Merge pull request #524 from alibaba-damo-academy/main
update dev_lyh
| | |
| | | }
|
| | | }
|
| | | ```
|
| | |
|
| | | ### 修改wsconnecter.js里asr接口地址
|
| | | wsconnecter.js里配置online asr服务地址路径,这里配置的是wss端口
|
| | | var Uri = "wss://xxx:xxx/wss/" |
| | | ## Acknowledge
|
| | | 1. This project is maintained by [FunASR community](https://github.com/alibaba-damo-academy/FunASR).
|
| | | 2. We acknowledge [AiHealthx](http://www.aihealthx.com/) for contributing the html5 demo. |
| | |
| | | /* 2021-2023 by zhaoming,mali aihealthx.com */
|
| | |
|
| | | function WebSocketConnectMethod( config ) { //定义socket连接方法类
|
| | | var Uri = "wss://111.205.137.58:5821/wss/" //设置wss asr online接口地址 如 wss://X.X.X.X:port/wss/
|
| | | var Uri = "wss://30.220.136.139:5921/" // var Uri = "wss://30.221.177.46:5921/" //设置wss asr online接口地址 如 wss://X.X.X.X:port/wss/
|
| | | var speechSokt;
|
| | | var connKeeperID;
|
| | |
|
| | |
| | | public: |
| | | virtual ~VadModel(){}; |
| | | virtual void InitVad(const std::string &vad_model, const std::string &vad_cmvn, const std::string &vad_config, int thread_num)=0; |
| | | virtual std::vector<std::vector<int>> Infer(const std::vector<float> &waves)=0; |
| | | virtual std::vector<std::vector<int>> Infer(std::vector<float> &waves, bool input_finished=true)=0; |
| | | virtual void ReadModel(const char* vad_model)=0; |
| | | virtual void LoadConfigFromYaml(const char* filename)=0; |
| | | virtual void FbankKaldi(float sample_rate, std::vector<std::vector<float>> &vad_feats, |
| | | const std::vector<float> &waves)=0; |
| | | virtual void LfrCmvn(std::vector<std::vector<float>> &vad_feats)=0; |
| | | virtual void Forward( |
| | | const std::vector<std::vector<float>> &chunk_feats, |
| | | std::vector<std::vector<float>> *out_prob)=0; |
| | | std::vector<float> &waves)=0; |
| | | virtual void LoadCmvn(const char *filename)=0; |
| | | virtual void InitCache()=0; |
| | | }; |
| | |
| | | ### funasr-onnx-offline-rtf |
| | | ```shell |
| | | ./funasr-onnx-offline-rtf --model-dir <string> [--quantize <string>] |
| | | [--vad-dir <string>] [--vad-quant <string>] |
| | | [--punc-dir <string>] [--punc-quant <string>] |
| | | --wav-path <string> --thread-num <int32_t> |
| | | [--] [--version] [-h] |
| | | Where: |
| | |
| | | (required) the model path, which contains model.onnx, config.yaml, am.mvn |
| | | --quantize <string> |
| | | false (Default), load the model of model.onnx in model_dir. If set true, load the model of model_quant.onnx in model_dir |
| | | |
| | | --vad-dir <string> |
| | | the vad model path, which contains model.onnx, vad.yaml, vad.mvn |
| | | --vad-quant <string> |
| | | false (Default), load the model of model.onnx in vad_dir. If set true, load the model of model_quant.onnx in vad_dir |
| | | |
| | | --punc-dir <string> |
| | | the punc model path, which contains model.onnx, punc.yaml |
| | | --punc-quant <string> |
| | | false (Default), load the model of model.onnx in punc_dir. If set true, load the model of model_quant.onnx in punc_dir |
| | | |
| | | --wav-path <string> |
| | | (required) the input could be: |
| | | wav_path, e.g.: asr_example.wav; |
| | |
| | | } |
| | | |
| | | // get 4 caches outputs,each size is 128*19 |
| | | for (int i = 1; i < 5; i++) { |
