Merge pull request #1750 from zh794390558/spx

[speechx] fix model params path name

demos/streaming_asr_server/README.md

@@ -352,4 +352,4 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav
         [2022-04-21 15:59:08,024] [    INFO] - receive msg={'asr_results': '我认为跑步最重要的就是给我带来了身体健康'}
         [2022-04-21 15:59:12,883] [    INFO] - final receive msg={'status': 'ok', 'signal': 'finished', 'asr_results': '我认为跑步最重要的就是给我带来了身体健康'}
         [2022-04-21 15:59:12,884] [    INFO] - 我认为跑步最重要的就是给我带来了身体健康
-  ```
\ No newline at end of file
+  ```

demos/streaming_asr_server/README_cn.md

@@ -353,4 +353,4 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav
         [2022-04-21 15:59:08,024] [    INFO] - receive msg={'asr_results': '我认为跑步最重要的就是给我带来了身体健康'}
         [2022-04-21 15:59:12,883] [    INFO] - final receive msg={'status': 'ok', 'signal': 'finished', 'asr_results': '我认为跑步最重要的就是给我带来了身体健康'}
         [2022-04-21 15:59:12,884] [    INFO] - 我认为跑步最重要的就是给我带来了身体健康
-  ```
\ No newline at end of file
+  ```

speechx/examples/ds2_ol/README.md

@@ -4,6 +4,8 @@ Please go to `aishell` to test it.
 
 * aishell
 Deepspeech2 Streaming Decoding under aishell dataset.
+* websocket
+Streaming ASR with websocket.
 
 The below is for developing and offline testing:
 * nnet

speechx/examples/ds2_ol/aishell/README.md

@@ -33,4 +33,4 @@ LM: [wenetspeech](http://paddlespeech.bj.bcebos.com/speechx/examples/ds2_ol/aish
 Overall -> 10.93 % N=104765 C=93410 S=9780 D=1575 I=95
 Mandarin -> 10.93 % N=104762 C=93410 S=9779 D=1573 I=95
 Other -> 100.00 % N=3 C=0 S=1 D=2 I=0
-```
\ No newline at end of file
+```

speechx/examples/ds2_ol/aishell/run.sh

@@ -119,7 +119,7 @@ if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
         pushd $wfst
         wget -c https://paddlespeech.bj.bcebos.com/s2t/paddle_asr_online/aishell_graph.zip
         unzip aishell_graph.zip
-        mv aishell_graph/* 
+        mv aishell_graph/* $wfst
         popd
     fi
 

speechx/examples/ds2_ol/decoder/ctc-prefix-beam-search-decoder-ol.cc

@@ -76,7 +76,7 @@ int main(int argc, char* argv[]) {
 
     ppspeech::ModelOptions model_opts;
     model_opts.model_path = model_path;
-    model_opts.params_path = model_params;
+    model_opts.param_path = model_params;
     model_opts.cache_shape = FLAGS_model_cache_names;
     model_opts.input_names = FLAGS_model_input_names;
     model_opts.output_names = FLAGS_model_output_names;

speechx/examples/ds2_ol/decoder/recognizer_test_main.cc

@@ -60,6 +60,7 @@ int main(int argc, char* argv[]) {
             for (int i = 0; i < cur_chunk_size; ++i) {
                 wav_chunk(i) = waveform(sample_offset + i);
             }
+            // wav_chunk = waveform.Range(sample_offset + i, cur_chunk_size);
 
             recognizer.Accept(wav_chunk);
             if (cur_chunk_size < chunk_sample_size) {
@@ -67,8 +68,10 @@ int main(int argc, char* argv[]) {
             }
             recognizer.Decode();
 
