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: 18.9850
Cer: 4.2170
Real Time Factor: 0.0697

Quick Inference

python

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-E3-lus-v2026.06")
model = Qwen2AudioForConditionalGeneration.from_pretrained("andrewbawitlung/qwen3-asr-0.6b-mizonal3-E3-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	0.12	0.7348	1.0950	66.07	21.06	1.51e-05	33.00
400	0.24	0.2043	0.4573	36.26	9.28	1.98e-05	9.00
600	0.36	0.1685	0.3415	27.38	6.61	1.95e-05	8.50
800	0.49	0.1129	0.3107	24.91	5.80	1.92e-05	5.50
1000	0.61	0.1239	0.2829	22.88	5.29	1.89e-05	4.62
1200	0.73	0.0943	0.2676	21.77	5.23	1.86e-05	5.00
1400	0.85	0.0740	0.2625	21.12	4.89	1.82e-05	4.53
1600	0.97	0.0765	0.2584	21.08	4.94	1.79e-05	3.56
1800	1.09	0.0375	0.2681	20.73	4.72	1.76e-05	4.09
2000	1.21	0.0411	0.2577	19.94	4.56	1.73e-05	1.96
2200	1.34	0.0347	0.2623	19.56	4.43	1.70e-05	5.03
2400	1.46	0.0325	0.2672	19.84	4.52	1.67e-05	5.09
2600	1.58	0.0306	0.2619	19.00	4.37	1.64e-05	2.80
2800	1.70	0.0282	0.2636	19.22	4.31	1.61e-05	2.92
3000	1.82	0.0266	0.2612	18.38	4.24	1.58e-05	2.70
3200	1.94	0.0183	0.2616	19.08	4.44	1.55e-05	4.00
3400	2.06	0.0107	0.2810	22.51	8.80	1.51e-05	2.66
3600	2.19	0.0158	0.2766	18.65	4.23	1.48e-05	2.73
3800	2.31	0.0114	0.2837	18.78	4.33	1.45e-05	1.63
4000	2.43	0.0135	0.2862	18.95	4.31	1.42e-05	6.94
4200	2.55	0.0123	0.2905	18.46	4.24	1.39e-05	4.53
4400	2.67	0.0073	0.2967	18.55	4.31	1.36e-05	2.66
4600	2.79	0.0073	0.3002	18.21	4.28	1.33e-05	2.28
4800	2.91	0.0074	0.3017	18.03	4.21	1.30e-05	1.87
5000	3.04	0.0042	0.3065	18.30	4.29	1.27e-05	0.58
5200	3.16	0.0034	0.3183	18.12	4.16	1.24e-05	0.80
5400	3.28	0.0027	0.3178	18.19	4.22	1.20e-05	1.76
5600	3.40	0.0039	0.3146	18.23	4.24	1.17e-05	0.49
5800	3.52	0.0024	0.3200	18.22	4.28	1.14e-05	0.29
6000	3.64	0.0015	0.3197	17.88	4.23	1.11e-05	0.43
6200	3.76	0.0021	0.3237	17.94	4.30	1.08e-05	0.99
6400	3.89	0.0020	0.3280	17.95	4.25	1.05e-05	0.85
6600	4.01	0.0013	0.3216	17.83	4.16	1.02e-05	0.29
6800	4.13	0.0014	0.3293	17.89	4.14	9.88e-06	0.28
7000	4.25	0.0011	0.3299	18.00	4.15	9.57e-06	0.18
7200	4.37	0.0014	0.3329	17.83	4.10	9.26e-06	0.31
7400	4.49	0.0019	0.3337	17.90	4.11	8.95e-06	2.44
7600	4.61	0.0013	0.3339	17.78	4.10	8.64e-06	0.44
7800	4.74	0.0014	0.3368	17.94	4.16	8.33e-06	0.31
8000	4.86	0.0017	0.3385	18.04	4.20	8.02e-06	0.25
8200	4.98	0.0014	0.3357	17.88	4.17	7.71e-06	0.37
8400	5.10	0.0009	0.3368	17.97	4.23	7.40e-06	0.24
8600	5.22	0.0009	0.3375	17.87	4.28	7.09e-06	0.14
8800	5.34	0.0012	0.3406	17.96	4.18	6.78e-06	0.16
9000	5.46	0.0010	0.3402	17.88	4.24	6.47e-06	0.30
9200	5.59	0.0009	0.3400	17.93	4.23	6.16e-06	0.21
9400	5.71	0.0013	0.3426	17.92	4.25	5.85e-06	0.23
9600	5.83	0.0012	0.3423	17.92	4.14	5.54e-06	0.23
9800	5.95	0.0010	0.3415	17.91	4.18	5.23e-06	0.18
10000	6.07	0.0010	0.3428	17.87	4.25	4.92e-06	0.22
10200	6.19	0.0012	0.3426	17.89	4.21	4.61e-06	0.23
10400	6.31	0.0010	0.3425	17.77	4.08	4.30e-06	0.20
10600	6.44	0.0010	0.3431	17.91	4.28	3.99e-06	0.13
10800	6.56	0.0011	0.3425	18.05	4.28	3.68e-06	0.22
11000	6.68	0.0011	0.3433	17.97	4.26	3.37e-06	0.32
11200	6.80	0.0012	0.3430	17.94	4.29	3.06e-06	0.27
11400	6.92	0.0023	0.3430	18.03	4.26	2.75e-06	0.19
11600	7.04	0.0010	0.3430	17.96	4.30	2.44e-06	0.26
11800	7.16	0.0010	0.3426	18.04	4.25	2.13e-06	0.22
12000	7.29	0.0010	0.3428	18.05	4.27	1.82e-06	0.17
12200	7.41	0.0010	0.3425	17.85	4.23	1.51e-06	0.43
12400	7.53	0.0009	0.3433	17.95	4.24	1.20e-06	0.21
12600	7.65	0.0009	0.3424	17.98	4.23	8.94e-07	0.39
12800	7.77	0.0010	0.3431	17.97	4.24	5.84e-07	0.24
13000	7.89	0.0009	0.3430	18.08	4.31	2.74e-07	0.60