| | | float* data = vad_ort_outputs[i].GetTensorMutableData<float>(); |
| | | memcpy(in_cache_[i-1].data(), data, sizeof(float) * 128*19); |
| | | } |
| | | // for (int i = 1; i < 5; i++) { |
| | | // float* data = vad_ort_outputs[i].GetTensorMutableData<float>(); |
| | | // memcpy(in_cache_[i-1].data(), data, sizeof(float) * 128*19); |
| | | // } |
| | | } |
| | | |
| | | void FsmnVad::FbankKaldi(float sample_rate, std::vector<std::vector<float>> &vad_feats, |
| | | const std::vector<float> &waves) { |
| | | std::vector<float> &waves) { |
| | | knf::OnlineFbank fbank(fbank_opts); |
| | | |
| | | fbank.AcceptWaveform(sample_rate, &waves[0], waves.size()); |
| | | std::vector<float> buf(waves.size()); |
| | | for (int32_t i = 0; i != waves.size(); ++i) { |
| | | buf[i] = waves[i] * 32768; |
| | | } |
| | | fbank.AcceptWaveform(sample_rate, buf.data(), buf.size()); |
| | | int32_t frames = fbank.NumFramesReady(); |
| | | for (int32_t i = 0; i != frames; ++i) { |
| | | const float *frame = fbank.GetFrame(i); |
| | |
| | | } |
| | | |
| | | std::vector<std::vector<int>> |
| | | FsmnVad::Infer(const std::vector<float> &waves) { |
| | | FsmnVad::Infer(std::vector<float> &waves, bool input_finished) { |
| | | std::vector<std::vector<float>> vad_feats; |
| | | std::vector<std::vector<float>> vad_probs; |
| | | FbankKaldi(vad_sample_rate_, vad_feats, waves); |
| | |
| | | ~FsmnVad(); |
| | | void Test(); |
| | | void InitVad(const std::string &vad_model, const std::string &vad_cmvn, const std::string &vad_config, int thread_num); |
| | | std::vector<std::vector<int>> Infer(const std::vector<float> &waves); |
| | | std::vector<std::vector<int>> Infer(std::vector<float> &waves, bool input_finished=true); |
| | | void Reset(); |
| | | |
| | | private: |
| | |
| | | std::vector<const char *> *in_names, std::vector<const char *> *out_names); |
| | | |
| | | void FbankKaldi(float sample_rate, std::vector<std::vector<float>> &vad_feats, |
| | | const std::vector<float> &waves); |
| | | std::vector<float> &waves); |
| | | |
| | | void LfrCmvn(std::vector<std::vector<float>> &vad_feats); |
| | | |
| | |
| | | // warm up |
| | | for (size_t i = 0; i < 1; i++) |
| | | { |
| | | FUNASR_RESULT result=FunASRInfer(asr_handle, wav_list[0].c_str(), RASR_NONE, NULL, 16000); |
| | | FUNASR_RESULT result=FunOfflineInfer(asr_handle, wav_list[0].c_str(), RASR_NONE, NULL, 16000); |
| | | } |
| | | |
| | | while (true) { |
| | |
| | | } |
| | | |
| | | gettimeofday(&start, NULL); |
| | | FUNASR_RESULT result=FunASRInfer(asr_handle, wav_list[i].c_str(), RASR_NONE, NULL, 16000); |
| | | FUNASR_RESULT result=FunOfflineInfer(asr_handle, wav_list[i].c_str(), RASR_NONE, NULL, 16000); |
| | | |
| | | gettimeofday(&end, NULL); |
| | | seconds = (end.tv_sec - start.tv_sec); |
| | |
| | | TCLAP::CmdLine cmd("funasr-onnx-offline-rtf", ' ', "1.0"); |
| | | TCLAP::ValueArg<std::string> model_dir("", MODEL_DIR, "the model path, which contains model.onnx, config.yaml, am.mvn", true, "", "string"); |
| | | TCLAP::ValueArg<std::string> quantize("", QUANTIZE, "false (Default), load the model of model.onnx in model_dir. If set true, load the model of model_quant.onnx in model_dir", false, "false", "string"); |
| | | TCLAP::ValueArg<std::string> vad_dir("", VAD_DIR, "the vad model path, which contains model.onnx, vad.yaml, vad.mvn", false, "", "string"); |