+            // no overlap
             sample_offset += cur_chunk_size;
         }
+
         std::string result;
         result = recognizer.GetFinalResult();
         recognizer.Reset();

speechx/examples/ds2_ol/decoder/wfst-decoder-ol.cc

@@ -79,7 +79,7 @@ int main(int argc, char* argv[]) {
 
     ppspeech::ModelOptions model_opts;
     model_opts.model_path = model_graph;
-    model_opts.params_path = model_params;
+    model_opts.param_path = model_params;
     model_opts.cache_shape = FLAGS_model_cache_names;
     model_opts.input_names = FLAGS_model_input_names;
     model_opts.output_names = FLAGS_model_output_names;

speechx/examples/ds2_ol/websocket/.gitignore

+data
+exp

speechx/examples/ds2_ol/websocket/CMakeLists.txt

@@ -6,5 +6,4 @@ target_link_libraries(websocket_server_main PUBLIC frontend kaldi-feat-common nn
 
 add_executable(websocket_client_main ${CMAKE_CURRENT_SOURCE_DIR}/websocket_client_main.cc)
 target_include_directories(websocket_client_main PRIVATE ${SPEECHX_ROOT} ${SPEECHX_ROOT}/kaldi)
-target_link_libraries(websocket_client_main PUBLIC frontend kaldi-feat-common nnet decoder fst utils gflags glog kaldi-base kaldi-matrix kaldi-util kaldi-decoder websocket ${DEPS})
-
+target_link_libraries(websocket_client_main PUBLIC frontend kaldi-feat-common nnet decoder fst utils gflags glog kaldi-base kaldi-matrix kaldi-util kaldi-decoder websocket ${DEPS})
\ No newline at end of file

speechx/examples/ds2_ol/websocket/path.sh

+# This contains the locations of binarys build required for running the examples.
+
+SPEECHX_ROOT=$PWD/../../..
+SPEECHX_EXAMPLES=$SPEECHX_ROOT/build/examples
+
+SPEECHX_TOOLS=$SPEECHX_ROOT/tools
+TOOLS_BIN=$SPEECHX_TOOLS/valgrind/install/bin
+
+[ -d $SPEECHX_EXAMPLES ] || { echo "Error: 'build/examples' directory not found. please ensure that the project build successfully"; }
+
+export LC_AL=C
+
+SPEECHX_BIN=$SPEECHX_EXAMPLES/ds2_ol/websocket:$SPEECHX_EXAMPLES/ds2_ol/feat
+export PATH=$PATH:$SPEECHX_BIN:$TOOLS_BIN

speechx/examples/ds2_ol/websocket/websocket_client_main.cc

@@ -26,7 +26,6 @@ using kaldi::int16;
 int main(int argc, char* argv[]) {
     gflags::ParseCommandLineFlags(&argc, &argv, false);
     google::InitGoogleLogging(argv[0]);
-    ppspeech::WebSocketClient client(FLAGS_host, FLAGS_port);
 
     kaldi::SequentialTableReader<kaldi::WaveHolder> wav_reader(
         FLAGS_wav_rspecifier);
@@ -36,6 +35,8 @@ int main(int argc, char* argv[]) {
     const int chunk_sample_size = streaming_chunk * sample_rate;
 
     for (; !wav_reader.Done(); wav_reader.Next()) {
+        ppspeech::WebSocketClient client(FLAGS_host, FLAGS_port);
+
         client.SendStartSignal();
         std::string utt = wav_reader.Key();
         const kaldi::WaveData& wave_data = wav_reader.Value();
@@ -74,9 +75,8 @@ int main(int argc, char* argv[]) {
         std::string result = client.GetResult();
         LOG(INFO) << "utt: " << utt << " " << result;
 
-
         client.Join();
-        return 0;
     }
+
     return 0;
 }

speechx/examples/ds2_ol/aishell/websocket_server.sh → speechx/examples/ds2_ol/websocket/websocket_server.sh

@@ -19,38 +19,29 @@ ckpt_dir=$data/model
 model_dir=$ckpt_dir/exp/deepspeech2_online/checkpoints/
 vocb_dir=$ckpt_dir/data/lang_char/
 
