ModelPrésentéExterne

Qwen3 Asr 0.6b Mizonal3 E1 Lus v2026.06

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

Sourcehuggingface.co/andrewbawitlung/qwen3-asr-0.6b-mizonal3-E1-lus-v2026.06/tree/50129a5c1f426d77db33095ef42f98049007ad6e

TypeText modelModèle de baseQwen/Qwen3-ASR-0.6BVersionmodelBClasse

Source du modèle

Type: Text model
Taille du fichier: 1,46 GB
Modèle de base: Qwen/Qwen3-ASR-0.6B
Version: vmodel
Mis à jour: 19 oct. 2026
Téléchargements: 13
Source: Hugging Face

Description de la source

Mizo Automatic Speech Recognition (ASR) Models v3.0

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

qwen3-asr-0.6b-mizonal3-E1-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: 22.1262
Cer: 5.2278
Real Time Factor: 0.0700

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-E1-lus-v2026.06")
model = Qwen2AudioForConditionalGeneration.from_pretrained("andrewbawitlung/qwen3-asr-0.6b-mizonal3-E1-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	0.3531	0.5678	42.06	10.79	1.95e-05	11.44
400	0.73	0.2215	0.3834	30.35	7.45	1.86e-05	9.12
600	1.09	0.1323	0.3356	26.59	6.25	1.76e-05	8.75
800	1.46	0.1034	0.3069	24.98	6.01	1.67e-05	19.88
1000	1.82	0.1023	0.2834	22.98	5.39	1.58e-05	4.03
1200	2.19	0.0753	0.2766	22.37	5.15	1.48e-05	5.62
1400	2.55	0.0566	0.2757	21.13	4.87	1.39e-05	4.72
1600	2.91	0.0573	0.2684	21.40	4.96	1.30e-05	4.88
1800	3.28	0.0370	0.2821	21.37	5.04	1.20e-05	3.75
2000	3.64	0.0337	0.2879	20.84	4.87	1.11e-05	4.03
2200	4.01	0.0390	0.2856	23.75	8.29	1.02e-05	5.44
2400	4.37	0.0189	0.3053	21.12	5.04	9.26e-06	2.84
2600	4.74	0.0178	0.3009	21.28	5.09	8.33e-06	2.69
2800	5.10	0.0101	0.3246	21.28	5.16	7.40e-06	2.14
3000	5.46	0.0122	0.3235	21.41	5.19	6.47e-06	2.41
3200	5.83	0.0102	0.3276	21.17	5.07	5.54e-06	2.19
3400	6.19	0.0059	0.3391	21.34	5.25	4.61e-06	0.97
3600	6.56	0.0074	0.3414	21.40	5.22	3.68e-06	1.15
3800	6.92	0.0084	0.3392	21.04	5.16	2.76e-06	2.12
4000	7.29	0.0079	0.3447	21.31	5.19	1.83e-06	1.74
4200	7.65	0.0082	0.3457	21.46	5.21	8.97e-07	3.44

Sources

1 source

Hugging FacePrincipalVersion 50129a5c1f426d77db33095ef42f98049007ad6e

huggingface.co/andrewbawitlung/qwen3-asr-0.6b-mizonal3-E1-lus-v2026.06/tree/50129a5c1f426d77db33095ef42f98049007ad6e

Vérifié 19 oct.

Artefacts du modèle

1 artefact

model.safetensors

safetensors · 1,46 GB · SHA-256 892185f59009…1993 · Hugging Face

Download file

Extraits de sources

3 extraits

Readme3 aoûthuggingface.co/andrewbawitlung/qwen3-asr-0.6b-mizonal3-E1-lus-v2026.06

--- 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-E1-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: 22.1262 - name: Cer type: cer value: 5.2278 - name: Real Time Factor type: rtf value: 0.0700 --- ![Mizo Automatic Speech Recognition (ASR) Models v3.0](banner.jpg) > **Disclaimer / Notice:** Details for these are in Peer Review and publications of the paper will be made available soon for more details.  # qwen3-asr-0.6b-mizonal3-E1-lus-v2026.06 This model is a fine-tuned version of [Qwen/Qwen3-ASR-0.6B](https://huggingface.co/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: 22.1262 - Cer: 5.2278 - Real Time Factor: 0.0700 ## 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-E1-lus-v2026.06") model = Qwen2AudioForConditionalGeneration.from_pretrained("andrewbawitlung/qwen3-asr-0.6b-mizonal3-E1-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_deco...

Note19 oct.huggingface.co/andrewbawitlung/qwen3-asr-0.6b-mizonal3-E1-lus-v2026.06

Source context: 13 downloads · 0 likes · Pipeline automatic-speech-recognition · Repo andrewbawitlung/qwen3-asr-0.6b-mizonal3-E1-lus-v2026.06

Note3 aoûthuggingface.co/andrewbawitlung/qwen3-asr-0.6b-mizonal3-E1-lus-v2026.06

Source context: 46 downloads · 0 likes · Pipeline automatic-speech-recognition · Repo andrewbawitlung/qwen3-asr-0.6b-mizonal3-E1-lus-v2026.06