xlsr-1b-mizonal3-E4-lus-v2026.06

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b 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: 20.9777
• Cer: 4.0982
• Real Time Factor: 0.0040

Quick Inference

import torch
import librosa
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC

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

processor = Wav2Vec2Processor.from_pretrained("andrewbawitlung/xlsr-1b-mizonal3-E4-lus-v2026.06")
model = Wav2Vec2ForCTC.from_pretrained("andrewbawitlung/xlsr-1b-mizonal3-E4-lus-v2026.06").to(device)

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

with torch.no_grad():
    logits = model(input_values).logits

predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)[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/xlsr-1b-mizonal3-E1-lus-v2026.06
E2 (Noise)	andrewbawitlung/xlsr-1b-mizonal3-E2-lus-v2026.06
E3 (Speed)	andrewbawitlung/xlsr-1b-mizonal3-E3-lus-v2026.06
E4 (SpecAug)	andrewbawitlung/xlsr-1b-mizonal3-E4-lus-v2026.06
E5 (Combined)	andrewbawitlung/xlsr-1b-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	0.91	1.0146	0.4622	49.46	12.67	1.49e-04	1.09
500	1.82	0.7421	0.3725	41.69	10.20	2.99e-04	1.04
750	2.73	0.5903	0.2855	34.39	7.92	2.56e-04	0.78
1000	3.64	0.4915	0.2410	28.43	6.38	2.12e-04	1.05
1250	4.55	0.3748	0.1973	23.82	5.13	1.68e-04	0.65
1500	5.46	0.3166	0.1746	21.37	4.50	1.24e-04	0.73
1750	6.36	0.2762	0.1640	21.24	4.35	7.96e-05	1.26
2000	7.27	0.2495	0.1607	18.83	3.88	3.55e-05	0.89