support ngram and fst hotword for 2pass-offline (#1205)

Yabin Li

2023-12-26 b8825902d93d5017e44828316062dc8306b7ddcd

support ngram and fst hotword for 2pass-offline (#1205)

 runtime/onnxruntime/bin/funasr-onnx-2pass-rtf.cpp

@@ -44,12 +44,17 @@
}

void runReg(FUNASR_HANDLE tpass_handle, std::vector<int> chunk_size, vector<string> wav_list, vector<string> wav_ids, int audio_fs,
            float* total_length, long* total_time, int core_id, ASR_TYPE asr_mode_, string nn_hotwords_) {
            float* total_length, long* total_time, int core_id, ASR_TYPE asr_mode_, string nn_hotwords_,
            float glob_beam, float lat_beam, float am_scale, int inc_bias, unordered_map<string, int> hws_map) {
    
    struct timeval start, end;
    long seconds = 0;
    float n_total_length = 0.0f;
    long n_total_time = 0;

    FUNASR_DEC_HANDLE decoder_handle = FunASRWfstDecoderInit(tpass_handle, ASR_TWO_PASS, glob_beam, lat_beam, am_scale);
    // load hotwords list and build graph
    FunWfstDecoderLoadHwsRes(decoder_handle, inc_bias, hws_map);
       
    std::vector<std::vector<float>> hotwords_embedding = CompileHotwordEmbedding(tpass_handle, nn_hotwords_, ASR_TWO_PASS);
    
@@ -90,7 +95,8 @@
                } else {
                    is_final = false;
            }
            FUNASR_RESULT result = FunTpassInferBuffer(tpass_handle, tpass_online_handle, speech_buff+sample_offset, step, punc_cache, is_final, sampling_rate_, "pcm", (ASR_TYPE)asr_mode_, hotwords_embedding);
            FUNASR_RESULT result = FunTpassInferBuffer(tpass_handle, tpass_online_handle, speech_buff+sample_offset, step, punc_cache, is_final, 
                                                        sampling_rate_, "pcm", (ASR_TYPE)asr_mode_, hotwords_embedding, true, decoder_handle);
            if (result)
            {
                FunASRFreeResult(result);
@@ -139,7 +145,8 @@
                    is_final = false;
            }
            gettimeofday(&start, NULL);
            FUNASR_RESULT result = FunTpassInferBuffer(tpass_handle, tpass_online_handle, speech_buff+sample_offset, step, punc_cache, is_final, sampling_rate_, "pcm", (ASR_TYPE)asr_mode_, hotwords_embedding);
            FUNASR_RESULT result = FunTpassInferBuffer(tpass_handle, tpass_online_handle, speech_buff+sample_offset, step, punc_cache, is_final, 
                                                        sampling_rate_, "pcm", (ASR_TYPE)asr_mode_, hotwords_embedding, true, decoder_handle);
            gettimeofday(&end, NULL);
            seconds = (end.tv_sec - start.tv_sec);
            long taking_micros = ((seconds * 1000000) + end.tv_usec) - (start.tv_usec);
@@ -197,6 +204,8 @@
            *total_time = n_total_time;
        }
    }
    FunWfstDecoderUnloadHwsRes(decoder_handle);
    FunASRWfstDecoderUninit(decoder_handle);
    FunTpassOnlineUninit(tpass_online_handle);
}

@@ -215,6 +224,11 @@
    TCLAP::ValueArg<std::string>    punc_dir("", PUNC_DIR, "the punc online model path, which contains model.onnx, punc.yaml", false, "", "string");
    TCLAP::ValueArg<std::string>    punc_quant("", PUNC_QUANT, "true (Default), load the model of model.onnx in punc_dir. If set true, load the model of model_quant.onnx in punc_dir", false, "true", "string");
    TCLAP::ValueArg<std::string>    itn_dir("", ITN_DIR, "the itn model(fst) path, which contains zh_itn_tagger.fst and zh_itn_verbalizer.fst", false, "", "string");
    TCLAP::ValueArg<std::string>    lm_dir("", LM_DIR, "the lm model path, which contains compiled models: TLG.fst, config.yaml, lexicon.txt ", false, "", "string");
    TCLAP::ValueArg<float>    global_beam("", GLOB_BEAM, "the decoding beam for beam searching ", false, 3.0, "float");
    TCLAP::ValueArg<float>    lattice_beam("", LAT_BEAM, "the lattice generation beam for beam searching ", false, 3.0, "float");
    TCLAP::ValueArg<float>    am_scale("", AM_SCALE, "the acoustic scale for beam searching ", false, 10.0, "float");
    TCLAP::ValueArg<std::int32_t>   fst_inc_wts("", FST_INC_WTS, "the fst hotwords incremental bias", false, 20, "int32_t");

