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
File size: 1,193 Bytes
262fa3f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 | # Experiment Runners
This directory contains the maintained entrypoints for LatentASR experiments.
All runners default to Qwen/Qwen3-ASR-0.6B and `./latent_qwen_asr_best.pth`.
The checkpoint itself is intentionally ignored by Git.
## Runners
- `run_english_clean_and_noise.sh`: English clean suite plus optional SNR noise sweeps.
- `run_multilingual_streaming.sh`: FLEURS 30 and MLS public7 streaming evaluation.
- `run_threshold_sweep.sh`: Value-head halting threshold sweep.
- `run_all_training_modes.sh`: Sequential baseline, prompt tuning, LoRA, and LatentASR training.
- `run_legacy_generalization.sh`: older SpeechTest generalization/noise runner kept for reproducibility.
Root-level scripts with the old names are thin compatibility wrappers around
these maintained files.
## Common Environment Variables
- `LATENT_CKPT`: latent adapter checkpoint path.
- `MODEL_ID`: HuggingFace model ID.
- `PYTHON_BIN`: Python executable override.
- `MAX_SAMPLES_PER_CONFIG`: per-config cap; `0` means full split.
- `OUT_DIR`: output directory.
- `RESUME=1`: skip JSON outputs that already exist.
Full experimental JSON/log outputs are written under `eval_runs/` and are not
committed to Git.
|