HotBias

Model weights for HotBias: Hotword Biasing for ASR with Cross-Domain Generalization and Cross-Model Transfer.

Official code and usage: QwenAudio/FunResearch/HotBias.

HotBias adds gated, adaptive hotword biases to frozen ASR models. A pretrained biasing core can be transferred to another recognizer by aligning target-specific input projections.

Checkpoints

File	Purpose
source_transfer.pt	Pretrained HotBias core and source input projections, exported from the Qwen3-based source system. Use it to initialize target alignment.
nano_aligned.pt	Aligned HotBias weights for Fun-ASR-Nano-2512.
firered_aligned.pt	Aligned HotBias weights for FireRedASR2-AED.

These files contain HotBias parameters. Obtain the corresponding ASR backbone and tokenizer separately. All three checkpoints share the same frozen biasing core.

Training data

The HotBias module was trained on speech data from AISHELL-2, and generated hotword candidate lists.

Download and use

Download from Hugging Face:

from huggingface_hub import snapshot_download

model_dir = snapshot_download('coming245/HotBias')
print(model_dir)

The same files are also available on ModelScope:

from modelscope import snapshot_download

model_dir = snapshot_download('coming245/HotBias')
print(model_dir)

Pass nano_aligned.pt or firered_aligned.pt from the downloaded directory to the corresponding inference script. For example, from the HotBias code directory:

python fun_asr_nano/infer.py \
  --model-dir /path/to/Fun-ASR-Nano-2512 \
  --nano-model-code /path/to/Fun-ASR/model.py \
  --checkpoint /path/to/hotbias-checkpoints/nano_aligned.pt \
  --audio /path/to/audio.wav \
  --hotword '池州' --hotword '赤州'

See the GitHub README for installation, FireRed inference, and target alignment. See checkpoint details for compatibility and reproduction notes for evaluation scope.

File integrity

From the downloaded directory, run:

sha256sum -c SHA256SUMS