Automatic Speech Recognition
Transformers
qwen3-asr
latent-reasoning
test-time-compute
parameter-efficient
Instructions to use voidful/latentASR with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use voidful/latentASR with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("automatic-speech-recognition", model="voidful/latentASR")# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("voidful/latentASR", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Checkpoints
This directory contains the released LatentASR adapter checkpoint:
latentASR_adapter.pth
The checkpoint stores only the lightweight LatentASR modules:
init_projdelta_projstep_projstep_embedlog_scalevalue_headinjection_gate
It does not include the frozen Qwen/Qwen3-ASR-0.6B backbone. Evaluation and
inference load the base model from Hugging Face and then attach this adapter.
Checkpoint metadata:
- Source run:
activation_500_epoch10.pth - Base model:
Qwen/Qwen3-ASR-0.6B - Latent budget:
N=4 - Training set size: 500 utterances
- Adapter parameters: 5,251,077
- SHA256:
f0ce39fa5e6952fced6992508f3e2b32ea8467442b545678781a0f04e64f2430