xuning / sherpaonnx

0.Introduction

Java wrapper com.k2fsa.sherpa.onnx.OnlineRecognizer for sherpa-onnx. Java is a cross-platform language; you can build jni .so lib according to your system, and then use the same java api for all your platform. now support multiple threads for websocket server

Depend on:
  Openjdk 1.8

---

1.Compile libsherpa-onnx-jni.so

Compile sherpa-onnx/jni/jni.cc according to your system. Example for Ubuntu 18.04 LTS, Openjdk 1.8.0_362:

  git clone https://github.com/k2-fsa/sherpa-onnx
  cd sherpa-onnx
  mkdir build
  cd build
  cmake -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -DSHERPA_ONNX_ENABLE_JNI=ON ..
  make -j6

---

2.Download asr model files

click here for more detail

3.Config model config.cfg

/change model path in config.cfg according to your env/

  #model config 
  sample_rate=16000 
  feature_dim=80
  rule1_min_trailing_silence=2.4
  rule2_min_trailing_silence=1.2
  rule3_min_utterance_length=20
  encoder=/sherpa-onnx/build/bin/sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20/encoder-epoch-99-avg-1.onnx
  decoder=/sherpa-onnx/build/bin/sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20/decoder-epoch-99-avg-1.onnx
  joiner=/sherpa-onnx/build/bin/sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20/joiner-epoch-99-avg-1.onnx
  tokens=/sherpa-onnx/build/bin/sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20/tokens.txt
  num_threads=4
  enable_endpoint_detection=false
  decoding_method=greedy_search
  max_active_paths=4

  #websocket server config
  port=8890
  #number of threads pool for network io
  connection_thread_num=16 
  #number of threads pool for stream
  stream_thread_num=16 
  #number of threads pool for decoder 
  decoder_thread_num=16 
  #size of streams for parallel decoding
  parallel_decoder_num=16
  #time(ms) idle for decoder thread when no job
  decoder_time_idle=10
  #time(ms) out for connection data
  deocder_time_out=3000

---

4.A simple java example

refer to java_api_example for more detail.

    import com.k2fsa.sherpa.onnx.OnlineRecognizer;
    import com.k2fsa.sherpa.onnx.OnlineStream;
    String cfgpath=appdir+"/modelconfig.cfg";
    OnlineRecognizer.setSoPath(soPath);   //set so lib path

    OnlineRecognizer rcgOjb = new OnlineRecognizer();   //create a recognizer
    rcgOjb = new OnlineRecognizer(cfgFile);    //set model config file 
    CreateStream streamObj=rcgOjb.CreateStream();       //create a stream for read wav data
    float[] buffer = rcgOjb.readWavFile(wavfilename); // read data from file
    streamObj.acceptWaveform(buffer); // feed stream with data
    streamObj.inputFinished(); // tell engine you done with all data
    OnlineStream ssObj[] = new OnlineStream[1];
    while (rcgOjb.isReady(streamObj)) { // engine is ready for unprocessed data
                ssObj[0] = streamObj;
                rcgOjb.decodeStreams(ssObj); // decode for multiple stream
                // rcgOjb.DecodeStream(streamObj);   // decode for single stream
            }

    String recText = "simple:" + rcgOjb.getResult(streamObj) + "\n";
    byte[] utf8Data = recText.getBytes(StandardCharsets.UTF_8);
    System.out.println(new String(utf8Data));
    rcgOjb.reSet(streamObj);
    rcgOjb.releaseStream(streamObj); // release stream
    rcgOjb.release(); // release recognizer

---

5.Makefile

OS Ubuntu 18.04 LTS Build package path: /sherpa-onnx/java-api-examples/lib/sherpaonnx.jar

5.1 Build

    cd sherpa-onnx/java-api-examples
    make all

5.2 Run DecodeFile example

    make runfile

5.3 Run DecodeMic example

    make runmic

---

6.WebSocket Server

support multiple threads for websocket server 6.0 Protocol for communication 1) client connect to server

   ws client -> srv ws address
   ws address example: ws://127.0.0.1:8889/

2) client send 16k pcm_s16le binary stream data to server

   PCM   sampleRate 16000
         single channel
         sampleSize 16bit
         little endian
         type short

3) client send "Done" text to server when all data is sent

    ws_socket.send("Done")

4) client will receive json message from server whenever asr engine decoded new text

   json example: {"text":"甚至出现交易几乎停滞的情况","eof":false"}

6.1 Build

    cd sherpa-onnx/java-api-examples
    make all

6.2 Run srv example

usage: AsrWebsocketServer soPath modelCfgPath

    make runsrv  /**change path in Makefile according to your env**/

6.3 Run multiple threads client example

usage: AsrWebsocketClient soPath srvIp srvPort wavPath numThreads

json result example: {"text":"甚至出现交易几乎停滞的情况","eof":"true"}

    make runclient  /**change path in Makefile according to your env**/