[TTS]Add TTS Paddle-Lite x86 inference (#2667)

* Add export2lite, test=tts

* add tts paddlelite x86 inference, test=tts

* update released_model.md, test=tts

* add paddlelite in setup.py

* update

examples/aishell3/tts3/README.md

@@ -226,6 +226,9 @@ The static model can be downloaded here:
 The ONNX model can be downloaded here:
 - [fastspeech2_aishell3_onnx_1.1.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_aishell3_onnx_1.1.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+- [fastspeech2_aishell3_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_aishell3_pdlite_1.3.0.zip)
+
 FastSpeech2 checkpoint contains files listed below.
 
 ```text

examples/aishell3/tts3/local/lite_predict.sh

+#!/bin/bash
+
+train_output_path=$1
+
+stage=0
+stop_stage=0
+
+# pwgan
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    python3 ${BIN_DIR}/../lite_predict.py \
+        --inference_dir=${train_output_path}/pdlite \
+        --am=fastspeech2_aishell3 \
+        --voc=pwgan_aishell3 \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/lite_infer_out \
+        --phones_dict=dump/phone_id_map.txt \
+        --speaker_dict=dump/speaker_id_map.txt \
+        --spk_id=0
+fi
+
+# hifigan
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    python3 ${BIN_DIR}/../lite_predict.py \
+        --inference_dir=${train_output_path}/pdlite \
+        --am=fastspeech2_aishell3 \
+        --voc=hifigan_aishell3 \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/lite_infer_out \
+        --phones_dict=dump/phone_id_map.txt \
+        --speaker_dict=dump/speaker_id_map.txt \
+        --spk_id=0
+fi

examples/aishell3/tts3/run.sh

@@ -60,11 +60,11 @@ if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
 fi
 
 if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
-    # This model is not supported, because 3 ops are not supported on 'arm'. These unsupported ops are: 'round, set_value, share_data'.
-    # This model is not supported, because 4 ops are not supported on 'x86'. These unsupported ops are: 'matmul_v2, round, set_value, share_data'.
-    # ./local/export2lite.sh ${train_output_path} inference pdlite fastspeech2_aishell3 x86
-    # x86 ok, arm Segmentation fault
-    # ./local/export2lite.sh ${train_output_path} inference pdlite pwgan_aishell3 x86  
-    # x86 ok, arm ok
-    ./local/export2lite.sh ${train_output_path} inference pdlite hifigan_aishell3 x86
+    ./local/export2lite.sh ${train_output_path} inference pdlite fastspeech2_aishell3 x86
+    ./local/export2lite.sh ${train_output_path} inference pdlite pwgan_aishell3 x86
+    # ./local/export2lite.sh ${train_output_path} inference pdlite hifigan_aishell3 x86
+fi
+
+if [ ${stage} -le 8 ] && [ ${stop_stage} -ge 8 ]; then
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/lite_predict.sh ${train_output_path} || exit -1
 fi

examples/aishell3/voc1/README.md

@@ -139,6 +139,9 @@ The static model can be downloaded here:
 The ONNX model can be downloaded here:
 - [pwgan_aishell3_onnx_1.1.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/pwgan/pwgan_aishell3_onnx_1.1.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+- [pwgan_aishell3_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/pwgan/pwgan_aishell3_pdlite_1.3.0.zip)
+
 Model | Step | eval/generator_loss | eval/log_stft_magnitude_loss:| eval/spectral_convergence_loss 
 :-------------:| :------------:| :-----: | :-----: | :--------:
 default| 1(gpu) x 400000|1.968762|0.759008|0.218524

examples/aishell3/voc5/README.md

@@ -122,6 +122,9 @@ The static model can be downloaded here:
 The ONNX model can be downloaded here:
 - [hifigan_aishell3_onnx_1.1.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/hifigan/hifigan_aishell3_onnx_1.1.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+- [hifigan_aishell3_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/hifigan/hifigan_aishell3_pdlite_1.3.0.zip)
+
 Model | Step | eval/generator_loss | eval/mel_loss| eval/feature_matching_loss
 :-------------:| :------------:| :-----: | :-----: | :--------:
 default| 1(gpu) x 2500000|24.060|0.1068|7.499

examples/csmsc/tts2/README.md

@@ -230,6 +230,9 @@ The static model can be downloaded here:
 The ONNX model can be downloaded here:
 - [speedyspeech_csmsc_onnx_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/speedyspeech/speedyspeech_csmsc_onnx_0.2.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+- [speedyspeech_csmsc_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/speedyspeech/speedyspeech_csmsc_pdlite_1.3.0.zip)
+
 
 Model | Step | eval/loss | eval/l1_loss | eval/duration_loss | eval/ssim_loss
 :-------------:| :------------:| :-----: | :-----: | :--------:|:--------:

examples/csmsc/tts2/local/lite_predict.sh

+#!/bin/bash
+
+train_output_path=$1
+
+stage=0
+stop_stage=0
+
+# pwgan
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    python3 ${BIN_DIR}/../lite_predict.py \
+        --inference_dir=${train_output_path}/pdlite \
+        --am=speedyspeech_csmsc \
+        --voc=pwgan_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/lite_infer_out \
+        --phones_dict=dump/phone_id_map.txt \
+        --tones_dict=dump/tone_id_map.txt
+fi
+
+# for more GAN Vocoders
+# multi band melgan
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    python3 ${BIN_DIR}/../lite_predict.py \
+        --inference_dir=${train_output_path}/pdlite \
+        --am=speedyspeech_csmsc \
+        --voc=mb_melgan_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/lite_infer_out \