    TCLAP::ValueArg<std::string>    asr_mode("", ASR_MODE, "offline, online, 2pass", false, "2pass", "string");
    TCLAP::ValueArg<std::int32_t>   onnx_thread("", "model-thread-num", "onnxruntime SetIntraOpNumThreads", false, 1, "int32_t");
@@ -231,6 +245,11 @@
    cmd.add(punc_dir);
    cmd.add(punc_quant);
    cmd.add(itn_dir);
    cmd.add(lm_dir);
    cmd.add(global_beam);
    cmd.add(lattice_beam);
    cmd.add(am_scale);
    cmd.add(fst_inc_wts);
    cmd.add(wav_path);
    cmd.add(audio_fs);
    cmd.add(asr_mode);
@@ -248,6 +267,7 @@
    GetValue(punc_dir, PUNC_DIR, model_path);
    GetValue(punc_quant, PUNC_QUANT, model_path);
    GetValue(itn_dir, ITN_DIR, model_path);
    GetValue(lm_dir, LM_DIR, model_path);
    GetValue(wav_path, WAV_PATH, model_path);
    GetValue(asr_mode, ASR_MODE, model_path);

@@ -271,6 +291,14 @@
    {
        LOG(ERROR) << "FunTpassInit init failed";
        exit(-1);
    }
    float glob_beam = 3.0f;
    float lat_beam = 3.0f;
    float am_sc = 10.0f;
    if (lm_dir.isSet()) {
        glob_beam = global_beam.getValue();
        lat_beam = lattice_beam.getValue();
        am_sc = am_scale.getValue();
    }

    gettimeofday(&end, NULL);
@@ -321,7 +349,8 @@
    int rtf_threds = thread_num_.getValue();
    for (int i = 0; i < rtf_threds; i++)
    {
        threads.emplace_back(thread(runReg, tpass_hanlde, chunk_size, wav_list, wav_ids, audio_fs.getValue(), &total_length, &total_time, i, (ASR_TYPE)asr_mode_, nn_hotwords_));
        threads.emplace_back(thread(runReg, tpass_hanlde, chunk_size, wav_list, wav_ids, audio_fs.getValue(), &total_length, &total_time, i, (ASR_TYPE)asr_mode_, nn_hotwords_,
                                    glob_beam, lat_beam, am_sc, fst_inc_wts.getValue(), hws_map));
    }

    for (auto& thread : threads)

 runtime/onnxruntime/bin/funasr-onnx-2pass.cpp

@@ -51,6 +51,11 @@
    TCLAP::ValueArg<std::string>    punc_dir("", PUNC_DIR, "the punc online model path, which contains model.onnx, punc.yaml", false, "", "string");
    TCLAP::ValueArg<std::string>    punc_quant("", PUNC_QUANT, "true (Default), load the model of model.onnx in punc_dir. If set true, load the model of model_quant.onnx in punc_dir", false, "true", "string");
    TCLAP::ValueArg<std::string>    itn_dir("", ITN_DIR, "the itn model(fst) path, which contains zh_itn_tagger.fst and zh_itn_verbalizer.fst", false, "", "string");
    TCLAP::ValueArg<std::string>    lm_dir("", LM_DIR, "the lm model path, which contains compiled models: TLG.fst, config.yaml, lexicon.txt ", false, "", "string");
    TCLAP::ValueArg<float>    global_beam("", GLOB_BEAM, "the decoding beam for beam searching ", false, 3.0, "float");
    TCLAP::ValueArg<float>    lattice_beam("", LAT_BEAM, "the lattice generation beam for beam searching ", false, 3.0, "float");
    TCLAP::ValueArg<float>    am_scale("", AM_SCALE, "the acoustic scale for beam searching ", false, 10.0, "float");
    TCLAP::ValueArg<std::int32_t>   fst_inc_wts("", FST_INC_WTS, "the fst hotwords incremental bias", false, 20, "int32_t");
    TCLAP::ValueArg<std::string>    asr_mode("", ASR_MODE, "offline, online, 2pass", false, "2pass", "string");
    TCLAP::ValueArg<std::int32_t>   onnx_thread("", "model-thread-num", "onnxruntime SetIntraOpNumThreads", false, 1, "int32_t");

