You can deploy this model and start earning money today!

James3214153/
test_zipformer

Name: test_zipformer
Availability: InStock

automatic-speech-recognition

apache-2.0

Views

No views yet

Zipformer ASR Model (Trained on 250k Hours)

This repository contains a high-performance Automatic Speech Recognition (ASR) model based on the Zipformer architecture. The model has been trained on an extensive dataset of 250,000 hours of speech, providing exceptional robustness and generalization across various acoustic environments.

📖 Model Overview

Architecture: Zipformer (k2-fsa / icefall)
Data Scale: 250k Hours
Framework: lhotse & icefall
Key Features: Utilizes a multi-scale encoder structure (U-Net style) to balance high accuracy with computational efficiency.

⚙️ Installation & Environment Setup

Please refer to document for installation.

🛠️ Evaluation & Decoding

To evaluate the performance of this model, follow the instructions below.

1. Dataset Configuration

Before running the evaluation, modify the test_cuts function in multi_dataset.py to point to your specific test sets and data paths.

2. Run Decoding Script

Execute the following command to start the decoding process. If OOM encountered, reduce the parameter "max-duration".

bash

python ./decode.py \
    --exp-dir ./exp \
    --max-duration 3000 \
    --decoding-method greedy_search \
    --bpe-model ./bpe.model \
    --num-workers 8 \
    --num-encoder-layers 2,2,4,5,4,2 \
    --feedforward-dim 512,768,1536,2048,1536,768 \
    --encoder-dim 192,256,512,768,512,256 \
    --encoder-unmasked-dim 192,192,256,320,256,192

James3214153/test_zipformer