【PaddleSpeech进阶】语音合成-onnx模型使用

发布于:2023-01-11 ⋅ 阅读:(391) ⋅ 点赞:(0)

很多同学对PaddleSpeech发布的语音合成onnx模型的使用比较感兴趣,这篇教程将教会你如何使用PaddleSpeech提供的语音合成预训练模型完成推理工作。

0. PaddleSpeech 介绍

🚀 PaddleSpeech 是 all-in-one 的语音算法工具箱,包含多种领先国际水平的语音算法与预训练模型。你可以从中选择各种语音处理工具以及预训练模型,支持语音识别,语音合成,声音分类,声纹识别,标点恢复,语音翻译等多种功能,PaddleSpeech Server模块可帮助用户快速在服务器上部署语音服务。PaddleSpeech团队发表的论文 An Easy-to-Use All-in-One Speech Toolkit 入选 NAACL2022 ,荣获 NAACL2022 Best Demo Award

喜欢的同学可以点个 ⭐️star⭐️ 支持我们,PaddleSpeech传送门:https://github.com/PaddlePaddle/PaddleSpeech

🎁 学习过程中你遇到任何问题,可以加入PaddleSpeech的交流群进行讨论。

 

1. 初步认识onnxruntime推理流程

使用onnxruntime推理PaddleSpeech提供的语音合成onnx模型只需要四个步骤:

  1. 文本前端
  2. 加载模型,创建Session
  3. 模型推理
  4. 音频保存

2. 配置PaddleSpeech开发环境

你可以通过PaddleSpeech的源码进行安装

In [ ]

# 安装PaddleSpeech
!git clone https://gitee.com/paddlepaddle/PaddleSpeech.git
%cd PaddleSpeech
!pip install pytest-runner
!pip install .

In [ ]

# aistudio会报错: paddlespeech 的 repo中存在失效软链接
# 执行下面这行命令!!
!find -L /home/aistudio -type l -delete

In [ ]

# 下载模型模型
%cd /home/aistudio/work
!wget https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_cnndecoder_csmsc_streaming_onnx_1.0.0.zip
!wget https://paddlespeech.bj.bcebos.com/Parakeet/released_models/mb_melgan/mb_melgan_csmsc_onnx_0.2.0.zip
!unzip fastspeech2_cnndecoder_csmsc_streaming_onnx_1.0.0.zip
!unzip mb_melgan_csmsc_onnx_0.2.0.zip

In [ ]


# 下载nltk数据包,如果项目中有就不用下载了
%cd /home/aistudio
!wget -P data https://paddlespeech.bj.bcebos.com/Parakeet/tools/nltk_data.tar.gz
!tar zxvf data/nltk_data.tar.gz

3. TTS文本前端

PaddleSpeech提供的文本前端可以帮助我们把中文文本转换成模型推理需要的音素序列

In [3]

phones_dict = "/home/aistudio/work/fastspeech2_cnndecoder_csmsc_streaming_onnx_1.0.0/phone_id_map.txt"

In [ ]

from paddlespeech.t2s.frontend.zh_frontend import Frontend

frontend = Frontend(
                phone_vocab_path=phones_dict,
                tone_vocab_path=None)

In [ ]

text = "今天天气真的很不错,我想出去玩!"
input_ids = frontend.get_input_ids(
                text,
                merge_sentences=True, # 是否按符号拆分句子
                get_tone_ids=False)
input_ids = input_ids['phone_ids']
print(input_ids)

4. 加载模型,创建Onnxruntime Session

创建onnxruntime的session,用于推理

In [ ]

import onnxruntime as ort

# 模型路径
onnx_am_encoder = "/home/aistudio/work/fastspeech2_cnndecoder_csmsc_streaming_onnx_1.0.0/fastspeech2_csmsc_am_encoder_infer.onnx"
onnx_am_decoder = "/home/aistudio/work/fastspeech2_cnndecoder_csmsc_streaming_onnx_1.0.0/fastspeech2_csmsc_am_decoder.onnx"
onnx_am_postnet = "/home/aistudio/work/fastspeech2_cnndecoder_csmsc_streaming_onnx_1.0.0/fastspeech2_csmsc_am_postnet.onnx"
onnx_voc_melgan = "/home/aistudio/work/mb_melgan_csmsc_onnx_0.2.0/mb_melgan_csmsc.onnx"

# 用CPU推理
providers = ['CPUExecutionProvider']

# 配置ort session
sess_options = ort.SessionOptions()

# 创建session
am_encoder_infer_sess = ort.InferenceSession(onnx_am_encoder, providers=providers, sess_options=sess_options)
am_decoder_sess = ort.InferenceSession(onnx_am_decoder, providers=providers, sess_options=sess_options)
am_postnet_sess = ort.InferenceSession(onnx_am_postnet, providers=providers, sess_options=sess_options)
voc_melgan_sess = ort.InferenceSession(onnx_voc_melgan, providers=providers, sess_options=sess_options)

5. 模型推理

In [7]

# 辅助函数 denorm, 训练过程中mel输出经过了norm,使用过程中需要进行denorm
import numpy as np
am_stat_path = r"/home/aistudio/work/fastspeech2_cnndecoder_csmsc_streaming_onnx_1.0.0/speech_stats.npy"
am_mu, am_std = np.load(am_stat_path)

In [8]

from paddlespeech.server.utils.util import denorm
# 推理阶段封装
# 端到端合成:一次性把句子全部合成完毕
def inference(text):
    phone_ids = frontend.get_input_ids(text, merge_sentences=True, get_tone_ids=False)['phone_ids']
    orig_hs = am_encoder_infer_sess.run(None, input_feed={'text': phone_ids[0].numpy()})
    hs = orig_hs[0]
    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]
    mel = denorm(normalized_mel, am_mu, am_std)
    wav = voc_melgan_sess.run(output_names=None, input_feed={'logmel': mel})[0]
    return wav