@@ -65,6 +70,11 @@
    cmd.add(vad_quant);
    cmd.add(punc_dir);
    cmd.add(punc_quant);
    cmd.add(lm_dir);
    cmd.add(global_beam);
    cmd.add(lattice_beam);
    cmd.add(am_scale);
    cmd.add(fst_inc_wts);
    cmd.add(itn_dir);
    cmd.add(wav_path);
    cmd.add(audio_fs);
@@ -81,6 +91,7 @@
    GetValue(vad_quant, VAD_QUANT, model_path);
    GetValue(punc_dir, PUNC_DIR, model_path);
    GetValue(punc_quant, PUNC_QUANT, model_path);
    GetValue(lm_dir, LM_DIR, model_path);
    GetValue(itn_dir, ITN_DIR, model_path);
    GetValue(wav_path, WAV_PATH, model_path);
    GetValue(asr_mode, ASR_MODE, model_path);
@@ -106,6 +117,16 @@
        LOG(ERROR) << "FunTpassInit init failed";
        exit(-1);
    }
    float glob_beam = 3.0f;
    float lat_beam = 3.0f;
    float am_sc = 10.0f;
    if (lm_dir.isSet()) {
        glob_beam = global_beam.getValue();
        lat_beam = lattice_beam.getValue();
        am_sc = am_scale.getValue();
    }
    // init wfst decoder
    FUNASR_DEC_HANDLE decoder_handle = FunASRWfstDecoderInit(tpass_handle, ASR_TWO_PASS, glob_beam, lat_beam, am_sc);

    gettimeofday(&end, NULL);
    long seconds = (end.tv_sec - start.tv_sec);
@@ -145,6 +166,9 @@
        wav_list.emplace_back(wav_path_);
        wav_ids.emplace_back(default_id);
    }

    // load hotwords list and build graph
    FunWfstDecoderLoadHwsRes(decoder_handle, fst_inc_wts.getValue(), hws_map);

    std::vector<std::vector<float>> hotwords_embedding = CompileHotwordEmbedding(tpass_handle, nn_hotwords_, ASR_TWO_PASS);
    // init online features
@@ -191,7 +215,9 @@
                    is_final = false;
            }
            gettimeofday(&start, NULL);
            FUNASR_RESULT result = FunTpassInferBuffer(tpass_handle, tpass_online_handle, speech_buff+sample_offset, step, punc_cache, is_final, sampling_rate_, "pcm", (ASR_TYPE)asr_mode_, hotwords_embedding);
            FUNASR_RESULT result = FunTpassInferBuffer(tpass_handle, tpass_online_handle, 
                speech_buff+sample_offset, step, punc_cache, is_final, sampling_rate_, "pcm", 
                (ASR_TYPE)asr_mode_, hotwords_embedding, true, decoder_handle);
            gettimeofday(&end, NULL);
            seconds = (end.tv_sec - start.tv_sec);
            taking_micros += ((seconds * 1000000) + end.tv_usec) - (start.tv_usec);
@@ -235,10 +261,12 @@
            }
        }
    }
 

    FunWfstDecoderUnloadHwsRes(decoder_handle);
    LOG(INFO) << "Audio length: " << (double)snippet_time << " s";
    LOG(INFO) << "Model inference takes: " << (double)taking_micros / 1000000 <<" s";
    LOG(INFO) << "Model inference RTF: " << (double)taking_micros/ (snippet_time*1000000);
    FunASRWfstDecoderUninit(decoder_handle);
    FunTpassOnlineUninit(tpass_online_handle);
    FunTpassUninit(tpass_handle);
    return 0;

 runtime/onnxruntime/bin/funasr-onnx-offline-rtf.cpp

@@ -54,7 +54,6 @@
    // warm up
    for (size_t i = 0; i < 1; i++)
    {
        FunOfflineReset(asr_handle, decoder_handle);
        FUNASR_RESULT result=FunOfflineInfer(asr_handle, wav_list[0].c_str(), RASR_NONE, NULL, hotwords_embedding, audio_fs, true, decoder_handle);
        if(result){
            FunASRFreeResult(result);

