README.md

metadata

language:
  - lus
license: apache-2.0
pipeline_tag: automatic-speech-recognition
base_model: Qwen/Qwen3-ASR-0.6B
tags:
  - generated_from_trainer
datasets:
  - andrewbawitlung/MiZonal-v3.0
metrics:
  - wer
  - cer
model-index:
  - name: qwen3-asr-0.6b-mizonal3-E5-lus-v2026.06
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: MiZonal v3.0
          type: andrewbawitlung/MiZonal-v3.0
          config: default
          split: test
        metrics:
          - name: Wer
            type: wer
            value: 18.6414
          - name: Cer
            type: cer
            value: 4.2134
          - name: Real Time Factor
            type: rtf
            value: 0.0718

qwen3-asr-0.6b-mizonal3-E5-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: 18.6414
• Cer: 4.2134
• Real Time Factor: 0.0718

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-E5-lus-v2026.06")
model = Qwen2AudioForConditionalGeneration.from_pretrained("andrewbawitlung/qwen3-asr-0.6b-mizonal3-E5-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	0.09	1.9158	1.5851	76.79	25.27	1.13e-05	90.00
400	0.18	0.2754	0.5047	37.90	9.54	1.99e-05	8.38
600	0.27	0.2056	0.3755	30.17	7.52	1.97e-05	7.00
800	0.36	0.1508	0.3205	25.75	6.06	1.95e-05	5.41
1000	0.46	0.1193	0.2937	24.42	5.88	1.92e-05	7.34
1200	0.55	0.1207	0.2833	23.31	5.47	1.90e-05	5.91
1400	0.64	0.1156	0.2642	22.26	5.12	1.88e-05	6.03
1600	0.73	0.0874	0.2507	20.53	4.72	1.86e-05	3.44
1800	0.82	0.0909	0.2488	19.69	4.73	1.83e-05	5.31
2000	0.91	0.0897	0.2517	20.01	4.69	1.81e-05	5.47
2200	1.00	0.0499	0.2411	18.98	4.30	1.79e-05	4.53
2400	1.09	0.0346	0.2440	18.28	4.21	1.76e-05	5.09
2600	1.18	0.0506	0.2463	19.26	4.44	1.74e-05	5.19
2800	1.28	0.0384	0.2533	19.08	4.43	1.72e-05	4.81
3000	1.37	0.0339	0.2554	18.82	4.26	1.69e-05	3.86
3200	1.46	0.0280	0.2577	18.61	4.33	1.67e-05	4.31
3400	1.55	0.0316	0.2543	18.33	4.22	1.65e-05	6.47
3600	1.64	0.0226	0.2546	18.08	4.13	1.62e-05	1.62
3800	1.73	0.0278	0.2602	18.05	4.18	1.60e-05	3.23
4000	1.82	0.0200	0.2637	18.03	4.14	1.58e-05	3.69
4200	1.91	0.0190	0.2629	17.76	4.00	1.55e-05	5.22
4400	2.00	0.0121	0.2688	18.23	4.11	1.53e-05	3.03
4600	2.09	0.0099	0.2832	18.19	4.18	1.51e-05	3.30
4800	2.19	0.0117	0.2814	18.17	4.10	1.48e-05	6.50
5000	2.28	0.0093	0.2842	17.73	4.23	1.46e-05	3.62
5200	2.37	0.0089	0.2876	18.03	4.21	1.44e-05	1.44
5400	2.46	0.0107	0.2814	17.81	4.16	1.41e-05	1.63
5600	2.55	0.0079	0.2831	17.24	3.92	1.39e-05	1.46
5800	2.64	0.0064	0.2927	17.58	3.95	1.37e-05	4.53
6000	2.73	0.0111	0.2926	17.63	3.98	1.34e-05	7.41
6200	2.82	0.0050	0.2963	17.72	4.12	1.32e-05	1.07
6400	2.91	0.0050	0.2935	17.27	3.96	1.30e-05	0.96
6600	3.01	0.0025	0.2995	17.06	3.90	1.27e-05	0.81
