README.md

metadata

language:
  - lus
license: apache-2.0
base_model: Qwen/Qwen3-ASR-0.6B
tags:
  - generated_from_trainer
datasets:
  - andrewbawitlung/mizonal-v3
metrics:
  - wer
  - cer
model-index:
  - name: qwen3-asr-0.6b-mizonal3-E2-lus-v2026.06
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: MiZonal v3
          type: andrewbawitlung/mizonal-v3
          config: default
          split: test
        metrics:
          - name: Wer
            type: wer
            value: 19.4954
          - name: Cer
            type: cer
            value: 4.4174
          - name: Real Time Factor
            type: rtf
            value: 0.066

qwen3-asr-0.6b-mizonal3-E2-lus-v2026.06

This model is a fine-tuned version of Qwen/Qwen3-ASR-0.6B on the MiZonal v3 dataset. Note: ~1 hour of conversational speech was added to this dataset version.

It achieves the following results on the evaluation set:

• Wer: 19.4954
• Cer: 4.4174
• Real Time Factor: 0.0660

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-E2-lus-v2026.06")
model = Qwen2AudioForConditionalGeneration.from_pretrained("andrewbawitlung/qwen3-asr-0.6b-mizonal3-E2-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	0.18	0.45	0.75	0.53	0.16	1.99e-05	14.50
400	0.36	0.27	0.40	0.32	0.08	1.95e-05	10.81
600	0.55	0.17	0.34	0.28	0.06	1.90e-05	7.56
800	0.73	0.21	0.30	0.25	0.06	1.86e-05	5.78
1000	0.91	0.14	0.29	0.24	0.05	1.81e-05	7.38
1200	1.09	0.09	0.27	0.23	0.05	1.76e-05	5.66
1400	1.28	0.07	0.26	0.22	0.05	1.72e-05	3.84
1600	1.46	0.08	0.26	0.21	0.05	1.67e-05	4.44
1800	1.64	0.08	0.25	0.20	0.05	1.62e-05	8.81
2000	1.82	0.06	0.25	0.20	0.05	1.58e-05	4.59
2200	2.00	0.05	0.24	0.20	0.05	1.53e-05	3.84
2400	2.19	0.04	0.26	0.20	0.05	1.48e-05	5.66
2600	2.37	0.03	0.26	0.19	0.05	1.44e-05	3.44
2800	2.55	0.03	0.27	0.19	0.04	1.39e-05	7.22
3000	2.73	0.03	0.26	0.20	0.05	1.34e-05	4.22
3200	2.91	0.05	0.26	0.19	0.04	1.30e-05	4.34
3400	3.10	0.02	0.28	0.19	0.04	1.25e-05	4.47
3600	3.28	0.01	0.29	0.19	0.04	1.20e-05	2.52
3800	3.46	0.02	0.29	0.19	0.04	1.16e-05	2.94
4000	3.64	0.01	0.29	0.19	0.04	1.11e-05	2.06
4200	3.83	0.01	0.30	0.19	0.04	1.07e-05	4.19
4400	4.01	0.01	0.30	0.19	0.04	1.02e-05	5.16
4600	4.19	0.01	0.31	0.19	0.04	9.72e-06	1.23
4800	4.37	0.00	0.31	0.19	0.04	9.26e-06	1.21
5000	4.55	0.01	0.32	0.19	0.04	8.79e-06	0.94
5200	4.74	0.00	0.32	0.19	0.04	8.33e-06	1.05
5400	4.92	0.00	0.32	0.19	0.04	7.86e-06	0.99
5600	5.10	0.00	0.33	0.19	0.04	7.40e-06	0.46
5800	5.28	0.00	0.33	0.19	0.04	6.94e-06	0.70
6000	5.46	0.00	0.33	0.19	0.04	6.47e-06	0.84
6200	5.65	0.00	0.34	0.19	0.04	6.01e-06	0.60
6400	5.83	0.00	0.34	0.19	0.04	5.54e-06	2.16
6600	6.01	0.00	0.34	0.19	0.04	5.08e-06	0.46
6800	6.19	0.00	0.34	0.19	0.04	4.61e-06	0.35
7000	6.38	0.00	0.34	0.19	0.04	4.15e-06	0.38
7200	6.56	0.00	0.34	0.19	0.04	3.68e-06	0.61
7400	6.74	0.00	0.34	0.19	0.04	3.22e-06	0.41
7600	6.92	0.00	0.34	0.19	0.04	2.75e-06	0.35
7800	7.10	0.00	0.34	0.19	0.04	2.29e-06	0.43
8000	7.29	0.00	0.34	0.19	0.04	1.82e-06	0.52
8200	7.47	0.00	0.34	0.19	0.04	1.36e-06	0.28
8400	7.65	0.00	0.34	0.19	0.04	8.95e-07	0.33
8600	7.83	0.00	0.34	0.19	0.04	4.30e-07	0.74