PaddleSpeech/docs/source/asr/PPASR.md

(简体中文|English)

PP-ASR

Catalogue

• 1. Introduction
• 2. Characteristic
• 3. Tutorials
  • 3.1 Pre-trained Models
  • 3.2 Training
  • 3.3 Inference
  • 3.4 Service Deployment
  • 3.5 Customized Auto Speech Recognition and Deployment
• 4. Quick Start

1. Introduction

PP-ASR is a tool to provide ASR(Automatic speech recognition) function. It provides a variety of Chinese and English models and supports model training. It also supports model inference using the command line. In addition, PP-ASR supports the deployment of streaming models and customized ASR.

2. Characteristic

The basic process of ASR is shown in the figure below:

The main characteristics of PP-ASR are shown below:

• Provides pre-trained models on Chinese/English open source datasets: aishell(Chinese), wenetspeech(Chinese) and librispeech(English). The models include deepspeech2 and conformer/transformer.
• Support model training on Chinese/English datasets.
• Support model inference using the command line. You can use to use paddlespeech asr --model xxx --input xxx.wav to use the pre-trained model to do model inference.
• Support deployment of streaming ASR server. Besides ASR function, the server supports timestamp function.
• Support customized auto speech recognition and deployment.

3. Tutorials

3.1 Pre-trained Models

The support pre-trained model list: released_model.
The model with good effect are Ds2 Online Wenetspeech ASR0 Model and Conformer Online Wenetspeech ASR1 Model. Both two models support streaming ASR.
For more information about model design, you can refer to the aistudio tutorial:

• Deepspeech2
• Transformer

3.2 Training

The referenced script for model training is stored in examples and stored according to "examples/dataset/model". The dataset mainly supports aishell and librispeech. The model supports deepspeech2 and u2(conformer/transformer). The specific steps of executing the script are recorded in run.sh.

For more information, you can refer to asr1

3.3 Inference

PP-ASR supports use paddlespeech asr --model xxx --input xxx.wav to use the pre-trained model to do model inference after install paddlespeech by pip install paddlespeech.

Specific supported functions include:

• Prediction of single audio
• Use the pipe to predict multiple audio
• Support RTF calculation

For specific usage, please refer to: speech_recognition

3.4 Service Deployment

PP-ASR supports the service deployment of streaming ASR. Support the simultaneous use of speech recognition and punctuation processing.

Demo of ASR Server: streaming_asr_server

Display of using ASR server on Web page: streaming_asr_demo_video

For more information about service deployment, you can refer to the aistudio tutorial:

• Streaming service - model part
• Streaming service

3.5 Customized Auto Speech Recognition and Deployment

For customized auto speech recognition and deployment, PP-ASR provides feature extraction(fbank) => Inference model（Scoring Library）=> C++ program of TLG（WFST, token, lexion, grammer). For specific usage, please refer to: speechx
If you want to quickly use it, you can refer to custom_streaming_asr

For more information about customized auto speech recognition and deployment, you can refer to the aistudio tutorial:

• Customized Auto Speech Recognition

4. Quick Start

To use PP-ASR, you can see here install, It supplies three methods to install paddlespeech, which are Easy, Medium and Hard. If you want to experience the inference function of paddlespeech, you can use Easy installation method.