 runtime/onnxruntime/bin/funasr-onnx-offline.cpp

@@ -50,7 +50,7 @@
    TCLAP::ValueArg<std::string>    vad_quant("", VAD_QUANT, "true (Default), load the model of model.onnx in vad_dir. If set true, load the model of model_quant.onnx in vad_dir", false, "true", "string");
    TCLAP::ValueArg<std::string>    punc_dir("", PUNC_DIR, "the punc model path, which contains model.onnx, punc.yaml", false, "", "string");
    TCLAP::ValueArg<std::string>    punc_quant("", PUNC_QUANT, "true (Default), load the model of model.onnx in punc_dir. If set true, load the model of model_quant.onnx in punc_dir", false, "true", "string");
    TCLAP::ValueArg<std::string>    lm_dir("", LM_DIR, "the lm model path, which contains compiled models: TLG.fst, config.yaml ", false, "", "string");
    TCLAP::ValueArg<std::string>    lm_dir("", LM_DIR, "the lm model path, which contains compiled models: TLG.fst, config.yaml, lexicon.txt ", false, "", "string");
    TCLAP::ValueArg<float>    global_beam("", GLOB_BEAM, "the decoding beam for beam searching ", false, 3.0, "float");
    TCLAP::ValueArg<float>    lattice_beam("", LAT_BEAM, "the lattice generation beam for beam searching ", false, 3.0, "float");
    TCLAP::ValueArg<float>    am_scale("", AM_SCALE, "the acoustic scale for beam searching ", false, 10.0, "float");

 runtime/onnxruntime/include/funasrruntime.h

@@ -119,7 +119,7 @@
_FUNASRAPI FUNASR_RESULT    FunTpassInferBuffer(FUNASR_HANDLE handle, FUNASR_HANDLE online_handle, const char* sz_buf, 
                                                int n_len, std::vector<std::vector<std::string>> &punc_cache, bool input_finished=true, 
                                                int sampling_rate=16000, std::string wav_format="pcm", ASR_TYPE mode=ASR_TWO_PASS, 
                                                const std::vector<std::vector<float>> &hw_emb={{0.0}}, bool itn=true);
                                                const std::vector<std::vector<float>> &hw_emb={{0.0}}, bool itn=true, FUNASR_DEC_HANDLE dec_handle=nullptr);
_FUNASRAPI void                FunTpassUninit(FUNASR_HANDLE handle);
_FUNASRAPI void                FunTpassOnlineUninit(FUNASR_HANDLE handle);


 runtime/onnxruntime/src/audio.cpp

@@ -254,9 +254,9 @@
void Audio::WavResample(int32_t sampling_rate, const float *waveform,
                          int32_t n)
{
    LOG(INFO) << "Creating a resampler:\n"
              << "   in_sample_rate: "<< sampling_rate << "\n"
              << "   output_sample_rate: " << static_cast<int32_t>(dest_sample_rate);
    LOG(INFO) << "Creating a resampler: "
              << " in_sample_rate: "<< sampling_rate
              << " output_sample_rate: " << static_cast<int32_t>(dest_sample_rate);
    float min_freq =
        std::min<int32_t>(sampling_rate, dest_sample_rate);
    float lowpass_cutoff = 0.99 * 0.5 * min_freq;

 runtime/onnxruntime/src/funasrruntime.cpp

@@ -437,7 +437,7 @@
    _FUNASRAPI FUNASR_RESULT FunTpassInferBuffer(FUNASR_HANDLE handle, FUNASR_HANDLE online_handle, const char* sz_buf, 
                                                 int n_len, std::vector<std::vector<std::string>> &punc_cache, bool input_finished, 
                                                 int sampling_rate, std::string wav_format, ASR_TYPE mode, 
                                                 const std::vector<std::vector<float>> &hw_emb, bool itn)
                                                 const std::vector<std::vector<float>> &hw_emb, bool itn, FUNASR_DEC_HANDLE dec_handle)
    {
        funasr::TpassStream* tpass_stream = (funasr::TpassStream*)handle;
        funasr::TpassOnlineStream* tpass_online_stream = (funasr::TpassOnlineStream*)online_handle;
@@ -511,7 +511,12 @@
        // timestamp
        std::string cur_stamp = "[";        
        while(audio->FetchTpass(frame) > 0){
            string msg = ((funasr::Paraformer*)asr_handle)->Forward(frame->data, frame->len, frame->is_final, hw_emb);
            // dec reset
            funasr::WfstDecoder* wfst_decoder = (funasr::WfstDecoder*)dec_handle;
            if (wfst_decoder){
                wfst_decoder->StartUtterance();
            }
            string msg = ((funasr::Paraformer*)asr_handle)->Forward(frame->data, frame->len, frame->is_final, hw_emb, dec_handle);