+        --phones_dict=dump/phone_id_map.txt \
+        --tones_dict=dump/tone_id_map.txt
+fi
+
+# hifigan
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    python3 ${BIN_DIR}/../lite_predict.py \
+        --inference_dir=${train_output_path}/pdlite \
+        --am=speedyspeech_csmsc \
+        --voc=hifigan_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/lite_infer_out \
+        --phones_dict=dump/phone_id_map.txt \
+        --tones_dict=dump/tone_id_map.txt
+fi

examples/csmsc/tts2/run.sh

@@ -63,13 +63,12 @@ fi
 
 # must run after stage 3 (which stage generated static models)
 if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
-    # This model is not supported, because 3 ops are not supported on 'arm'. These unsupported ops are: 'round, set_value, share_data'.
-    # This model is not supported, because 4 ops are not supported on 'x86'. These unsupported ops are: 'matmul_v2, round, set_value, share_data'.
     ./local/export2lite.sh ${train_output_path} inference pdlite speedyspeech_csmsc x86
-    # x86 ok, arm Segmentation fault
-    # ./local/export2lite.sh ${train_output_path} inference pdlite pwgan_csmsc x86
-    # x86 ok, arm Segmentation fault
+    ./local/export2lite.sh ${train_output_path} inference pdlite pwgan_csmsc x86
     # ./local/export2lite.sh ${train_output_path} inference pdlite mb_melgan_csmsc x86
-    # x86 ok, arm ok
     # ./local/export2lite.sh ${train_output_path} inference pdlite hifigan_csmsc x86
 fi
+
+if [ ${stage} -le 8 ] && [ ${stop_stage} -ge 8 ]; then
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/lite_predict.sh ${train_output_path} || exit -1
+fi

examples/csmsc/tts3/README.md

@@ -238,6 +238,12 @@ The ONNX model can be downloaded here:
 - [fastspeech2_cnndecoder_csmsc_onnx_1.0.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_cnndecoder_csmsc_onnx_1.0.0.zip)
 - [fastspeech2_cnndecoder_csmsc_streaming_onnx_1.0.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_cnndecoder_csmsc_streaming_onnx_1.0.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+> please compile develop version of Paddle-Lite to export and run TTS models, cause TTS models are supported by https://github.com/PaddlePaddle/Paddle-Lite/pull/9587 and https://github.com/PaddlePaddle/Paddle-Lite/pull/9706
+- [fastspeech2_csmsc_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_csmsc_pdlite_1.3.0.zip)
+- [fastspeech2_cnndecoder_csmsc_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_cnndecoder_csmsc_pdlite_1.3.0.zip)
+- [fastspeech2_cnndecoder_csmsc_streaming_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_cnndecoder_csmsc_streaming_pdlite_1.3.0.zip)
+
 Model | Step | eval/loss | eval/l1_loss | eval/duration_loss | eval/pitch_loss| eval/energy_loss 
 :-------------:| :------------:| :-----: | :-----: | :--------: |:--------:|:---------:
 default| 2(gpu) x 76000|1.0991|0.59132|0.035815|0.31915|0.15287|

examples/csmsc/tts3/local/export2lite.sh

@@ -7,12 +7,12 @@ valid_targets=$5
 model_name=${model%_*}
 echo model_name: ${model_name}
 
-
+suffix=${valid_targets%,*}
 
 mkdir -p ${train_output_path}/${output_dir}
 
 paddle_lite_opt \
     --model_file ${train_output_path}/${model_dir}/${model}.pdmodel \
     --param_file  ${train_output_path}/${model_dir}/${model}.pdiparams \
-    --optimize_out ${train_output_path}/${output_dir}/${model}_${valid_targets} \
+    --optimize_out ${train_output_path}/${output_dir}/${model}_${suffix} \
     --valid_targets ${valid_targets}

examples/csmsc/tts3/local/lite_predict.sh

+#!/bin/bash
+
+train_output_path=$1
+
+stage=0
+stop_stage=0
+
+# pwgan
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    python3 ${BIN_DIR}/../lite_predict.py \
+        --inference_dir=${train_output_path}/pdlite \
+        --am=fastspeech2_csmsc \
+        --voc=pwgan_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/lite_infer_out \
+        --phones_dict=dump/phone_id_map.txt
+fi
+
+# for more GAN Vocoders
+# multi band melgan
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    python3 ${BIN_DIR}/../lite_predict.py \
+        --inference_dir=${train_output_path}/pdlite \
+        --am=fastspeech2_csmsc \
+        --voc=mb_melgan_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/lite_infer_out \
+        --phones_dict=dump/phone_id_map.txt
+fi
+
+# hifigan
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    python3 ${BIN_DIR}/../lite_predict.py \
+        --inference_dir=${train_output_path}/pdlite \
+        --am=fastspeech2_csmsc \
+        --voc=hifigan_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/lite_infer_out \
+        --phones_dict=dump/phone_id_map.txt
+fi

examples/csmsc/tts3/local/lite_predict_streaming.sh

+#!/bin/bash
+
+train_output_path=$1
+
+stage=0
+stop_stage=0
+
+# pwgan
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    python3 ${BIN_DIR}/../lite_predict_streaming.py \
+        --inference_dir=${train_output_path}/pdlite_streaming \
+        --am=fastspeech2_csmsc \
+        --am_stat=dump/train/speech_stats.npy \
+        --voc=pwgan_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/lite_infer_out_streaming \
+        --phones_dict=dump/phone_id_map.txt \
+        --am_streaming=True
+fi
+
+# for more GAN Vocoders
+# multi band melgan
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    python3 ${BIN_DIR}/../lite_predict_streaming.py \
+        --inference_dir=${train_output_path}/pdlite_streaming \
+        --am=fastspeech2_csmsc \
+        --am_stat=dump/train/speech_stats.npy \
