parent
149b1fb1fa
commit
dafe7c3657
@ -0,0 +1,47 @@
|
||||
#!/bin/bash
|
||||
|
||||
train_output_path=$1
|
||||
|
||||
stage=2
|
||||
stop_stage=2
|
||||
|
||||
# pwgan
|
||||
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
|
||||
python3 ${BIN_DIR}/../inference_streaming.py \
|
||||
--inference_dir=${train_output_path}/inference_streaming \
|
||||
--am=fastspeech2_csmsc \
|
||||
--am_stat=dump/train/speech_stats.npy \
|
||||
--voc=pwgan_csmsc \
|
||||
--text=${BIN_DIR}/../sentences.txt \
|
||||
--output_dir=${train_output_path}/pd_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 ${BIN_DIR}/../inference_streaming.py \
|
||||
--inference_dir=${train_output_path}/inference_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}/pd_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 ${BIN_DIR}/../inference_streaming.py \
|
||||
--inference_dir=${train_output_path}/inference_streaming \
|
||||
--am=fastspeech2_csmsc \
|
||||
--am_stat=dump/train/speech_stats.npy \
|
||||
--voc=hifigan_csmsc \
|
||||
--text=${BIN_DIR}/../sentences.txt \
|
||||
--output_dir=${train_output_path}/pd_infer_out_streaming \
|
||||
--phones_dict=dump/phone_id_map.txt \
|
||||
--am_streaming=True
|
||||
fi
|
||||
|
@ -0,0 +1,19 @@
|
||||
train_output_path=$1
|
||||
|
||||
stage=0
|
||||
stop_stage=0
|
||||
|
||||
# e2e, synthesize from text
|
||||
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
|
||||
python3 ${BIN_DIR}/../ort_predict_streaming.py \
|
||||
--inference_dir=${train_output_path}/inference_onnx_streaming \
|
||||
--am=fastspeech2_csmsc \
|
||||
--am_stat=dump/train/speech_stats.npy \
|
||||
--voc=hifigan_csmsc \
|
||||
--output_dir=${train_output_path}/onnx_infer_out_streaming \
|
||||
--text=${BIN_DIR}/../csmsc_test.txt \
|
||||
--phones_dict=dump/phone_id_map.txt \
|
||||
--device=cpu \
|
||||
--cpu_threads=2 \
|
||||
--am_streaming=True
|
||||
fi
|
@ -0,0 +1,224 @@
|
||||
# Copyright (c) 2021 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_am_sublayer_output
|
||||
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_predictor
|
||||
from paddlespeech.t2s.exps.syn_utils import get_sentences
|
||||
from paddlespeech.t2s.exps.syn_utils import get_streaming_am_output
|
||||
from paddlespeech.t2s.exps.syn_utils import get_streaming_am_predictor
|
||||
from paddlespeech.t2s.exps.syn_utils import get_voc_output
|
||||
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
|
||||
parser.add_argument(
|
||||
"--device",
|
||||
default="gpu",
|
||||
choices=["gpu", "cpu"],
|
||||
help="Device selected for inference.", )
|
||||
# streaming related
|
||||
parser.add_argument(
|
||||
"--am_streaming",
|
||||
type=str2bool,
|
||||
default=False,
|
||||
help="whether use streaming acoustic model")
|
||||
parser.add_argument(
|
||||
"--chunk_size", type=int, default=42, help="chunk 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(args)
|
||||
|
||||
# am_predictor
|
||||
am_encoder_infer_predictor, am_decoder_predictor, am_postnet_predictor = get_streaming_am_predictor(
|
||||
args)
|
||||
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_predictor(args, filed='voc')
|
||||
|
||||
output_dir = Path(args.output_dir)
|
||||
output_dir.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
sentences = get_sentences(args)
|
||||
|
||||
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_streaming_am_output(
|
||||
args,
|
||||
am_encoder_infer_predictor=am_encoder_infer_predictor,
|
||||
am_decoder_predictor=am_decoder_predictor,
|
||||
am_postnet_predictor=am_postnet_predictor,
|
||||
frontend=frontend,
|
||||
merge_sentences=merge_sentences,
|
||||
input=sentence)
|
||||
mel = denorm(normalized_mel, am_mu, am_std)
|
||||
wav = get_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
|
||||
chunk_size = args.chunk_size
|
||||
pad_size = args.pad_size
|
||||
get_tone_ids = False
|
||||
for utt_id, sentence in sentences:
|
||||
with timer() as t:
|
||||
# frontend
|
||||
if args.lang == 'zh':
|
||||
input_ids = frontend.get_input_ids(
|
||||
sentence,
|
||||
merge_sentences=merge_sentences,
|
||||
get_tone_ids=get_tone_ids)
|
||||
phone_ids = input_ids["phone_ids"]
|
||||
else:
|
||||
print("lang should be 'zh' here!")
