Omni Asr Ctc 300m v2 Mizonal3 E4 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.
Model source
- Kind
- Other
- File size
- 1.21 GB
- Base model
- facebook/omniASR-CTC-300M
- Version
- vSdp 00
- Updated
- Oct 19, 2026
- Source
- Hugging Face
Source description
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.
omni-asr-ctc-300m-v2-mizonal3-E4-lus-v2026.06
This model is a fine-tuned version of facebook/omniASR-CTC-300M on the MiZonal v3.0 dataset.
It achieves the following results on the evaluation set:
- Wer: 18.0524
- Cer: 3.2913
- Real Time Factor: 0.0065
Quick Inference
# Google Colab Ready Quick Inference
# 1. Install dependencies:
!pip install huggingface_hub omnilingual-asr torchaudio
# 2. Download the model checkpoint directly from Hugging Face:
import torch
from huggingface_hub import hf_hub_download
model_path = hf_hub_download(
repo_id="andrewbawitlung/omni-asr-ctc-300m-v2-mizonal3-E4-lus-v2026.06",
filename="sdp_00.pt"
)
# 3. Load the base model pipeline and inject fine-tuned weights:
from omnilingual_asr.models.inference.pipeline import ASRInferencePipeline
print("Loading base model pipeline...")
pipeline = ASRInferencePipeline(model_card="omniASR_CTC_300M_v2")
print("Loading fine-tuned checkpoint weights...")
checkpoint = torch.load(model_path, map_location="cpu")
state_dict = checkpoint['model'] if 'model' in checkpoint else checkpoint
pipeline.model.load_state_dict(state_dict, strict=False)
# 4. Run transcription (make sure audio files are present in the specified paths):
audio_files = ["/content/segment_003.wav", "/content/segment_108.wav"]
lang = ["lus_Latn", "lus_Latn"]
transcriptions = pipeline.transcribe(audio_files, lang=lang, batch_size=2)
print("Transcription 0:", transcriptions[0])
print("Transcription 1:", transcriptions[1])
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/omni-asr-ctc-300m-v2-mizonal3-E1-lus-v2026.06 |
| E2 (Noise) | andrewbawitlung/omni-asr-ctc-300m-v2-mizonal3-E2-lus-v2026.06 |
| E3 (Speed) | andrewbawitlung/omni-asr-ctc-300m-v2-mizonal3-E3-lus-v2026.06 |
| E4 (SpecAug) | andrewbawitlung/omni-asr-ctc-300m-v2-mizonal3-E4-lus-v2026.06 |
| E5 (Combined) | andrewbawitlung/omni-asr-ctc-300m-v2-mizonal3-E5-lus-v2026.06 |
Training hyperparameters
Training hyperparameters not available in bin format.
Training results
| step | epoch | train_loss | eval_loss | eval_wer | eval_cer | learning_rate | grad_norm |
|---|---|---|---|---|---|---|---|
| 100 | 8.0000 | 86.7633 | 234.2761 | ||||
| 200 | 15.0000 | 59.5095 | 77.1118 | ||||
| 300 | 22.0000 | 50.3034 | 44.1160 | 42.1618 | 16.6563 | 85.3836 | |
| 400 | 29.0000 | 45.7736 | 91.6643 | ||||
| 500 | 36.0000 | 42.8647 | 78.6585 | ||||
| 600 | 43.0000 | 40.6568 | 38.6385 | 35.3988 | 14.1109 | 69.1629 | |
| 700 | 50.0000 | 38.3338 | 61.5000 | ||||
| 800 | 58.0000 | 36.6251 | 63.2135 | ||||
| 900 | 65.0000 | 36.2486 | 35.9211 | 32.5840 | 12.9047 | 71.4102 | |
| 1000 | 72.0000 | 34.8354 | 69.2921 | ||||
| 1100 | 79.0000 | 34.3219 | 70.9686 | ||||
| 1200 | 86.0000 | 33.5431 | 35.7565 | 30.2776 | 12.6705 | 79.8076 | |