+        --voc=mb_melgan_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/lite_infer_out_streaming \
+        --phones_dict=dump/phone_id_map.txt \
+        --am_streaming=True
+fi
+
+# hifigan
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    python3 ${BIN_DIR}/../lite_predict_streaming.py \
+        --inference_dir=${train_output_path}/pdlite_streaming \
+        --am=fastspeech2_csmsc \
+        --am_stat=dump/train/speech_stats.npy \
+        --voc=hifigan_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/lite_infer_out_streaming \
+        --phones_dict=dump/phone_id_map.txt \
+        --am_streaming=True
+fi
+

examples/csmsc/tts3/run.sh

@@ -64,13 +64,15 @@ fi
 
 # must run after stage 3 (which stage generated static models)
 if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
-    # This model is not supported, because 3 ops are not supported on 'arm'. These unsupported ops are: 'round, set_value, share_data'.
-    # This model is not supported, because 4 ops are not supported on 'x86'. These unsupported ops are: 'matmul_v2, round, set_value, share_data'.
+    # NOTE by yuantian 2022.11.21: please compile develop version of Paddle-Lite to export and run TTS models,
+    #                   cause TTS models are supported by https://github.com/PaddlePaddle/Paddle-Lite/pull/9587 
+    #                   and https://github.com/PaddlePaddle/Paddle-Lite/pull/9706
     ./local/export2lite.sh ${train_output_path} inference pdlite fastspeech2_csmsc x86
-    # x86 ok, arm Segmentation fault
-    # ./local/export2lite.sh ${train_output_path} inference pdlite pwgan_csmsc x86
-    # x86 ok, arm Segmentation fault
+    ./local/export2lite.sh ${train_output_path} inference pdlite pwgan_csmsc x86
     # ./local/export2lite.sh ${train_output_path} inference pdlite mb_melgan_csmsc x86
-    # x86 ok, arm ok
     # ./local/export2lite.sh ${train_output_path} inference pdlite hifigan_csmsc x86
 fi
+
+if [ ${stage} -le 8 ] && [ ${stop_stage} -ge 8 ]; then
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/lite_predict.sh ${train_output_path} || exit -1
+fi

examples/csmsc/tts3/run_cnndecoder.sh

@@ -98,32 +98,27 @@ fi
 
 # must run after stage 3 (which stage generated static models)
 if [ ${stage} -le 11 ] && [ ${stop_stage} -ge 11 ]; then
-    # This model is not supported, because 3 ops are not supported on 'arm'. These unsupported ops are: 'round, set_value, share_data'.
-    # This model is not supported, because 4 ops are not supported on 'x86'. These unsupported ops are: 'matmul_v2, round, set_value, share_data'.
     ./local/export2lite.sh ${train_output_path} inference pdlite fastspeech2_csmsc x86
-    # x86 ok, arm Segmentation fault
-    # ./local/export2lite.sh ${train_output_path} inference pdlite pwgan_csmsc x86
-    # x86 ok, arm Segmentation fault
+    ./local/export2lite.sh ${train_output_path} inference pdlite pwgan_csmsc x86
     # ./local/export2lite.sh ${train_output_path} inference pdlite mb_melgan_csmsc x86
-    # x86 ok, arm ok
     # ./local/export2lite.sh ${train_output_path} inference pdlite hifigan_csmsc x86
 fi
 
-# must run after stage 5 (which stage generated static models)
 if [ ${stage} -le 12 ] && [ ${stop_stage} -ge 12 ]; then
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/lite_predict.sh ${train_output_path} || exit -1
+fi
+
+# must run after stage 5 (which stage generated static models)
+if [ ${stage} -le 13 ] && [ ${stop_stage} -ge 13 ]; then
     # streaming acoustic model
-    # This model is not supported, because 3 ops are not supported on 'arm'. These unsupported ops are: 'round, set_value, share_data'.
-    # This model is not supported, because 4 ops are not supported on 'x86'. These unsupported ops are: 'matmul_v2, round, set_value, share_data'.
-    # ./local/export2lite.sh ${train_output_path} inference pdlite fastspeech2_csmsc x86
     ./local/export2lite.sh ${train_output_path} inference_streaming pdlite_streaming fastspeech2_csmsc_am_encoder_infer x86
-    # x86 ok, arm Segmentation fault
     ./local/export2lite.sh ${train_output_path} inference_streaming pdlite_streaming fastspeech2_csmsc_am_decoder x86
-    # x86 ok, arm Segmentation fault
     ./local/export2lite.sh ${train_output_path} inference_streaming pdlite_streaming fastspeech2_csmsc_am_postnet x86
-    # x86 ok, arm Segmentation fault
-    # ./local/export2lite.sh ${train_output_path} inference_streaming pdlite_streaming pwgan_csmsc x86
-    # x86 ok, arm Segmentation fault
+    ./local/export2lite.sh ${train_output_path} inference_streaming pdlite_streaming pwgan_csmsc x86
     # ./local/export2lite.sh ${train_output_path} inference_streaming pdlite_streaming mb_melgan_csmsc x86
-    # x86 ok, arm ok
     # ./local/export2lite.sh ${train_output_path} inference_streaming pdlite_streaming hifigan_csmsc x86
 fi
+
+if [ ${stage} -le 14 ] && [ ${stop_stage} -ge 14 ]; then
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/lite_predict_streaming.sh ${train_output_path} || exit -1
+fi

examples/csmsc/voc1/README.md

@@ -136,6 +136,9 @@ The static model can be downloaded here:
 The ONNX model can be downloaded here:
 - [pwgan_csmsc_onnx_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/pwgan/pwgan_csmsc_onnx_0.2.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+- [pwgan_csmsc_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/pwgan/pwgan_csmsc_pdlite_1.3.0.zip)
+
 Model | Step | eval/generator_loss | eval/log_stft_magnitude_loss| eval/spectral_convergence_loss
 :-------------:| :------------:| :-----: | :-----: | :--------:
 default| 1(gpu) x 400000|1.948763|0.670098|0.248882

examples/csmsc/voc3/README.md

@@ -164,6 +164,9 @@ The static model can be downloaded here:
 The ONNX model can be downloaded here:
 - [mb_melgan_csmsc_onnx_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/mb_melgan/mb_melgan_csmsc_onnx_0.2.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+- [mb_melgan_csmsc_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/mb_melgan/mb_melgan_csmsc_pdlite_1.3.0.zip)
+
 Model | Step | eval/generator_loss | eval/log_stft_magnitude_loss|eval/spectral_convergence_loss |eval/sub_log_stft_magnitude_loss|eval/sub_spectral_convergence_loss
 :-------------:| :------------:| :-----: | :-----: | :--------:| :--------:| :--------:
 default| 1(gpu) x 1000000| 2.4851|0.71778 |0.2761 |0.66334 |0.2777|

examples/csmsc/voc5/README.md

@@ -121,6 +121,9 @@ The static model can be downloaded here:
 The ONNX model can be downloaded here:
 - [hifigan_csmsc_onnx_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/hifigan/hifigan_csmsc_onnx_0.2.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+- [hifigan_csmsc_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/hifigan/hifigan_csmsc_pdlite_1.3.0.zip)
+
 Model | Step | eval/generator_loss | eval/mel_loss| eval/feature_matching_loss
 :-------------:| :------------:| :-----: | :-----: | :--------:
 default| 1(gpu) x 2500000|24.927|0.1262|7.554

examples/ljspeech/tts3/README.md

@@ -221,6 +221,9 @@ The static model can be downloaded here:
 The ONNX model can be downloaded here:
 - [fastspeech2_ljspeech_onnx_1.1.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_ljspeech_onnx_1.1.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+- [fastspeech2_ljspeech_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_ljspeech_pdlite_1.3.0.zip)
+
 
 Model | Step | eval/loss | eval/l1_loss | eval/duration_loss | eval/pitch_loss| eval/energy_loss 
 :-------------:| :------------:| :-----: | :-----: | :--------: |:--------:|:---------:

examples/ljspeech/tts3/local/lite_predict.sh

+#!/bin/bash
+
+train_output_path=$1
+
+stage=0
+stop_stage=0
+
+# pwgan
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    python3 ${BIN_DIR}/../lite_predict.py \
+        --inference_dir=${train_output_path}/pdlite \
+        --am=fastspeech2_ljspeech \
+        --voc=pwgan_ljspeech \
+        --text=${BIN_DIR}/../sentences_en.txt \
+        --output_dir=${train_output_path}/lite_infer_out \
+        --phones_dict=dump/phone_id_map.txt \
+        --lang=en
+fi
+
+# hifigan
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    python3 ${BIN_DIR}/../lite_predict.py \
+        --inference_dir=${train_output_path}/pdlite \
+        --am=fastspeech2_ljspeech \
+        --voc=hifigan_ljspeech \
+        --text=${BIN_DIR}/../sentences_en.txt \
+        --output_dir=${train_output_path}/lite_infer_out \
+        --phones_dict=dump/phone_id_map.txt \
+        --lang=en
+fi

examples/ljspeech/tts3/run.sh

@@ -62,11 +62,11 @@ fi
 
 # must run after stage 3 (which stage generated static models)
 if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
-    # This model is not supported, because 3 ops are not supported on 'arm'. These unsupported ops are: 'round, set_value, share_data'.
-    # This model is not supported, because 4 ops are not supported on 'x86'. These unsupported ops are: 'matmul_v2, round, set_value, share_data'.
     ./local/export2lite.sh ${train_output_path} inference pdlite fastspeech2_ljspeech x86
-    # x86 ok, arm Segmentation fault
-    # ./local/export2lite.sh ${train_output_path} inference pdlite pwgan_ljspeech x86
-    # x86 ok, arm ok
+    ./local/export2lite.sh ${train_output_path} inference pdlite pwgan_ljspeech x86
     # ./local/export2lite.sh ${train_output_path} inference pdlite hifigan_ljspeech x86
+fi
+
+if [ ${stage} -le 8 ] && [ ${stop_stage} -ge 8 ]; then
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/lite_predict.sh ${train_output_path} || exit -1
 fi
\ No newline at end of file

examples/ljspeech/voc1/README.md

@@ -136,6 +136,9 @@ The static model can be downloaded here:
 The ONNX model can be downloaded here:
 - [pwgan_ljspeech_onnx_1.1.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/pwgan/pwgan_ljspeech_onnx_1.1.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+- [pwgan_ljspeech_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/pwgan/pwgan_ljspeech_pdlite_1.3.0.zip)
+
 
 Parallel WaveGAN checkpoint contains files listed below.
 

examples/ljspeech/voc5/README.md

@@ -121,6 +121,8 @@ The static model can be downloaded here:
 The ONNX model can be downloaded here:
 - [hifigan_ljspeech_onnx_1.1.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/hifigan/hifigan_ljspeech_onnx_1.1.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+- [hifigan_ljspeech_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/hifigan/hifigan_ljspeech_pdlite_1.3.0.zip)
 
 Model | Step | eval/generator_loss | eval/mel_loss| eval/feature_matching_loss
 :-------------:| :------------:| :-----: | :-----: | :--------:

examples/vctk/tts3/README.md

@@ -224,6 +224,9 @@ The static model can be downloaded here:
 The ONNX model can be downloaded here:
 - [fastspeech2_vctk_onnx_1.1.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_vctk_onnx_1.1.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+- [fastspeech2_vctk_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_vctk_pdlite_1.3.0.zip)
+
 FastSpeech2 checkpoint contains files listed below.
 ```text
 fastspeech2_vctk_ckpt_1.2.0