|
||||
phones = phone_ids[0].numpy()
|
||||
# acoustic model
|
||||
orig_hs = get_am_sublayer_output(
|
||||
am_encoder_infer_predictor, input=phones)
|
||||
|
||||
if args.am_streaming:
|
||||
hss = get_chunks(orig_hs, chunk_size, pad_size)
|
||||
chunk_num = len(hss)
|
||||
mel_list = []
|
||||
for i, hs in enumerate(hss):
|
||||
am_decoder_output = get_am_sublayer_output(
|
||||
am_decoder_predictor, input=hs)
|
||||
am_postnet_output = get_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:(chunk_size + pad_size) -
|
||||
sub_mel.shape[0]]
|
||||
mel_list.append(sub_mel)
|
||||
mel = np.concatenate(mel_list, axis=0)
|
||||
|
||||
else:
|
||||
am_decoder_output = get_am_sublayer_output(
|
||||
am_decoder_predictor, input=orig_hs)
|
||||
|
||||
am_postnet_output = get_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_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()
|
@ -0,0 +1,233 @@
|
||||
# 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_sentences
|
||||
from paddlespeech.t2s.exps.syn_utils import get_sess
|
||||
from paddlespeech.t2s.exps.syn_utils import get_streaming_am_sess
|
||||
from paddlespeech.t2s.utils import str2bool
|
||||
|
||||
|
||||
def ort_predict(args):
|
||||
|
||||
# frontend
|
||||
frontend = get_frontend(args)
|
||||
|
||||
output_dir = Path(args.output_dir)
|
||||
output_dir.mkdir(parents=True, exist_ok=True)
|
||||
sentences = get_sentences(args)
|
||||
|
||||
am_name = args.am[:args.am.rindex('_')]
|
||||
am_dataset = args.am[args.am.rindex('_') + 1:]
|
||||
fs = 24000 if am_dataset != 'ljspeech' else 22050
|
||||
|
||||
# am
|
||||
am_encoder_infer_sess, am_decoder_sess, am_postnet_sess = get_streaming_am_sess(
|
||||
args)
|
||||
am_mu, am_std = np.load(args.am_stat)
|
||||
|
||||
# vocoder
|
||||
voc_sess = get_sess(args, filed='voc')
|
||||
|
||||
# frontend warmup
|
||||
# Loading model cost 0.5+ seconds
|
||||
if args.lang == 'zh':
|
||||
frontend.get_input_ids("你好,欢迎使用飞桨框架进行深度学习研究!", merge_sentences=True)
|
||||
else:
|
||||
print("lang should in be 'zh' here!")
|
||||
|
||||
# am warmup
|
||||
for T in [27, 38, 54]:
|
||||
phone_ids = np.random.randint(1, 266, size=(T, ))
|
||||
am_encoder_infer_sess.run(None, input_feed={'text': phone_ids})
|
||||
|
||||
am_decoder_input = np.random.rand(1, T * 15, 384).astype('float32')
|
||||
am_decoder_sess.run(None, input_feed={'xs': am_decoder_input})
|
||||
|
||||
am_postnet_input = np.random.rand(1, 80, T * 15).astype('float32')
|
||||
am_postnet_sess.run(None, input_feed={'xs': am_postnet_input})
|
||||
|
||||
# voc warmup
|
||||
for T in [227, 308, 544]:
|
||||
data = np.random.rand(T, 80).astype("float32")
|
||||
voc_sess.run(None, input_feed={"logmel": data})
|
||||
print("warm up done!")
|
||||
|
||||
N = 0
|
||||
T = 0
|
||||
merge_sentences = True
|
||||
get_tone_ids = False
|
||||
chunk_size = args.chunk_size
|
||||
pad_size = args.pad_size
|
||||
|
||||
for utt_id, sentence in sentences:
|
||||
with timer() as t:
|
||||
if args.lang == 'zh':
|
||||
input_ids = frontend.get_input_ids(
|
||||
sentence,
|
||||
merge_sentences=merge_sentences,
|
||||
get_tone_ids=get_tone_ids)
|
||||
phone_ids = input_ids["phone_ids"]
|
||||
else:
|
||||
print("lang should in be 'zh' here!")