-# output
-aishell_wav_scp=aishell_test.scp
-if [ ! -d $data/test ]; then
-    pushd $data
-    wget -c https://paddlespeech.bj.bcebos.com/s2t/paddle_asr_online/aishell_test.zip
-    unzip  aishell_test.zip
-    popd
-
-    realpath $data/test/*/*.wav > $data/wavlist
-    awk -F '/' '{ print $(NF) }' $data/wavlist | awk -F '.' '{ print $1 }' > $data/utt_id
-    paste $data/utt_id $data/wavlist > $data/$aishell_wav_scp
-fi
-
-
-if [ ! -d $ckpt_dir ]; then
-    mkdir -p $ckpt_dir
-    wget -P $ckpt_dir -c https://paddlespeech.bj.bcebos.com/s2t/aishell/asr0/asr0_deepspeech2_online_aishell_ckpt_0.2.0.model.tar.gz
-    tar xzfv $ckpt_dir/asr0_deepspeech2_online_aishell_ckpt_0.2.0.model.tar.gz -C $ckpt_dir
+if [ ! -f $ckpt_dir/data/mean_std.json ]; then
+        mkdir -p $ckpt_dir
+        pushd $ckpt_dir
+        wget -c https://paddlespeech.bj.bcebos.com/s2t/aishell/asr0/asr0_deepspeech2_online_aishell_ckpt_0.2.0.model.tar.gz
+        tar xzfv asr0_deepspeech2_online_aishell_ckpt_0.2.0.model.tar.gz 
+        popd
 fi
 
-
 export GLOG_logtostderr=1
 
 # 3. gen cmvn 
-cmvn=$PWD/cmvn.ark
+cmvn=$data/cmvn.ark
 cmvn-json2kaldi --json_file=$ckpt_dir/data/mean_std.json --cmvn_write_path=$cmvn
 
-text=$data/test/text
-graph_dir=./aishell_graph
-if [ ! -d $graph_dir ]; then
+
+wfst=$data/wfst/
+mkdir -p $wfst
+if [ ! -f $wfst/aishell_graph.zip ]; then
+    pushd $wfst
     wget -c https://paddlespeech.bj.bcebos.com/s2t/paddle_asr_online/aishell_graph.zip
-    unzip aishell_graph.zip 
+    unzip aishell_graph.zip
+    mv aishell_graph/* $wfst
+    popd
 fi
 
 # 5. test websocket server 
@@ -59,8 +50,8 @@ websocket_server_main \
     --model_path=$model_dir/avg_1.jit.pdmodel \
     --streaming_chunk=0.1 \
     --convert2PCM32=true \
-    --params_path=$model_dir/avg_1.jit.pdiparams \
-    --word_symbol_table=$graph_dir/words.txt \
+    --param_path=$model_dir/avg_1.jit.pdiparams \
+    --word_symbol_table=$data/wfst/words.txt \
     --model_output_names=softmax_0.tmp_0,tmp_5,concat_0.tmp_0,concat_1.tmp_0 \
-    --graph_path=$graph_dir/TLG.fst --max_active=7500 \
+    --graph_path=$data/wfst/TLG.fst --max_active=7500 \
     --acoustic_scale=1.2 