| | | TCLAP::ValueArg<std::string> vad_quant("", VAD_QUANT, "false (Default), load the model of model.onnx in vad_dir. If set true, load the model of model_quant.onnx in vad_dir", false, "false", "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, "false (Default), load the model of model.onnx in punc_dir. If set true, load the model of model_quant.onnx in punc_dir", false, "false", "string"); |
| | | |
| | | TCLAP::ValueArg<std::string> wav_path("", WAV_PATH, "the input could be: wav_path, e.g.: asr_example.wav; pcm_path, e.g.: asr_example.pcm; wav.scp, kaldi style wav list (wav_id \t wav_path)", true, "", "string"); |
| | | TCLAP::ValueArg<std::int32_t> thread_num("", THREAD_NUM, "multi-thread num for rtf", true, 0, "int32_t"); |
| | | |
| | | cmd.add(model_dir); |
| | | cmd.add(quantize); |
| | | cmd.add(vad_dir); |
| | | cmd.add(vad_quant); |
| | | cmd.add(punc_dir); |
| | | cmd.add(punc_quant); |
| | | cmd.add(wav_path); |
| | | cmd.add(thread_num); |
| | | cmd.parse(argc, argv); |
| | |
| | | std::map<std::string, std::string> model_path; |
| | | GetValue(model_dir, MODEL_DIR, model_path); |
| | | GetValue(quantize, QUANTIZE, model_path); |
| | | GetValue(vad_dir, VAD_DIR, model_path); |
| | | GetValue(vad_quant, VAD_QUANT, model_path); |
| | | GetValue(punc_dir, PUNC_DIR, model_path); |
| | | GetValue(punc_quant, PUNC_QUANT, model_path); |
| | | GetValue(wav_path, WAV_PATH, model_path); |
| | | |
| | | struct timeval start, end; |
| | | gettimeofday(&start, NULL); |
| | | FUNASR_HANDLE asr_handle=FunASRInit(model_path, 1); |
| | | FUNASR_HANDLE asr_handle=FunOfflineInit(model_path, 1); |
| | | |
| | | if (!asr_handle) |
| | | { |
| | |
| | | long modle_init_micros = ((seconds * 1000000) + end.tv_usec) - (start.tv_usec); |
| | | LOG(INFO) << "Model initialization takes " << (double)modle_init_micros / 1000000 << " s"; |
| | | |
| | | // read wav_scp |
| | | // read wav_path |
| | | vector<string> wav_list; |
| | | string wav_path_ = model_path.at(WAV_PATH); |
| | | if(is_target_file(wav_path_, "wav") || is_target_file(wav_path_, "pcm")){ |
| | |
| | | LOG(INFO) << "total_rtf " << (double)total_time/ (total_length*1000000); |
| | | LOG(INFO) << "speedup " << 1.0/((double)total_time/ (total_length*1000000)); |
| | | |
| | | FunASRUninit(asr_handle); |
| | | FunOfflineUninit(asr_handle); |
| | | return 0; |
| | | } |
| | |
| | | |
| | | vector<float> Paraformer::FbankKaldi(float sample_rate, const float* waves, int len) { |
| | | knf::OnlineFbank fbank_(fbank_opts); |
| | | fbank_.AcceptWaveform(sample_rate, waves, len); |
| | | std::vector<float> buf(len); |
| | | for (int32_t i = 0; i != len; ++i) { |
| | | buf[i] = waves[i] * 32768; |
| | | } |
| | | fbank_.AcceptWaveform(sample_rate, buf.data(), buf.size()); |
| | | //fbank_->InputFinished(); |
| | | int32_t frames = fbank_.NumFramesReady(); |
| | | int32_t feature_dim = fbank_opts.mel_opts.num_bins; |
| | |
| | | make |
| | | ``` |
| | | |
| | | #### Recipe |
| | | |
| | | set the model, data path and output_dir |
| | | |
| | | ```shell |
| | | ./bin/funasr-onnx-offline-rtf /path/to/model_dir /path/to/wav.scp quantize(true or false) thread_num |
| | | ``` |
| | | |
| | | The structure of /path/to/models_dir |
| | | ``` |