6800	3.10	0.0031	0.3090	17.32	3.94	1.25e-05	0.84
7000	3.19	0.0036	0.3094	17.17	3.92	1.23e-05	0.48
7200	3.28	0.0055	0.3113	17.28	3.99	1.20e-05	1.09
7400	3.37	0.0035	0.3116	17.32	3.95	1.18e-05	0.30
7600	3.46	0.0049	0.3128	16.95	3.90	1.16e-05	1.99
7800	3.55	0.0036	0.3111	17.19	4.00	1.13e-05	2.77
8000	3.64	0.0021	0.3138	17.22	3.90	1.11e-05	0.37
8200	3.73	0.0039	0.3096	17.13	3.84	1.09e-05	0.62
8400	3.83	0.0031	0.3160	17.15	3.90	1.07e-05	1.31
8600	3.92	0.0017	0.3132	17.40	3.91	1.04e-05	0.56
8800	4.01	0.0016	0.3194	17.06	3.91	1.02e-05	0.25
9000	4.10	0.0024	0.3214	17.25	3.90	9.95e-06	0.17
9200	4.19	0.0017	0.3217	17.32	3.96	9.72e-06	2.45
9400	4.28	0.0012	0.3262	17.29	3.95	9.49e-06	0.27
9600	4.37	0.0018	0.3231	17.26	3.92	9.26e-06	0.26
9800	4.46	0.0013	0.3260	17.61	4.01	9.03e-06	0.51
10000	4.55	0.0012	0.3224	17.25	3.91	8.79e-06	0.17
10200	4.64	0.0016	0.3262	17.41	3.98	8.56e-06	0.66
10400	4.74	0.0012	0.3246	17.46	3.97	8.33e-06	0.23
10600	4.83	0.0011	0.3277	17.25	3.92	8.10e-06	0.16
10800	4.92	0.0020	0.3261	17.47	3.97	7.86e-06	2.86
11000	5.01	0.0011	0.3268	17.26	3.94	7.63e-06	0.59
11200	5.10	0.0017	0.3292	17.27	3.93	7.40e-06	0.46
11400	5.19	0.0011	0.3290	17.31	3.97	7.17e-06	0.14
11600	5.28	0.0011	0.3295	17.16	3.94	6.93e-06	0.27
11800	5.37	0.0011	0.3300	17.25	3.91	6.70e-06	0.44
12000	5.46	0.0011	0.3300	17.18	3.95	6.47e-06	0.36
12200	5.56	0.0012	0.3291	17.29	3.96	6.24e-06	0.29
12400	5.65	0.0011	0.3295	17.25	3.92	6.00e-06	0.28
12600	5.74	0.0012	0.3286	17.27	3.88	5.77e-06	0.45
12800	5.83	0.0013	0.3303	17.20	3.92	5.54e-06	0.24
13000	5.92	0.0018	0.3291	17.24	3.95	5.31e-06	0.55
13200	6.01	0.0011	0.3308	17.22	3.92	5.08e-06	0.18
13400	6.10	0.0012	0.3307	17.29	3.92	4.84e-06	0.30
13600	6.19	0.0010	0.3317	17.25	3.97	4.61e-06	0.52
13800	6.28	0.0012	0.3317	17.21	3.96	4.38e-06	0.14
14000	6.38	0.0012	0.3307	17.21	3.92	4.15e-06	0.44
14200	6.47	0.0012	0.3311	17.26	3.93	3.91e-06	0.34
14400	6.56	0.0011	0.3316	17.24	3.99	3.68e-06	0.19
14600	6.65	0.0010	0.3308	17.30	3.94	3.45e-06	0.25
14800	6.74	0.0013	0.3311	17.22	3.92	3.22e-06	0.27
15000	6.83	0.0010	0.3320	17.15	3.95	2.98e-06	0.17
15200	6.92	0.0010	0.3307	17.28	3.93	2.75e-06	0.18
15400	7.01	0.0011	0.3308	17.32	3.91	2.52e-06	0.22
15600	7.10	0.0011	0.3305	17.26	3.92	2.29e-06	0.21
15800	7.19	0.0011	0.3317	17.29	3.92	2.06e-06	0.33
16000	7.29	0.0011	0.3316	17.22	3.93	1.82e-06	0.43
16200	7.38	0.0010	0.3312	17.32	3.93	1.59e-06	0.21
16400	7.47	0.0011	0.3314	17.30	3.94	1.36e-06	0.61
16600	7.56	0.0010	0.3313	17.38	3.93	1.13e-06	0.18
16800	7.65	0.0010	0.3313	17.29	3.92	8.93e-07	0.21
17000	7.74	0.0010	0.3310	17.19	3.90	6.61e-07	0.23
17200	7.83	0.0009	0.3306	17.41	3.98	4.29e-07	0.15
17400	7.92	0.0011	0.3309	17.26	3.91	1.96e-07	0.18