qwen3-asr-0.6b-mizonal3-E4-lus-v2026.06

This model is a fine-tuned version of Qwen/Qwen3-ASR-0.6B 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: 22.2146
• Cer: 5.2331
• Real Time Factor: 0.0685

Quick Inference

import torch
import librosa
from transformers import AutoProcessor, Qwen2AudioForConditionalGeneration

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

processor = AutoProcessor.from_pretrained("andrewbawitlung/qwen3-asr-0.6b-mizonal3-E4-lus-v2026.06")
model = Qwen2AudioForConditionalGeneration.from_pretrained("andrewbawitlung/qwen3-asr-0.6b-mizonal3-E4-lus-v2026.06").to(device)

audio, sr = librosa.load("your_audio.wav", sr=16000)

conversation = [
    {"role": "user", "content": [
        {"type": "audio", "audio_url": "your_audio.wav"},
        {"type": "text", "text": "Transcribe the audio:"}
    ]}
]
text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
inputs = processor(text=text, audios=[audio], return_tensors="pt", padding=True)
inputs.input_ids = inputs.input_ids.to(device)

with torch.no_grad():
    generate_ids = model.generate(**inputs, max_length=256)

generate_ids = generate_ids[:, inputs.input_ids.size(1):]
transcription = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[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/qwen3-asr-0.6b-mizonal3-E1-lus-v2026.06
E2 (Noise)	andrewbawitlung/qwen3-asr-0.6b-mizonal3-E2-lus-v2026.06
E3 (Speed)	andrewbawitlung/qwen3-asr-0.6b-mizonal3-E3-lus-v2026.06
E4 (SpecAug)	andrewbawitlung/qwen3-asr-0.6b-mizonal3-E4-lus-v2026.06
E5 (Combined)	andrewbawitlung/qwen3-asr-0.6b-mizonal3-E5-lus-v2026.06

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• 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
200	0.36	0.3447	0.5944	43.21	11.49	1.95e-05	27.88
400	0.73	0.2183	0.3971	32.67	8.08	1.86e-05	9.94
600	1.09	0.1382	0.3338	27.18	6.40	1.76e-05	9.44
800	1.46	0.0984	0.3060	24.81	5.92	1.67e-05	4.19
1000	1.82	0.1005	0.2859	23.03	5.36	1.58e-05	4.47
1200	2.19	0.0763	0.2803	21.74	5.16	1.48e-05	6.12
1400	2.55	0.0566	0.2762	21.13	4.91	1.39e-05	4.47
1600	2.91	0.0576	0.2692	21.14	4.81	1.30e-05	4.25
1800	3.28	0.0401	0.2816	20.85	4.85	1.20e-05	3.98
2000	3.64	0.0317	0.2860	20.30	4.69	1.11e-05	6.34
2200	4.01	0.0396	0.2856	20.52	4.71	1.02e-05	4.41
2400	4.37	0.0219	0.3018	20.54	4.82	9.26e-06	4.47
2600	4.74	0.0200	0.2963	20.79	4.90	8.33e-06	6.00
2800	5.10	0.0109	0.3231	20.76	4.80	7.40e-06	2.19
3000	5.46	0.0113	0.3205	20.75	4.85	6.47e-06	2.75
3200	5.83	0.0100	0.3252	20.54	4.78	5.54e-06	1.66
3400	6.19	0.0079	0.3390	21.07	4.99	4.61e-06	0.77
3600	6.56	0.0084	0.3384	20.97	4.97	3.68e-06	1.29
3800	6.92	0.0083	0.3375	20.72	4.85	2.76e-06	2.17
4000	7.29	0.0087	0.3430	20.73	4.89	1.83e-06	2.41
4200	7.65	0.0070	0.3444	20.72	4.91	8.97e-07	3.73