molind/conformer-transducer-be-mlx

Conformer-Transducer (RNN-T) Belarusian (MLX)

Code: molind/mlx-conformer

NVIDIA's Conformer-Transducer Large model for Belarusian speech recognition, packaged for MLX inference on Apple Silicon.

Original model: nvidia/stt_be_conformer_transducer_large

Results

Dataset	WER	Speed
CommonVoice 24.0 test (500 samples)	6.29%	7.6 samples/s

Best open-source Belarusian ASR result we are aware of on CommonVoice 24.0.

Usage

pip install mlx numpy pyyaml sentencepiece torch

git clone https://github.com/molind/mlx-conformer
cd mlx-conformer

python mlx_conformer.py \
    --download nvidia/stt_be_conformer_transducer_large \
    --output models

python mlx_conformer.py \
    --model models/stt_be_conformer_transducer_large \
    --type transducer \
    --audio test.mp3

Architecture

• 17 Conformer encoder layers, d_model=512, 8 heads
• LSTM prediction network (1 layer, 640 hidden)
• Joint network (640 hidden, ReLU)
• 1024 BPE vocabulary + blank
• ~120M parameters

License

Original model by NVIDIA, licensed under CC-BY-4.0.