            std::vector<std::string> msg_vec = funasr::split(msg, '|');  // split with timestamp
            if(msg_vec.size()==0){
@@ -762,8 +767,14 @@
            funasr::OfflineStream* offline_stream = (funasr::OfflineStream*)handle;
            funasr::Paraformer* paraformer = (funasr::Paraformer*)offline_stream->asr_handle.get();
            if (paraformer->lm_)
                mm = new funasr::WfstDecoder(paraformer->lm_.get(),
                    paraformer->GetPhoneSet(), paraformer->GetLmVocab(), glob_beam, lat_beam, am_scale);
        } else if (asr_type == ASR_TWO_PASS){
            funasr::TpassStream* tpass_stream = (funasr::TpassStream*)handle;
            funasr::Paraformer* paraformer = (funasr::Paraformer*)tpass_stream->asr_handle.get();
            if (paraformer->lm_)
                mm = new funasr::WfstDecoder(paraformer->lm_.get(), 
                    paraformer->GetPhoneSet(), paraformer->GetVocab(), glob_beam, lat_beam, am_scale);
                    paraformer->GetPhoneSet(), paraformer->GetLmVocab(), glob_beam, lat_beam, am_scale);
        }
        return mm;
    }

 runtime/onnxruntime/src/paraformer.cpp

@@ -193,8 +193,7 @@
        lm_ = std::shared_ptr<fst::Fst<fst::StdArc>>(
            fst::Fst<fst::StdArc>::Read(lm_file));
        if (lm_){
            if (vocab) { delete vocab; }
            vocab = new Vocab(lm_cfg_file.c_str(), lex_file.c_str());
            lm_vocab = new Vocab(lm_cfg_file.c_str(), lex_file.c_str());
            LOG(INFO) << "Successfully load lm file " << lm_file;
        }else{
            LOG(ERROR) << "Failed to load lm file " << lm_file;
@@ -309,6 +308,9 @@
{
    if(vocab){
        delete vocab;
    }
    if(lm_vocab){
        delete lm_vocab;
    }
    if(seg_dict){
        delete seg_dict;
@@ -687,6 +689,11 @@
    return vocab;
}

Vocab* Paraformer::GetLmVocab()
{
    return lm_vocab;
}

PhoneSet* Paraformer::GetPhoneSet()
{
    return phone_set_;

 runtime/onnxruntime/src/paraformer.h

@@ -20,6 +20,7 @@
    */
    private:
        Vocab* vocab = nullptr;
        Vocab* lm_vocab = nullptr;
        SegDict* seg_dict = nullptr;
        PhoneSet* phone_set_ = nullptr;
        //const float scale = 22.6274169979695;
@@ -65,6 +66,7 @@
        string FinalizeDecode(WfstDecoder* &wfst_decoder,
                          bool is_stamp=false, std::vector<float> us_alphas={0}, std::vector<float> us_cif_peak={0});
        Vocab* GetVocab();
        Vocab* GetLmVocab();
        PhoneSet* GetPhoneSet();
        
        knf::FbankOptions fbank_opts_;

 runtime/onnxruntime/src/tpass-stream.cpp

@@ -66,6 +66,20 @@
        LOG(ERROR) <<"Can not find offline-model-dir or online-model-dir";
        exit(-1);
    }
    