| 1300 | 93.0000 | 32.5158 | 73.5882 | ||||
| 1400 | 100.0000 | 31.9301 | 89.3828 | ||||
| 1500 | 108.0000 | 31.4754 | 36.5236 | 31.3428 | 12.7627 | 76.0098 | |
| 1600 | 115.0000 | 31.0447 | 77.8923 | ||||
| 1700 | 122.0000 | 30.5563 | 56.5680 | ||||
| 1800 | 129.0000 | 30.0866 | 35.4210 | 30.0039 | 11.9858 | 61.9671 | |
| 1900 | 136.0000 | 29.8622 | 83.1070 | ||||
| 2000 | 143.0000 | 30.0428 | 63.8978 | ||||
| 2100 | 150.0000 | 29.7930 | 34.3638 | 28.5575 | 11.4111 | 61.9098 | |
| 2200 | 158.0000 | 29.4431 | 69.0937 | ||||
| 2300 | 165.0000 | 29.6798 | 64.0933 | ||||
| 2400 | 172.0000 | 29.2460 | 35.1561 | 28.9973 | 11.6044 | 65.5170 |
Sources
1 source
Sources
1 sourceVerified Oct 19
Model artifacts
1 artifact
Model artifacts
1 artifactSource excerpts
2 excerpts
Source excerpts
2 excerpts--- language: - lus license: apache-2.0 pipeline_tag: automatic-speech-recognition base_model: facebook/omniASR-CTC-300M tags: - generated_from_trainer datasets: - andrewbawitlung/MiZonal-v3.0 metrics: - wer - cer model-index: - name: omni-asr-ctc-300m-v2-mizonal3-E4-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: 18.0524 - name: Cer type: cer value: 3.2913 - name: Real Time Factor type: rtf value: 0.0065 ---  > **Disclaimer / Notice:** Details for these are in Peer Review and publications of the paper will be made available soon for more details. # omni-asr-ctc-300m-v2-mizonal3-E4-lus-v2026.06 This model is a fine-tuned version of [facebook/omniASR-CTC-300M](https://huggingface.co/facebook/omniASR-CTC-300M) on the **MiZonal v3.0** dataset. It achieves the following results on the evaluation set: - Wer: 18.0524 - Cer: 3.2913 - Real Time Factor: 0.0065 ## Quick Inference ```python # Google Colab Ready Quick Inference # 1. Install dependencies: !pip install huggingface_hub omnilingual-asr torchaudio # 2. Download the model checkpoint directly from Hugging Face: import torch from huggingface_hub import hf_hub_download model_path = hf_hub_download( repo_id="andrewbawitlung/omni-asr-ctc-300m-v2-mizonal3-E4-lus-v2026.06", filename="sdp_00.pt" ) # 3. Load the base model pipeline and inject fine-tuned weights: from omnilingual_asr.models.inference.pipeline import ASRInferencePipeline print("Loading base model pipeline...") pipeline = ASRInferencePipeline(model_card="omniASR_CTC_300M_v2") print("Loading fine-tuned checkpoint weights...") checkpoint = torch.load(model_path, map_location="cpu") state_dict = checkpoint['model'] if 'model' in checkpoint else checkpoint pipeline.model.load_state_dict(state_dict, strict=False) # 4. Run transcription (make sure audio files are present in the specified paths): audio_files = ["/content/segment_003.wav", "/content/segment_108.wav"] lang = ["lus_Latn", "lus_Latn"] transcriptions = pipeline.transcribe(audio_files, lang=lang, batch_size=2) print("Transcription 0:", transcriptions[0]) print("Transcription 1:", transcriptions[1]) ``` ## Model descripti...
Source context: 0 downloads · 0 likes · Pipeline automatic-speech-recognition · Repo andrewbawitlung/omni-asr-ctc-300m-v2-mizonal3-E4-lus-v2026.06