|
||||
# merge_sentences=True here, so we only use the first item of phone_ids
|
||||
phone_ids = phone_ids[0].numpy()
|
||||
orig_hs = am_encoder_infer_sess.run(
|
||||
None, input_feed={'text': phone_ids})
|
||||
if args.am_streaming:
|
||||
hss = get_chunks(orig_hs[0], chunk_size, pad_size)
|
||||
chunk_num = len(hss)
|
||||
mel_list = []
|
||||
for i, hs in enumerate(hss):
|
||||
am_decoder_output = am_decoder_sess.run(
|
||||
None, input_feed={'xs': hs})
|
||||
am_postnet_output = am_postnet_sess.run(
|
||||
None,
|
||||
input_feed={
|
||||
'xs': np.transpose(am_decoder_output[0], (0, 2, 1))
|
||||
})
|
||||
am_output_data = am_decoder_output + np.transpose(
|
||||
am_postnet_output[0], (0, 2, 1))
|
||||
normalized_mel = am_output_data[0][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:(chunk_size + pad_size) -
|
||||
sub_mel.shape[0]]
|
||||
mel_list.append(sub_mel)
|
||||
mel = np.concatenate(mel_list, axis=0)
|
||||
else:
|
||||
am_decoder_output = am_decoder_sess.run(
|
||||
None, input_feed={'xs': orig_hs[0]})
|
||||
am_postnet_output = am_postnet_sess.run(
|
||||
None,
|
||||
input_feed={
|
||||
'xs': np.transpose(am_decoder_output[0], (0, 2, 1))
|
||||
})
|
||||
am_output_data = am_decoder_output + np.transpose(
|
||||
am_postnet_output[0], (0, 2, 1))
|
||||
normalized_mel = am_output_data[0]
|
||||
mel = denorm(normalized_mel, am_mu, am_std)
|
||||
mel = mel[0]
|
||||
# vocoder
|
||||
|
||||
wav = voc_sess.run(output_names=None, input_feed={'logmel': mel})
|
||||
|
||||
N += len(wav[0])
|
||||
T += t.elapse
|
||||
speed = len(wav[0]) / t.elapse
|
||||
rtf = fs / speed
|
||||
sf.write(
|
||||
str(output_dir / (utt_id + ".wav")),
|
||||
np.array(wav)[0],
|
||||
samplerate=fs)
|
||||
print(
|
||||
f"{utt_id}, mel: {mel.shape}, wave: {len(wav[0])}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
|
||||
)
|
||||
print(f"generation speed: {N / T}Hz, RTF: {fs / (N / T) }")
|
||||
|
||||
|
||||
def parse_args():
|
||||
parser = argparse.ArgumentParser(description="Infernce with onnxruntime.")
|
||||
# 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.")
|
||||
|
||||
# voc
|
||||
parser.add_argument(
|
||||
'--voc',
|
||||
type=str,
|
||||
default='hifigan_csmsc',
|
||||
choices=['hifigan_csmsc', 'mb_melgan_csmsc', 'pwgan_csmsc'],
|
||||
help='Choose vocoder type of tts task.')
|
||||
# other
|
||||
parser.add_argument(
|
||||
"--inference_dir", type=str, help="dir to save inference models")
|
||||
parser.add_argument(
|
||||
"--text",
|
||||
type=str,
|
||||
help="text to synthesize, a 'utt_id sentence' pair per line")
|
||||
parser.add_argument("--output_dir", type=str, help="output dir")
|
||||
parser.add_argument(
|
||||
'--lang',
|
||||
type=str,
|
||||
default='zh',
|
||||
help='Choose model language. zh or en')
|
||||
|
||||
# inference
|
||||
parser.add_argument(
|
||||
"--use_trt",
|
||||
type=str2bool,
|
||||
default=False,
|
||||
help="Whether to use inference engin TensorRT.", )
|
||||
|
||||
parser.add_argument(
|
||||
"--device",
|
||||
default="gpu",
|
||||
choices=["gpu", "cpu"],
|
||||
help="Device selected for inference.", )
|
||||
parser.add_argument('--cpu_threads', type=int, default=1)
|
||||
|
||||
# streaming related
|
||||
parser.add_argument(
|
||||
"--am_streaming",
|
||||
type=str2bool,
|
||||
default=False,
|
||||
help="whether use streaming acoustic model")
|
||||
parser.add_argument(
|
||||
"--chunk_size", type=int, default=42, help="chunk 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
|
||||
|
||||
|
||||
def main():
|
||||
args = parse_args()
|
||||
|
||||
ort_predict(args)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
Loading…
Reference in new issue