    // Lm resource
    if (model_path.find(LM_DIR) != model_path.end() && model_path.at(LM_DIR) != "") {
        string fst_path, lm_config_path, lex_path;
        fst_path = PathAppend(model_path.at(LM_DIR), LM_FST_RES);
        lm_config_path = PathAppend(model_path.at(LM_DIR), LM_CONFIG_NAME);
        lex_path = PathAppend(model_path.at(LM_DIR), LEX_PATH);
        if (access(lex_path.c_str(), F_OK) != 0 )
        {
            LOG(ERROR) << "Lexicon.txt file is not exist, please use the latest version. Skip load LM model.";
        }else{
            asr_handle->InitLm(fst_path, lm_config_path, lex_path);
        }
    }

    // PUNC model
    if(model_path.find(PUNC_DIR) != model_path.end()){

 runtime/websocket/bin/funasr-wss-client-2pass.cpp

@@ -192,7 +192,10 @@
    funasr::Audio audio(1);
    int32_t sampling_rate = audio_fs;
    std::string wav_format = "pcm";
    if (funasr::IsTargetFile(wav_path.c_str(), "pcm")) {
    if (funasr::IsTargetFile(wav_path.c_str(), "wav")) {
      if (!audio.LoadWav(wav_path.c_str(), &sampling_rate, false)) 
        return;
    } else if (funasr::IsTargetFile(wav_path.c_str(), "pcm")) {
      if (!audio.LoadPcmwav(wav_path.c_str(), &sampling_rate, false)) return;
    } else {
      wav_format = "others";

 runtime/websocket/bin/funasr-wss-server-2pass.cpp

@@ -16,6 +16,7 @@
// hotwords
std::unordered_map<std::string, int> hws_map_;
int fst_inc_wts_=20;
float global_beam_, lattice_beam_, am_scale_;

using namespace std;
void GetValue(TCLAP::ValueArg<std::string>& value_arg, string key,
@@ -120,6 +121,14 @@
        "connection",
        false, "../../../ssl_key/server.key", "string");

    TCLAP::ValueArg<float>    global_beam("", GLOB_BEAM, "the decoding beam for beam searching ", false, 3.0, "float");
    TCLAP::ValueArg<float>    lattice_beam("", LAT_BEAM, "the lattice generation beam for beam searching ", false, 3.0, "float");
    TCLAP::ValueArg<float>    am_scale("", AM_SCALE, "the acoustic scale for beam searching ", false, 10.0, "float");

    TCLAP::ValueArg<std::string> lm_dir("", LM_DIR,
        "the LM model path, which contains compiled models: TLG.fst, config.yaml ", false, "damo/speech_ngram_lm_zh-cn-ai-wesp-fst", "string");
    TCLAP::ValueArg<std::string> lm_revision(
        "", "lm-revision", "LM model revision", false, "v1.0.2", "string");
    TCLAP::ValueArg<std::string> hotword("", HOTWORD,
        "the hotword file, one hotword perline, Format: Hotword Weight (could be: 阿里巴巴 20)", 
        false, "/workspace/resources/hotwords.txt", "string");
@@ -128,6 +137,10 @@

    // add file
    cmd.add(hotword);
    cmd.add(fst_inc_wts);
    cmd.add(global_beam);
    cmd.add(lattice_beam);
    cmd.add(am_scale);

    cmd.add(certfile);
    cmd.add(keyfile);
@@ -146,6 +159,8 @@
    cmd.add(punc_quant);
    cmd.add(itn_dir);
    cmd.add(itn_revision);
    cmd.add(lm_dir);
    cmd.add(lm_revision);

    cmd.add(listen_ip);
    cmd.add(port);
@@ -163,6 +178,7 @@
    GetValue(punc_dir, PUNC_DIR, model_path);
    GetValue(punc_quant, PUNC_QUANT, model_path);
    GetValue(itn_dir, ITN_DIR, model_path);
    GetValue(lm_dir, LM_DIR, model_path);
    GetValue(hotword, HOTWORD, model_path);

    GetValue(offline_model_revision, "offline-model-revision", model_path);
@@ -170,6 +186,11 @@
    GetValue(vad_revision, "vad-revision", model_path);
    GetValue(punc_revision, "punc-revision", model_path);
    GetValue(itn_revision, "itn-revision", model_path);
    GetValue(lm_revision, "lm-revision", model_path);

    global_beam_ = global_beam.getValue();
    lattice_beam_ = lattice_beam.getValue();
    am_scale_ = am_scale.getValue();