| | | config.yaml, am.mvn, model.onnx(or model_quant.onnx) |
| | | ``` |
| | | |
| | | ## [Paraformer-large](https://www.modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary) |
| | | ```shell |
| | | ./funasr-onnx-offline-rtf \ |
| | | --model-dir ./asrmodel/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch \ |
| | | --quantize true \ |
| | | --wav-path ./aishell1_test.scp \ |
| | | --thread-num 32 |
| | | |
| | | Node: '--quantize false' means fp32, otherwise it will be int8 |
| | | ``` |
| | | |
| | | Number of Parameter: 220M |
| | | |
| | |
| | | ### Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHz 32core-64processor without avx512_vnni |
| | | |
| | | | concurrent-tasks | processing time(s) | RTF | Speedup Rate | |
| | | |---------------------|--------------------|----------|--------------| |
| | | |---------------------|:------------------:|----------|:------------:| |
| | | | 1 (onnx fp32) | 2903s | 0.080404 | 12 | |
| | | | 1 (onnx int8) | 2714s | 0.075168 | 13 | |
| | | | 8 (onnx fp32) | 373s | 0.010329 | 97 | |
| | |
| | | | 96 (onnx fp32) | 115s | 0.003183 | 314 | |
| | | | 96 (onnx int8) | 80s | 0.002222 | 450 | |
| | | |
| | | ## [FSMN-VAD](https://www.modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/summary) + [Paraformer-large](https://www.modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary) + [CT-Transformer](https://www.modelscope.cn/models/damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch/summary) |
| | | |
| | | ```shell |
| | | ./funasr-onnx-offline-rtf \ |
| | | --model-dir ./asrmodel/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch \ |
| | | --quantize true \ |
| | | --vad-dir ./asrmodel/speech_fsmn_vad_zh-cn-16k-common-pytorch \ |
| | | --punc-dir ./asrmodel/punc_ct-transformer_zh-cn-common-vocab272727-pytorch \ |
| | | --wav-path ./aishell1_test.scp \ |
| | | --thread-num 32 |
| | | |
| | | Node: '--quantize false' means fp32, otherwise it will be int8 |
| | | ``` |
| | | |
| | | ### Intel(R) Xeon(R) Platinum 8369B CPU @ 2.90GHz 16core-32processor with avx512_vnni |
| | | |
| | | | concurrent-tasks | processing time(s) | RTF | Speedup Rate | |
| | | |---------------------|:------------------:|:--------:|:------------:| |
| | | | 1 (onnx fp32) | 2134s | 0.0591 | 17 | |
| | | | 1 (onnx int8) | 1047s | 0.029 | 34 | |
| | | | 8 (onnx fp32) | 273s | 0.007557 | 132 | |
| | | | 8 (onnx int8) | 132s | 0.003647 | 274 | |
| | | | 16 (onnx fp32) | 147s | 0.004061 | 246 | |
| | | | 16 (onnx int8) | 69s | 0.001916 | 521 | |
| | | | 32 (onnx fp32) | 133s | 0.003675 | 272 | |
| | | | 32 (onnx int8) | 65s | 0.001786 | 559 | |
| | | | 64 (onnx fp32) | 136s | 0.003767 | 265 | |
| | | | 64 (onnx int8) | 67s | 0.001867 | 535 | |
| | | | 96 (onnx fp32) | 137s | 0.003802 | 262 | |
| | | | 96 (onnx int8) | 69s | 0.001904 | 524 | |
| | | |
| | | |
| | | |
| | | ### Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHz 32core-64processor without avx512_vnni |
| | | |
| | | | concurrent-tasks | processing time(s) | RTF | Speedup Rate | |
| | | |---------------------|:------------------:|----------|:------------:| |
| | | | 1 (onnx fp32) | 3073s | 0.0851 | 12 | |
| | | | 1 (onnx int8) | 2840s | 0.0787 | 13 | |
| | | | 8 (onnx fp32) | 389s | 0.01079 | 93 | |