examples/vctk/tts3/local/lite_predict.sh

+#!/bin/bash
+
+train_output_path=$1
+
+stage=0
+stop_stage=0
+
+# pwgan
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    python3 ${BIN_DIR}/../lite_predict.py \
+        --inference_dir=${train_output_path}/pdlite \
+        --am=fastspeech2_vctk \
+        --voc=pwgan_vctk \
+        --text=${BIN_DIR}/../sentences_en.txt \
+        --output_dir=${train_output_path}/lite_infer_out \
+        --phones_dict=dump/phone_id_map.txt \
+        --speaker_dict=dump/speaker_id_map.txt \
+        --spk_id=0 \
+        --lang=en
+fi
+
+# hifigan
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    python3 ${BIN_DIR}/../lite_predict.py \
+        --inference_dir=${train_output_path}/pdlite \
+        --am=fastspeech2_vctk \
+        --voc=hifigan_vctk \
+        --text=${BIN_DIR}/../sentences_en.txt \
+        --output_dir=${train_output_path}/lite_infer_out \
+        --phones_dict=dump/phone_id_map.txt \
+        --speaker_dict=dump/speaker_id_map.txt \
+        --spk_id=0 \
+        --lang=en
+fi

examples/vctk/tts3/run.sh

@@ -61,11 +61,11 @@ fi
 
 # must run after stage 3 (which stage generated static models)
 if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
-    # This model is not supported, because 3 ops are not supported on 'arm'. These unsupported ops are: 'round, set_value, share_data'.
-    # This model is not supported, because 4 ops are not supported on 'x86'. These unsupported ops are: 'matmul_v2, round, set_value, share_data'.
     ./local/export2lite.sh ${train_output_path} inference pdlite fastspeech2_vctk x86
-    # x86 ok, arm Segmentation fault
-    # ./local/export2lite.sh ${train_output_path} inference pdlite pwgan_vctk x86
-    # x86 ok, arm ok
+    ./local/export2lite.sh ${train_output_path} inference pdlite pwgan_vctk x86
     # ./local/export2lite.sh ${train_output_path} inference pdlite hifigan_vctk x86
 fi
+
+if [ ${stage} -le 8 ] && [ ${stop_stage} -ge 8 ]; then
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/lite_predict.sh ${train_output_path} || exit -1
+fi
\ No newline at end of file

examples/vctk/voc1/README.md

@@ -141,6 +141,9 @@ The static model can be downloaded here:
 The ONNX model can be downloaded here:
 - [pwgan_vctk_onnx_1.1.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/pwgan/pwgan_vctk_onnx_1.1.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+- [pwgan_vctk_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/pwgan/pwgan_vctk_pdlite_1.3.0.zip)
+
 
 Parallel WaveGAN checkpoint contains files listed below.
 

examples/vctk/voc5/README.md

@@ -127,6 +127,9 @@ The static model can be downloaded here:
 The ONNX model can be downloaded here:
 - [hifigan_vctk_onnx_1.1.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/hifigan/hifigan_vctk_onnx_1.1.0.zip)
 
+The Paddle-Lite model can be downloaded here:
+- [hifigan_vctk_pdlite_1.3.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/hifigan/hifigan_vctk_pdlite_1.3.0.zip)
+
 
 Model | Step | eval/generator_loss | eval/mel_loss| eval/feature_matching_loss
 :-------------:| :------------:| :-----: | :-----: | :--------:

paddlespeech/t2s/exps/inference.py

@@ -145,7 +145,7 @@ def main():
     # warmup
     for utt_id, sentence in sentences[:3]:
         with timer() as t:
-            am_output_data = get_am_output(
+            mel = get_am_output(
                 input=sentence,
                 am_predictor=am_predictor,
                 am=args.am,
@@ -154,12 +154,11 @@ def main():
                 merge_sentences=merge_sentences,
                 speaker_dict=args.speaker_dict,
                 spk_id=args.spk_id, )
-            wav = get_voc_output(
-                voc_predictor=voc_predictor, input=am_output_data)
+            wav = get_voc_output(voc_predictor=voc_predictor, input=mel)
         speed = wav.size / t.elapse
         rtf = fs / speed
         print(
-            f"{utt_id}, mel: {am_output_data.shape}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
+            f"{utt_id}, mel: {mel.shape}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
         )
 
     print("warm up done!")
@@ -168,7 +167,7 @@ def main():
     T = 0
     for utt_id, sentence in sentences:
         with timer() as t:
-            am_output_data = get_am_output(
+            mel = get_am_output(
                 input=sentence,
                 am_predictor=am_predictor,
                 am=args.am,
@@ -177,8 +176,7 @@ def main():
                 merge_sentences=merge_sentences,
                 speaker_dict=args.speaker_dict,
                 spk_id=args.spk_id, )
-            wav = get_voc_output(
-                voc_predictor=voc_predictor, input=am_output_data)
+            wav = get_voc_output(voc_predictor=voc_predictor, input=mel)
 
         N += wav.size
         T += t.elapse
@@ -187,7 +185,7 @@ def main():
 
         sf.write(output_dir / (utt_id + ".wav"), wav, samplerate=fs)
         print(
-            f"{utt_id}, mel: {am_output_data.shape}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
+            f"{utt_id}, mel: {mel.shape}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
         )
 
         print(f"{utt_id} done!")