    // Download model form Modelscope
    try {
@@ -183,6 +204,7 @@
      std::string s_punc_path = model_path[PUNC_DIR];
      std::string s_punc_quant = model_path[PUNC_QUANT];
      std::string s_itn_path = model_path[ITN_DIR];
      std::string s_lm_path = model_path[LM_DIR];

      std::string python_cmd =
          "python -m funasr.utils.runtime_sdk_download_tool --type onnx --quantize True ";
@@ -241,11 +263,18 @@
        size_t found = s_offline_asr_path.find("speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404");
        if (found != std::string::npos) {
            model_path["offline-model-revision"]="v1.2.4";
        } else{
            found = s_offline_asr_path.find("speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404");
            if (found != std::string::npos) {
                model_path["offline-model-revision"]="v1.0.5";
            }
        }

        found = s_offline_asr_path.find("speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404");
        if (found != std::string::npos) {
            model_path["offline-model-revision"]="v1.0.5";
        }

        found = s_offline_asr_path.find("speech_paraformer-large_asr_nat-en-16k-common-vocab10020");
        if (found != std::string::npos) {
            model_path["model-revision"]="v1.0.0";
            s_itn_path="";
            s_lm_path="";
        }

        if (access(s_offline_asr_path.c_str(), F_OK) == 0) {
@@ -332,6 +361,49 @@
        LOG(INFO) << "ASR online model is not set, use default.";
      }

      if (!s_lm_path.empty() && s_lm_path != "NONE" && s_lm_path != "none") {
          std::string python_cmd_lm;
          std::string down_lm_path;
          std::string down_lm_model;

          if (access(s_lm_path.c_str(), F_OK) == 0) {
              // local
              python_cmd_lm = python_cmd + " --model-name " + s_lm_path +
                                  " --export-dir ./ " + " --model_revision " +
                                  model_path["lm-revision"] + " --export False ";
              down_lm_path = s_lm_path;
          } else {
              // modelscope
              LOG(INFO) << "Download model: " << s_lm_path
                          << " from modelscope : "; 
              python_cmd_lm = python_cmd + " --model-name " +
                      s_lm_path +
                      " --export-dir " + s_download_model_dir +
                      " --model_revision " + model_path["lm-revision"]
                      + " --export False "; 
              down_lm_path  =
                      s_download_model_dir +
                      "/" + s_lm_path;
          }

          int ret = system(python_cmd_lm.c_str());
          if (ret != 0) {
              LOG(INFO) << "Failed to download model from modelscope. If you set local lm model path, you can ignore the errors.";
          }
          down_lm_model = down_lm_path + "/TLG.fst";

          if (access(down_lm_model.c_str(), F_OK) != 0) {
              LOG(ERROR) << down_lm_model << " do not exists.";
              exit(-1);
          } else {
              model_path[LM_DIR] = down_lm_path;
              LOG(INFO) << "Set " << LM_DIR << " : " << model_path[LM_DIR];
          }
      } else {
          LOG(INFO) << "LM model is not set, not executed.";
          model_path[LM_DIR] = "";
      }

      if (!s_punc_path.empty()) {
        std::string python_cmd_punc;
        std::string down_punc_path;

 runtime/websocket/bin/websocket-server-2pass.cpp

@@ -18,6 +18,7 @@

extern std::unordered_map<std::string, int> hws_map_;
extern int fst_inc_wts_;
extern float global_beam_, lattice_beam_, am_scale_;

context_ptr WebSocketServer::on_tls_init(tls_mode mode,
                                         websocketpp::connection_hdl hdl,
@@ -102,7 +103,8 @@
    bool itn,
    int audio_fs,
    std::string wav_format,
    FUNASR_HANDLE& tpass_online_handle) {
    FUNASR_HANDLE& tpass_online_handle,
    FUNASR_DEC_HANDLE& decoder_handle) {
  // lock for each connection
  if(!tpass_online_handle){
    scoped_lock guard(thread_lock);
@@ -131,7 +133,7 @@
                                       subvector.data(), subvector.size(),
                                       punc_cache, false, audio_fs,
                                       wav_format, (ASR_TYPE)asr_mode_,
                                       hotwords_embedding, itn);
                                       hotwords_embedding, itn, decoder_handle);

