_loss	eval_wer	eval_cer	learning_rate	grad_norm
250	0.46	0.6866	0.6110	34.26	12.72	1.49e-04	7.83
500	0.91	0.8029	0.8721	63.79	36.05	2.99e-04	20.60
750	1.37	0.6327	0.7277	36.49	17.06	2.81e-04	5.54
1000	1.82	0.4858	0.6910	34.48	14.43	2.62e-04	3.86
1250	2.28	0.3171	0.6293	30.86	12.28	2.42e-04	3.36
1500	2.73	0.2640	0.6324	30.76	11.94	2.23e-04	3.04
1750	3.19	0.1781	0.6204	35.20	15.72	2.04e-04	2.49
2000	3.64	0.1584	0.6050	28.89	11.35	1.84e-04	2.56
2250	4.10	0.0972	0.6102	28.05	11.92	1.65e-04	1.73
2500	4.55	0.0894	0.5966	25.34	9.41	1.46e-04	1.68
2750	5.01	0.0659	0.6184	24.96	9.42	1.27e-04	0.96
3000	5.46	0.0524	0.6331	25.60	10.95	1.07e-04	1.43
3250	5.92	0.0345	0.6023	23.26	8.42	8.81e-05	0.72
3500	6.38	0.0192	0.5784	24.83	10.80	6.88e-05	0.25
3750	6.83	0.0097	0.6011	21.91	8.17	4.96e-05	0.13
4000	7.29	0.0038	0.5964	20.29	7.09	3.03e-05	0.12
4250	7.74	0.0020	0.6115	19.98	6.81	1.10e-05	0.01