| | | | 8 (onnx int8) | 355s | 0.0098 | 101 | |
| | | | 16 (onnx fp32) | 199s | 0.005513 | 181 | |
| | | | 16 (onnx int8) | 171s | 0.004784 | 210 | |
| | | | 32 (onnx fp32) | 113s | 0.00314 | 318 | |
| | | | 32 (onnx int8) | 92s | 0.00255 | 391 | |
| | | | 64 (onnx fp32) | 115s | 0.0032 | 312 | |
| | | | 64 (onnx int8) | 81s | 0.002232 | 448 | |
| | | | 96 (onnx fp32) | 117s | 0.003257 | 307 | |
| | | | 96 (onnx int8) | 81s | 0.002258 | 442 | |
| | |
| | | ncpu=args.ncpu, |
| | | model_revision='v1.0.4') |
| | | |
| | | # vad |
| | | inference_pipeline_vad = pipeline( |
| | | task=Tasks.voice_activity_detection, |
| | | model=args.vad_model, |
| | | model_revision=None, |
| | | output_dir=None, |
| | | batch_size=1, |
| | | mode='online', |
| | | ngpu=args.ngpu, |
| | | ncpu=1, |
| | | ) |
| | | |
| | | print("model loaded") |
| | | |
| | | |
| | | |
| | | async def ws_serve(websocket, path): |
| | | frames = [] |
| | | frames_asr_online = [] |
| | | global websocket_users |
| | | websocket_users.add(websocket) |
| | | websocket.param_dict_asr_online = {"cache": dict()} |
| | | websocket.param_dict_vad = {'in_cache': dict()} |
| | | websocket.wav_name = "microphone" |
| | | print("new user connected",flush=True) |
| | | try: |
| | |
| | | if "is_speaking" in messagejson: |
| | | websocket.is_speaking = messagejson["is_speaking"] |
| | | websocket.param_dict_asr_online["is_final"] = not websocket.is_speaking |
| | | websocket.param_dict_vad["is_final"] = not websocket.is_speaking |
| | | # need to fire engine manually if no data received any more |
| | | if not websocket.is_speaking: |
| | | await async_asr_online(websocket,b"") |
| | |
| | | if len(frames_asr_online) > 0 or not isinstance(message, str): |
| | | if not isinstance(message,str): |
| | | frames_asr_online.append(message) |
| | | # frames.append(message) |
| | | # duration_ms = len(message) // 32 |
| | | # websocket.vad_pre_idx += duration_ms |
| | | speech_start_i, speech_end_i = await async_vad(websocket, message) |
| | | websocket.is_speaking = not speech_end_i |
| | | |
| | | if len(frames_asr_online) % websocket.chunk_interval == 0 or not websocket.is_speaking: |
| | | websocket.param_dict_asr_online["is_final"] = not websocket.is_speaking |
| | | audio_in = b"".join(frames_asr_online) |
| | | # if not websocket.is_speaking: |
| | | #padding 0.5s at end gurantee that asr engine can fire out last word |
| | | # audio_in=audio_in+b''.join(np.zeros(int(16000*0.5),dtype=np.int16)) |
| | | await async_asr_online(websocket,audio_in) |
| | | frames_asr_online = [] |
| | | |
| | |
| | | await websocket.send(message) |
| | | |
| | | |
| | | async def async_vad(websocket, audio_in): |
| | | segments_result = inference_pipeline_vad(audio_in=audio_in, param_dict=websocket.param_dict_vad) |
| | | |
| | | speech_start = False |
| | | speech_end = False |
| | | |
| | | if len(segments_result) == 0 or len(segments_result["text"]) > 1: |
| | | return speech_start, speech_end |
| | | if segments_result["text"][0][0] != -1: |
| | | speech_start = segments_result["text"][0][0] |
| | | if segments_result["text"][0][1] != -1: |
| | | speech_end = True |
| | | return speech_start, speech_end |
| | | |
| | | if len(args.certfile)>0: |
| | | ssl_context = ssl.SSLContext(ssl.PROTOCOL_TLS_SERVER) |
| | | |