paddlespeech/t2s/exps/lite_predict.py

+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+import argparse
+from pathlib import Path
+
+import soundfile as sf
+from timer import timer
+
+from paddlespeech.t2s.exps.syn_utils import get_frontend
+from paddlespeech.t2s.exps.syn_utils import get_lite_am_output
+from paddlespeech.t2s.exps.syn_utils import get_lite_predictor
+from paddlespeech.t2s.exps.syn_utils import get_lite_voc_output
+from paddlespeech.t2s.exps.syn_utils import get_sentences
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description="Paddle Infernce with acoustic model & vocoder.")
+    # acoustic model
+    parser.add_argument(
+        '--am',
+        type=str,
+        default='fastspeech2_csmsc',
+        choices=[
+            'speedyspeech_csmsc',
+            'fastspeech2_csmsc',
+            'fastspeech2_aishell3',
+            'fastspeech2_ljspeech',
+            'fastspeech2_vctk',
+            'fastspeech2_mix',
+        ],
+        help='Choose acoustic model type of tts task.')
+    parser.add_argument(
+        "--phones_dict", type=str, default=None, help="phone vocabulary file.")
+    parser.add_argument(
+        "--tones_dict", type=str, default=None, help="tone vocabulary file.")
+    parser.add_argument(
+        "--speaker_dict", type=str, default=None, help="speaker id map file.")
+    parser.add_argument(
+        '--spk_id',
+        type=int,
+        default=0,
+        help='spk id for multi speaker acoustic model')
+    # voc
+    parser.add_argument(
+        '--voc',
+        type=str,
+        default='pwgan_csmsc',
+        choices=[
+            'pwgan_csmsc',
+            'pwgan_aishell3',
+            'pwgan_ljspeech',
+            'pwgan_vctk',
+            'mb_melgan_csmsc',
+            'hifigan_csmsc',
+            'hifigan_aishell3',
+            'hifigan_ljspeech',
+            'hifigan_vctk',
+        ],
+        help='Choose vocoder type of tts task.')
+    # other
+    parser.add_argument(
+        '--lang',
+        type=str,
+        default='zh',
+        help='Choose model language. zh or en or mix')
+    parser.add_argument(
+        "--text",
+        type=str,
+        help="text to synthesize, a 'utt_id sentence' pair per line")
+    parser.add_argument(
+        "--inference_dir", type=str, help="dir to save inference models")
+    parser.add_argument("--output_dir", type=str, help="output dir")
+
+    args, _ = parser.parse_known_args()
+    return args
+
+
+# only inference for models trained with csmsc now
+def main():
+    args = parse_args()
+
+    # frontend
+    frontend = get_frontend(
+        lang=args.lang,
+        phones_dict=args.phones_dict,
+        tones_dict=args.tones_dict)
+
+    # am_predictor
+    am_predictor = get_lite_predictor(
+        model_dir=args.inference_dir, model_file=args.am + "_x86.nb")
+    # model: {model_name}_{dataset}
+    am_dataset = args.am[args.am.rindex('_') + 1:]
+
+    # voc_predictor
+    voc_predictor = get_lite_predictor(
+        model_dir=args.inference_dir, model_file=args.voc + "_x86.nb")
+
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    sentences = get_sentences(text_file=args.text, lang=args.lang)
+
+    merge_sentences = True
+    fs = 24000 if am_dataset != 'ljspeech' else 22050
+    # warmup
+    for utt_id, sentence in sentences[:3]:
+        with timer() as t:
+            mel = get_lite_am_output(
+                input=sentence,
+                am_predictor=am_predictor,
+                am=args.am,
+                frontend=frontend,
+                lang=args.lang,
+                merge_sentences=merge_sentences,
+                speaker_dict=args.speaker_dict,
+                spk_id=args.spk_id, )
+            wav = get_lite_voc_output(voc_predictor=voc_predictor, input=mel)
+        speed = wav.size / t.elapse
+        rtf = fs / speed
+        print(
+            f"{utt_id}, mel: {mel.shape}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
+        )
+
+    print("warm up done!")
+
+    N = 0
+    T = 0
+    for utt_id, sentence in sentences:
+        with timer() as t:
+            mel = get_lite_am_output(
+                input=sentence,
+                am_predictor=am_predictor,
+                am=args.am,
+                frontend=frontend,
+                lang=args.lang,
+                merge_sentences=merge_sentences,
+                speaker_dict=args.speaker_dict,
+                spk_id=args.spk_id, )
+            wav = get_lite_voc_output(voc_predictor=voc_predictor, input=mel)
+
+        N += wav.size
+        T += t.elapse
+        speed = wav.size / t.elapse
+        rtf = fs / speed
+
+        sf.write(output_dir / (utt_id + ".wav"), wav, samplerate=fs)
+        print(
+            f"{utt_id}, mel: {mel.shape}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
+        )
+
+        print(f"{utt_id} done!")
+    print(f"generation speed: {N / T}Hz, RTF: {fs / (N / T) }")
+
+
+if __name__ == "__main__":
+    main()