6. 音频保存

In [10]

# 保存为wav,播放体验
import soundfile as sf
import time
text = "欢迎使用飞桨语音合成系统,测试一下合成效果。"
t1 = time.time()
wav = inference(text)
print("合成耗时:", time.time() - t1)
sf.write("demo.wav", wav, samplerate=24000)
合成耗时: 6.246196746826172

In [11]

import IPython.display as dp
dp.Audio("demo.wav")
<IPython.lib.display.Audio object>

7. 流式语音合成

流式语音合成需要流式播放才能起到展示效果,思路上是把各个流程进行分片,然后再分块合成,播放器同时流式播放。

流式播放需要声卡支持,建议放到自己的笔记本上进行播放,aistudio 上不便于展示,只展示拼接在一起的结果,不进行流式播放展示

将 streaming_tts.py下载到本机,按上面的步骤下载好模型,安装好PaddleSpeech即可(注意nltk_data,下载速度较慢,建议按上面方式提前下载好)

需要安装 pyaudio

In [12]

# 配置流式参数
import math
from paddlespeech.server.utils.util import get_chunks


voc_block = 36
voc_pad = 14
am_block = 72
am_pad = 12
voc_upsample = 300


def depadding(data, chunk_num, chunk_id, block, pad, upsample):
    """ 
    Streaming inference removes the result of pad inference
    """
    front_pad = min(chunk_id * block, pad)
    # first chunk
    if chunk_id == 0:
        data = data[:block * upsample]
    # last chunk
    elif chunk_id == chunk_num - 1:
        data = data[front_pad * upsample:]
    # middle chunk
    else:
        data = data[front_pad * upsample:(front_pad + block) * upsample]

    return data


def inference_stream(text):
    input_ids = frontend.get_input_ids(
                text,
                merge_sentences=False,
                get_tone_ids=False)
    phone_ids = input_ids["phone_ids"]
    print(phone_ids)
    for i in range(len(phone_ids)):
        # 先分句
        # am 
        voc_chunk_id = 0
        orig_hs = am_encoder_infer_sess.run(
            None, input_feed={'text': phone_ids[i].numpy()})
        orig_hs = orig_hs[0]

        # streaming voc chunk info
        mel_len = orig_hs.shape[1]
        voc_chunk_num = math.ceil(mel_len / voc_block)
        start = 0
        end = min(voc_block + voc_pad, mel_len)

        # streaming am
        hss = get_chunks(orig_hs, am_block, am_pad, "am")
        am_chunk_num = len(hss)
        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)
            sub_mel = depadding(sub_mel, am_chunk_num, i,
                                    am_block, am_pad, 1)

            if i == 0:
                mel_streaming = sub_mel
            else:
                mel_streaming = np.concatenate(
                    (mel_streaming, sub_mel), axis=0)

            # streaming voc
            # 当流式AM推理的mel帧数大于流式voc推理的chunk size,开始进行流式voc 推理
            while (mel_streaming.shape[0] >= end and
                    voc_chunk_id < voc_chunk_num):
                voc_chunk = mel_streaming[start:end, :]

                sub_wav = voc_melgan_sess.run(
                    output_names=None, input_feed={'logmel': voc_chunk})
                sub_wav = depadding(
                    sub_wav[0], voc_chunk_num, voc_chunk_id,
                    voc_block, voc_pad, voc_upsample)

                yield sub_wav

                voc_chunk_id += 1
                start = max(
                    0, voc_chunk_id * voc_block - voc_pad)
                end = min(
                    (voc_chunk_id + 1) * voc_block + voc_pad,
                    mel_len)

In [13]

text = "欢迎使用飞桨语音合成系统,测试一下合成效果。"
wavs = []
t1 = time.time()
for sub_wav in inference_stream(text):
    print("响应时间:", time.time() - t1)
    t1 = time.time()
    wavs.append(sub_wav.flatten())
wav = np.concatenate(wavs)
print(wav.shape)
sf.write("demo_stream.wav",data=wav, samplerate=24000)
[Tensor(shape=[21], dtype=int64, place=Place(cpu), stop_gradient=True,
       [71 , 199, 126, 177, 115, 138, 69 , 46 , 151, 89 , 241, 120, 71 , 42 ,
        39 , 57 , 260, 75 , 182, 163, 179]), Tensor(shape=[16], dtype=int64, place=Place(cpu), stop_gradient=True,
       [38 , 44 , 177, 116, 73 , 260, 80 , 71 , 42 , 39 , 57 , 260, 99 , 70 ,
        232, 179])]
响应时间: 1.4043552875518799
响应时间: 1.4002647399902344
响应时间: 1.2040159702301025
响应时间: 0.8999142646789551
响应时间: 0.8972508907318115
响应时间: 1.100409746170044
响应时间: 0.19948339462280273
响应时间: 1.9018454551696777
响应时间: 1.303241491317749
响应时间: 0.9031462669372559
响应时间: 1.1989772319793701
响应时间: 0.7002658843994141
(110100,)

In [14]

dp.Audio("demo_stream.wav")
<IPython.lib.display.Audio object>

8. 总结

通过上面教程的学习,相信你已经熟练掌握了如何使用onnxruntime来推理PaddleSpeech提供的预训练onnx模型,希望本教程能对你有所帮助。

遇到问题,可以加入 PaddleSpeech 技术交流群,和 PaddleSpeech 的开发者们一起交流讨论。

 

本文含有隐藏内容,请 开通VIP 后查看