        } else {
          scoped_lock guard(thread_lock);
@@ -168,7 +170,7 @@
                                       buffer.data(), buffer.size(), punc_cache,
                                       is_final, audio_fs,
                                       wav_format, (ASR_TYPE)asr_mode_,
                                       hotwords_embedding, itn);
                                       hotwords_embedding, itn, decoder_handle);
        } else {
          scoped_lock guard(thread_lock);
          msg["access_num"]=(int)msg["access_num"]-1;     
@@ -241,6 +243,9 @@
    data_msg->msg["audio_fs"] = 16000; // default is 16k
    data_msg->msg["access_num"] = 0; // the number of access for this object, when it is 0, we can free it saftly
    data_msg->msg["is_eof"]=false; // if this connection is closed
    FUNASR_DEC_HANDLE decoder_handle =
      FunASRWfstDecoderInit(tpass_handle, ASR_TWO_PASS, global_beam_, lattice_beam_, am_scale_);
    data_msg->decoder_handle = decoder_handle;
    data_msg->punc_cache =
        std::make_shared<std::vector<std::vector<std::string>>>(2);
      data_msg->strand_ =    std::make_shared<asio::io_context::strand>(io_decoder_);
@@ -267,6 +272,9 @@
  // finished and avoid access freed tpass_online_handle
  unique_lock guard_decoder(*(data_msg->thread_lock));
  if (data_msg->msg["access_num"]==0 && data_msg->msg["is_eof"]==true) {
    FunWfstDecoderUnloadHwsRes(data_msg->decoder_handle);
    FunASRWfstDecoderUninit(data_msg->decoder_handle);
    data_msg->decoder_handle = nullptr;
    FunTpassOnlineUninit(data_msg->tpass_online_handle);
    data_msg->tpass_online_handle = nullptr;
      data_map.erase(hdl);
@@ -431,7 +439,7 @@
            nn_hotwords += " " + pair.first;
            LOG(INFO) << pair.first << " : " << pair.second;
        }
        // FunWfstDecoderLoadHwsRes(msg_data->decoder_handle, fst_inc_wts_, merged_hws_map);
        FunWfstDecoderLoadHwsRes(msg_data->decoder_handle, fst_inc_wts_, merged_hws_map);

        // nn
        std::vector<std::vector<float>> new_hotwords_embedding = CompileHotwordEmbedding(tpass_handle, nn_hotwords, ASR_TWO_PASS);
@@ -483,7 +491,8 @@
                        msg_data->msg["itn"],
                        msg_data->msg["audio_fs"],
                        msg_data->msg["wav_format"],
                        std::ref(msg_data->tpass_online_handle)));
                        std::ref(msg_data->tpass_online_handle),
                        std::ref(msg_data->decoder_handle)));
              msg_data->msg["access_num"]=(int)(msg_data->msg["access_num"])+1;
        }
        catch (std::exception const &e)
@@ -530,7 +539,8 @@
                                  msg_data->msg["itn"],
                                  msg_data->msg["audio_fs"],
                                  msg_data->msg["wav_format"],
                                  std::ref(msg_data->tpass_online_handle)));
                                  std::ref(msg_data->tpass_online_handle),
                                  std::ref(msg_data->decoder_handle)));
              msg_data->msg["access_num"]=(int)(msg_data->msg["access_num"])+1;
            }
          }

 runtime/websocket/bin/websocket-server-2pass.h

@@ -60,7 +60,8 @@
  FUNASR_HANDLE tpass_online_handle=NULL;
  std::string online_res = "";
  std::string tpass_res = "";
  std::shared_ptr<asio::io_context::strand>  strand_; // for data execute in order 
  std::shared_ptr<asio::io_context::strand>  strand_; // for data execute in order
  FUNASR_DEC_HANDLE decoder_handle=NULL; 
} FUNASR_MESSAGE;

// See https://wiki.mozilla.org/Security/Server_Side_TLS for more details about
@@ -123,7 +124,8 @@
                  bool itn,
                  int audio_fs,
                  std::string wav_format,
                  FUNASR_HANDLE& tpass_online_handle);
                  FUNASR_HANDLE& tpass_online_handle,
                  FUNASR_DEC_HANDLE& decoder_handle);

  void initAsr(std::map<std::string, std::string>& model_path, int thread_num);
  void on_message(websocketpp::connection_hdl hdl, message_ptr msg);