speechx/speechx/decoder/param.h

@@ -15,7 +15,6 @@
 #pragma once
 
 #include "base/common.h"
-
 #include "decoder/ctc_beam_search_decoder.h"
 #include "decoder/ctc_tlg_decoder.h"
 #include "frontend/audio/feature_pipeline.h"
@@ -24,7 +23,7 @@ DEFINE_string(cmvn_file, "", "read cmvn");
 DEFINE_double(streaming_chunk, 0.1, "streaming feature chunk size");
 DEFINE_bool(convert2PCM32, true, "audio convert to pcm32");
 DEFINE_string(model_path, "avg_1.jit.pdmodel", "paddle nnet model");
-DEFINE_string(params_path, "avg_1.jit.pdiparams", "paddle nnet model param");
+DEFINE_string(param_path, "avg_1.jit.pdiparams", "paddle nnet model param");
 DEFINE_string(word_symbol_table, "words.txt", "word symbol table");
 DEFINE_string(graph_path, "TLG", "decoder graph");
 DEFINE_double(acoustic_scale, 1.0, "acoustic scale");
@@ -37,13 +36,16 @@ DEFINE_int32(receptive_field_length,
 DEFINE_int32(downsampling_rate,
              4,
              "two CNN(kernel=5) module downsampling rate.");
+DEFINE_string(
+    model_input_names,
+    "audio_chunk,audio_chunk_lens,chunk_state_h_box,chunk_state_c_box",
+    "model input names");
 DEFINE_string(model_output_names,
-              "save_infer_model/scale_0.tmp_1,save_infer_model/"
-              "scale_1.tmp_1,save_infer_model/scale_2.tmp_1,save_infer_model/"
-              "scale_3.tmp_1",
+              "softmax_0.tmp_0,tmp_5,concat_0.tmp_0,concat_1.tmp_0",
               "model output names");
 DEFINE_string(model_cache_names, "5-1-1024,5-1-1024", "model cache names");
 
+
 namespace ppspeech {
 // todo refactor later
 FeaturePipelineOptions InitFeaturePipelineOptions() {
@@ -67,7 +69,7 @@ FeaturePipelineOptions InitFeaturePipelineOptions() {
 ModelOptions InitModelOptions() {
     ModelOptions model_opts;
     model_opts.model_path = FLAGS_model_path;
-    model_opts.params_path = FLAGS_params_path;
+    model_opts.param_path = FLAGS_param_path;
     model_opts.cache_shape = FLAGS_model_cache_names;
     model_opts.output_names = FLAGS_model_output_names;
     return model_opts;

speechx/speechx/nnet/paddle_nnet.cc

@@ -49,7 +49,7 @@ void PaddleNnet::InitCacheEncouts(const ModelOptions& opts) {
 
 PaddleNnet::PaddleNnet(const ModelOptions& opts) : opts_(opts) {
     paddle_infer::Config config;
-    config.SetModel(opts.model_path, opts.params_path);
+    config.SetModel(opts.model_path, opts.param_path);
     if (opts.use_gpu) {
         config.EnableUseGpu(500, 0);
     }

speechx/speechx/nnet/paddle_nnet.h

@@ -11,25 +11,19 @@
 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 // See the License for the specific language governing permissions and
 // limitations under the License.
-
-
 #pragma once
-
-
+#include <numeric>
+#include "base/common.h"
 #include "kaldi/matrix/kaldi-matrix.h"
 #include "kaldi/util/options-itf.h"
-
-#include "base/common.h"
 #include "nnet/nnet_itf.h"
 #include "paddle_inference_api.h"
 
-#include <numeric>
-
 namespace ppspeech {
 
 struct ModelOptions {
     std::string model_path;
-    std::string params_path;
+    std::string param_path;
     int thread_num;
     bool use_gpu;
     bool switch_ir_optim;
@@ -41,7 +35,7 @@ struct ModelOptions {
     bool enable_profile;
     ModelOptions()
         : model_path("avg_1.jit.pdmodel"),
-          params_path("avg_1.jit.pdiparams"),
+          param_path("avg_1.jit.pdiparams"),
           thread_num(2),
           use_gpu(false),
           input_names(
@@ -59,7 +53,7 @@ struct ModelOptions {
 
     void Register(kaldi::OptionsItf* opts) {
         opts->Register("model-path", &model_path, "model file path");
-        opts->Register("model-params", &params_path, "params model file path");
+        opts->Register("model-param", &param_path, "params model file path");
         opts->Register("thread-num", &thread_num, "thread num");
         opts->Register("use-gpu", &use_gpu, "if use gpu");
         opts->Register("input-names", &input_names, "paddle input names");

utils/format_rsl.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import argparse
+
 import jsonlines