andrewbawitlung/qwen3-asr-0.6b-mizonal3-E5-lus-v2026.06

[image]

Disclaimer / Notice: Details for these are in Peer Review and publications of the paper will be made available soon for more details.

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

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

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 2e-05
• —trainbatchsize: 16
• —evalbatchsize: 8
• —seed: 42
• —optimizer: OptimizerNames.ADAMWTORCHFUSED
• —lrschedulertype: SchedulerType.LINEAR
• —num_epochs: 8

Training results