paddlespeech/t2s/exps/lite_predict_streaming.py

+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+import argparse
+from pathlib import Path
+
+import numpy as np
+import soundfile as sf
+from timer import timer
+
+from paddlespeech.t2s.exps.syn_utils import denorm
+from paddlespeech.t2s.exps.syn_utils import get_chunks
+from paddlespeech.t2s.exps.syn_utils import get_frontend
+from paddlespeech.t2s.exps.syn_utils import get_lite_am_sublayer_output
+from paddlespeech.t2s.exps.syn_utils import get_lite_predictor
+from paddlespeech.t2s.exps.syn_utils import get_lite_streaming_am_output
+from paddlespeech.t2s.exps.syn_utils import get_lite_voc_output
+from paddlespeech.t2s.exps.syn_utils import get_sentences
+from paddlespeech.t2s.exps.syn_utils import run_frontend
+from paddlespeech.t2s.utils import str2bool
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description="Paddle Infernce with acoustic model & vocoder.")
+    # acoustic model
+    parser.add_argument(
+        '--am',
+        type=str,
+        default='fastspeech2_csmsc',
+        choices=['fastspeech2_csmsc'],
+        help='Choose acoustic model type of tts task.')
+    parser.add_argument(
+        "--am_stat",
+        type=str,
+        default=None,
+        help="mean and standard deviation used to normalize spectrogram when training acoustic model."
+    )
+    parser.add_argument(
+        "--phones_dict", type=str, default=None, help="phone vocabulary file.")
+    parser.add_argument(
+        "--tones_dict", type=str, default=None, help="tone vocabulary file.")
+    parser.add_argument(
+        "--speaker_dict", type=str, default=None, help="speaker id map file.")
+    parser.add_argument(
+        '--spk_id',
+        type=int,
+        default=0,
+        help='spk id for multi speaker acoustic model')
+    # voc
+    parser.add_argument(
+        '--voc',
+        type=str,
+        default='pwgan_csmsc',
+        choices=['pwgan_csmsc', 'mb_melgan_csmsc', 'hifigan_csmsc'],
+        help='Choose vocoder type of tts task.')
+    # other
+    parser.add_argument(
+        '--lang',
+        type=str,
+        default='zh',
+        help='Choose model language. zh or en')
+    parser.add_argument(
+        "--text",
+        type=str,
+        help="text to synthesize, a 'utt_id sentence' pair per line")
+    parser.add_argument(
+        "--inference_dir", type=str, help="dir to save inference models")
+    parser.add_argument("--output_dir", type=str, help="output dir")
+    # inference
+
+    # streaming related
+    parser.add_argument(
+        "--am_streaming",
+        type=str2bool,
+        default=False,
+        help="whether use streaming acoustic model")
+    parser.add_argument(
+        "--block_size", type=int, default=42, help="block size of am streaming")
+    parser.add_argument(
+        "--pad_size", type=int, default=12, help="pad size of am streaming")
+
+    args, _ = parser.parse_known_args()
+    return args
+
+
+# only inference for models trained with csmsc now
+def main():
+    args = parse_args()
+
+    # frontend
+    frontend = get_frontend(
+        lang=args.lang,
+        phones_dict=args.phones_dict,
+        tones_dict=args.tones_dict)
+
+    # am_predictor
+    am_encoder_infer_predictor = get_lite_predictor(
+        model_dir=args.inference_dir,
+        model_file=args.am + "_am_encoder_infer" + "_x86.nb")
+    am_decoder_predictor = get_lite_predictor(
+        model_dir=args.inference_dir,
+        model_file=args.am + "_am_decoder" + "_x86.nb")
+    am_postnet_predictor = get_lite_predictor(
+        model_dir=args.inference_dir,
+        model_file=args.am + "_am_postnet" + "_x86.nb")
+    am_mu, am_std = np.load(args.am_stat)
+    # model: {model_name}_{dataset}
+    am_dataset = args.am[args.am.rindex('_') + 1:]
+
+    # voc_predictor
+    voc_predictor = get_lite_predictor(
+        model_dir=args.inference_dir, model_file=args.voc + "_x86.nb")
+
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    sentences = get_sentences(text_file=args.text, lang=args.lang)
+
+    merge_sentences = True
+
+    fs = 24000 if am_dataset != 'ljspeech' else 22050
+    # warmup
+    for utt_id, sentence in sentences[:3]:
+        with timer() as t:
+            normalized_mel = get_lite_streaming_am_output(
+                input=sentence,
+                am_encoder_infer_predictor=am_encoder_infer_predictor,
+                am_decoder_predictor=am_decoder_predictor,
+                am_postnet_predictor=am_postnet_predictor,
+                frontend=frontend,
+                lang=args.lang,
+                merge_sentences=merge_sentences, )
+            mel = denorm(normalized_mel, am_mu, am_std)
+            wav = get_lite_voc_output(voc_predictor=voc_predictor, input=mel)
+        speed = wav.size / t.elapse
+        rtf = fs / speed
+        print(
+            f"{utt_id}, mel: {mel.shape}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
+        )
+
+    print("warm up done!")
+
+    N = 0
+    T = 0
+    block_size = args.block_size
+    pad_size = args.pad_size
+    get_tone_ids = False
+    for utt_id, sentence in sentences:
+        with timer() as t:
+            # frontend
+            frontend_dict = run_frontend(
+                frontend=frontend,
+                text=sentence,
+                merge_sentences=merge_sentences,
+                get_tone_ids=get_tone_ids,
+                lang=args.lang)
+            phone_ids = frontend_dict['phone_ids']
+            phones = phone_ids[0].numpy()
+            # acoustic model
+            orig_hs = get_lite_am_sublayer_output(
+                am_encoder_infer_predictor, input=phones)
+
+            if args.am_streaming:
+                hss = get_chunks(orig_hs, block_size, pad_size)
+                chunk_num = len(hss)
+                mel_list = []
+                for i, hs in enumerate(hss):
+                    am_decoder_output = get_lite_am_sublayer_output(
+                        am_decoder_predictor, input=hs)
+                    am_postnet_output = get_lite_am_sublayer_output(
+                        am_postnet_predictor,
+                        input=np.transpose(am_decoder_output, (0, 2, 1)))
+                    am_output_data = am_decoder_output + np.transpose(
+                        am_postnet_output, (0, 2, 1))
+                    normalized_mel = am_output_data[0]
+
+                    sub_mel = denorm(normalized_mel, am_mu, am_std)
+                    # clip output part of pad
+                    if i == 0:
+                        sub_mel = sub_mel[:-pad_size]
+                    elif i == chunk_num - 1:
+                        # 最后一块的右侧一定没有 pad 够
+                        sub_mel = sub_mel[pad_size:]
+                    else:
+                        # 倒数几块的右侧也可能没有 pad 够
+                        sub_mel = sub_mel[pad_size:(block_size + pad_size) -
+                                          sub_mel.shape[0]]
+                    mel_list.append(sub_mel)
+                mel = np.concatenate(mel_list, axis=0)
+
+            else:
+                am_decoder_output = get_lite_am_sublayer_output(
+                    am_decoder_predictor, input=orig_hs)
+                am_postnet_output = get_lite_am_sublayer_output(
+                    am_postnet_predictor,
+                    input=np.transpose(am_decoder_output, (0, 2, 1)))
+                am_output_data = am_decoder_output + np.transpose(
+                    am_postnet_output, (0, 2, 1))
+                normalized_mel = am_output_data[0]
+                mel = denorm(normalized_mel, am_mu, am_std)
+            # vocoder
+            wav = get_lite_voc_output(voc_predictor=voc_predictor, input=mel)
+
+        N += wav.size
+        T += t.elapse
+        speed = wav.size / t.elapse
+        rtf = fs / speed
+
+        sf.write(output_dir / (utt_id + ".wav"), wav, samplerate=24000)
+        print(
+            f"{utt_id}, mel: {mel.shape}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
+        )
+
+        print(f"{utt_id} done!")
+    print(f"generation speed: {N / T}Hz, RTF: {fs / (N / T) }")
+
+
+if __name__ == "__main__":
+    main()

