nl8590687/ASRT_SpeechRecognition

Project overview

An open-source Chinese speech recognition system that transcribes audio into Chinese characters using a two-stage deep learning pipeline. It provides pre-trained models, HTTP and gRPC API servers, and client SDKs for self-hosted, local-only deployment.

Project type

Model Runtime · Audio & Speech

Use cases

Learning & Education

Deployment

Refer to project documentation

License

GPL-3.0

Best for

• Developers and AI engineers who need an open-source, self-hosted system for Chinese speech recognition.

Key capabilities

• A deep learning acoustic model using DCNN, LSTM, an attention mechanism, and CTC to transcribe Chinese audio into pinyin sequences.
• An N-gram based maximum entropy hidden Markov model that converts Chinese pinyin sequences into Chinese characters.
• Provides speech recognition inference capabilities via an HTTP protocol API server.
• Provides speech recognition inference capabilities via a gRPC protocol API server.
• Allows users to train the speech recognition models using configured datasets.
• Tests and evaluates the speech recognition models against test datasets.

Limitations and risks

• Input audio maximum length is restricted to 16 seconds.
• Training is only supported on Linux, while Windows supports only inference.
• The best test model achieves approximately 85% pinyin accuracy.

Getting started

• Setup is rated as medium difficulty due to specific software dependencies and manual steps. The initial path to success involves cloning the repository, installing Python dependencies, downloading pre-trained models, and running asrserver_http.py or asrserver_grpc.py.

Evidence and sources

• GitHub project description: A Deep-Learning-Based Chinese Speech Recognition System 基于深度学习的中文语音识别系统
• README: 启动ASRT HTTP协议的API服务器启动请执行： ```shell $ python3 asrserver_http.py ``` 本地测试调用HTTP协议API服务是否成功： ```shell $ python3 client_http.py ``` 启动ASRT GRPC协议的API服务器启动请执行： ```shell $ python3 asrs…
• README: 本项目使用tensorFlow.keras基于深度卷积神经网络和长短时记忆神经网络、注意力机制以及CTC实现。
• README: 基于概率图的最大熵隐马尔可夫模型 输入为汉语拼音序列，输出为对应的汉字文本
• README: DCNN + CTC 其中，输入的音频的最大时间长度为16秒，输出为对应的汉语拼音序列

Reviewed at: 2026-07-16T11:02:21Z