🔥 FireRedTrainAlign: Chinese Speech Recognition Training and Inference Toolkit

This project provides training and inference code for the FireRedTeam/FireRedASR-AED-L model, supporting high-quality Chinese speech transcription tasks. The model adopts an Attention-based Encoder-Decoder (AED) architecture, suitable for Mandarin and certain dialects.

---

🚀 Project Features

• ✅ Supports the FireRedASR-AED-L model on Hugging Face
• ✅ Provides a complete training pipeline (data preprocessing, model training, and application)
• ✅ Supports beam search decoding and batch inference for multiple audio files
• ✅ Optimized for medium to long speech segments, applicable in various Chinese speech scenarios

---

⚙️ Installation & Environment Requirements

1️⃣ Create a Virtual Environment (Conda recommended)

git clone https://github.com/AUSTIN2526/FireRedTrainAlign.git
cd FireRedTrainAlign

conda create -n firered_asr python=3.10
conda activate firered_asr

pip install -r requirements.txt

2️⃣ Additional Dependencies

• ffmpeg: Required for audio format conversion and resampling
• CUDA (optional): For GPU-accelerated training or inference

---

🎯 Notes & Limitations

• It is recommended that audio length does not exceed 60 seconds to avoid decoding errors or repeated phrases
• Audio longer than 200 seconds may cause errors in positional encoding
• Batch inference and training are not fully tested yet
• Please place the FireRedASR-AED-L model in the project directory for use

---

📁 Preparing Training Data

Please prepare a .tsv or .txt text file in the following format, where each line represents one audio sample and its corresponding transcription:

ID\tTranscription text
ID\tTranscription text
ID\tTranscription text
...

---

📊 Model Performance (from the official paper)

Model Name	Parameters	AISHELL1	AISHELL2	WS Net	WS Meeting	Avg. CER
FireRedASR-AED-L	1.1B	0.55%	2.52%	4.88%	4.76%	3.18%

🔗 Speech-to-Text Integration: Supports med-voice-SHI-detector

This project can be integrated with med-voice-SHI-detector to automatically transcribe medical voice data into text and perform SHI (Speech-based Health Information) de-identification. The overall processing workflow is as follows:

1. Speech-to-Text (ASR): Use med-voice-SHI-detector to convert Chinese medical audio files into text.

2. SHI De-identification: Input the transcribed text into this project to identify and mask sensitive information such as names, healthcare institutions, and ID numbers.

3. Output Format: Generates de-identified text records suitable for downstream NLP tasks or data analysis.

👉 Installation & Usage Please refer to the official documentation of med-voice-SHI-detector for detailed instructions.