paddlespeech/t2s/exps/syn_utils.py

@@ -26,6 +26,8 @@ import paddle
 from paddle import inference
 from paddle import jit
 from paddle.static import InputSpec
+from paddlelite.lite import create_paddle_predictor
+from paddlelite.lite import MobileConfig
 from yacs.config import CfgNode
 
 from paddlespeech.t2s.datasets.data_table import DataTable
@@ -510,3 +512,105 @@ def get_sess(model_path: Optional[os.PathLike],
     sess = ort.InferenceSession(
         model_path, providers=providers, sess_options=sess_options)
     return sess
+
+
+# Paddle-Lite
+def get_lite_predictor(model_dir: Optional[os.PathLike]=None,
+                       model_file: Optional[os.PathLike]=None,
+                       cpu_threads: int=1):
+    config = MobileConfig()
+    config.set_model_from_file(str(Path(model_dir) / model_file))
+    predictor = create_paddle_predictor(config)
+    return predictor
+
+
+def get_lite_am_output(
+        input: str,
+        am_predictor,
+        am: str,
+        frontend: object,
+        lang: str='zh',
+        merge_sentences: bool=True,
+        speaker_dict: Optional[os.PathLike]=None,
+        spk_id: int=0, ):
+    am_name = am[:am.rindex('_')]
+    am_dataset = am[am.rindex('_') + 1:]
+    get_spk_id = False
+    get_tone_ids = False
+    if am_name == 'speedyspeech':
+        get_tone_ids = True
+    if am_dataset in {"aishell3", "vctk", "mix"} and speaker_dict:
+        get_spk_id = True
+        spk_id = np.array([spk_id])
+
+    frontend_dict = run_frontend(
+        frontend=frontend,
+        text=input,
+        merge_sentences=merge_sentences,
+        get_tone_ids=get_tone_ids,
+        lang=lang)
+
+    if get_tone_ids:
+        tone_ids = frontend_dict['tone_ids']
+        tones = tone_ids[0].numpy()
+        tones_handle = am_predictor.get_input(1)
+        tones_handle.from_numpy(tones)
+
+    if get_spk_id:
+        spk_id_handle = am_predictor.get_input(1)
+        spk_id_handle.from_numpy(spk_id)
+    phone_ids = frontend_dict['phone_ids']
+    phones = phone_ids[0].numpy()
+    phones_handle = am_predictor.get_input(0)
+    phones_handle.from_numpy(phones)
+    am_predictor.run()
+    am_output_handle = am_predictor.get_output(0)
+    am_output_data = am_output_handle.numpy()
+    return am_output_data
+
+
+def get_lite_voc_output(voc_predictor, input):
+    mel_handle = voc_predictor.get_input(0)
+    mel_handle.from_numpy(input)
+    voc_predictor.run()
+    voc_output_handle = voc_predictor.get_output(0)
+    wav = voc_output_handle.numpy()
+    return wav
+
+
+def get_lite_am_sublayer_output(am_sublayer_predictor, input):
+    input_handle = am_sublayer_predictor.get_input(0)
+    input_handle.from_numpy(input)
+
+    am_sublayer_predictor.run()
+    am_sublayer_handle = am_sublayer_predictor.get_output(0)
+    am_sublayer_output = am_sublayer_handle.numpy()
+    return am_sublayer_output
+
+
+def get_lite_streaming_am_output(input: str,
+                                 am_encoder_infer_predictor,
+                                 am_decoder_predictor,
+                                 am_postnet_predictor,
+                                 frontend,
+                                 lang: str='zh',
+                                 merge_sentences: bool=True):
+    get_tone_ids = False
+    frontend_dict = run_frontend(
+        frontend=frontend,
+        text=input,
+        merge_sentences=merge_sentences,
+        get_tone_ids=get_tone_ids,
+        lang=lang)
+    phone_ids = frontend_dict['phone_ids']
+    phones = phone_ids[0].numpy()
+    am_encoder_infer_output = get_lite_am_sublayer_output(
+        am_encoder_infer_predictor, input=phones)
+    am_decoder_output = get_lite_am_sublayer_output(
+        am_decoder_predictor, input=am_encoder_infer_output)
+    am_postnet_output = get_lite_am_sublayer_output(
+        am_postnet_predictor, input=np.transpose(am_decoder_output, (0, 2, 1)))
+    am_output_data = am_decoder_output + np.transpose(am_postnet_output,
+                                                      (0, 2, 1))
+    normalized_mel = am_output_data[0]
+    return normalized_mel

setup.py

@@ -75,6 +75,7 @@ base = [
     "braceexpand",
     "pyyaml",
     "pybind11",
+    "paddlelite",
     "paddleslim==2.3.4",
 ]