zerovox
1.0.0
Zerovox是用於實時和嵌入式使用的文本對語音（TTS）系統。
Zerovox完全離線運行，確保隱私和脫離雲服務的獨立性。它是完全免費和開源的，邀請社區貢獻和建議。
Zerovox以FastSpeech2的形式建模，更進一步，使用零拍揚聲器克隆，利用全球樣式令牌（GST）和揚聲器條件層歸一化（SCLN）進行有效的揚聲器嵌入。該系統從單個模型中支持英語和德語語音生成，並在廣泛的數據集中訓練。 Zerovox是基於音素的，利用發音詞典來確保精確的單詞發音，利用英語的CMU字典，以及Zamiaspeech項目中的德語自定義字典，其中使用的音素集也來自其中。
Zerovox可以用作LLM的TTS後端，實現實時互動，並作為家庭自動化系統（例如家庭助理）易於安裝的TTS系統。由於它不像FastSpeech2那樣是無助的，其輸出通常易於控制和可預測。
許可證：Zerovox是Apache 2，並根據MIT許可證獲得了其他項目（請參閱下面的信用部分）的許多零件。
請注意：模型仍處於Alpha階段，仍在訓練。
https://huggingface.co/spaces/gooooofy/zerovox-demo
當前的Zerovox培訓語料庫統計：
german audio corpus: 16679 speakers, 475.3 hours audio
english audio corpus: 19899 speakers, 358.7 hours audio
（1/5）準備語料庫：
pushd configs/corpora/cv_de_100
./gen_cv.sh
popd
（2/5）準備對齊：
utils/prepare_align.py configs/corpora/cv_de_100
（3/5）OOV：
utils/oovtool.py -a -m zerovox-g2p-autoreg-zamia-de configs/corpora/cv_de_100
（4/5）對齊：
utils/align.py --kaldi-model=tts_de_kaldi_zamia_4 configs/corpora/cv_de_100
（5/5）預處理：
utils/preprocess.py configs/corpora/cv_de_100
utils/train_tts.py
--head=2 --reduction=1 --expansion=2 --kernel-size=5 --n-blocks=3 --block-depth=3
--accelerator=gpu --threads=24 --batch-size=32 --val_epochs=8
--infer-device=cpu
--lr=0.0001 --warmup_epochs=25
--hifigan-checkpoint=VCTK_V2
--out-folder=models/tts_de_zerovox_base_1
configs/corpora/cv_de_100
configs/corpora/de_hui/de_hui_*.yaml
configs/corpora/de_thorsten.yaml
utils/train_kaldi.py --model-name=tts_de_kaldi_zamia_4 --num-jobs=12 configs/corpora/cv_de_100
運行訓練：
scripts/train_g2p_de_autoreg.sh
最初是基於Rowel Atienza的效力
https://github.com/roatienza/efficientspeech
@inproceedings{atienza2023efficientspeech,
title={EfficientSpeech: An On-Device Text to Speech Model},
author={Atienza, Rowel},
booktitle={ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
pages={1--5},
year={2023},
organization={IEEE}
}
FastSpeech2編碼器和解碼器是從Chung-Ming Chien實施FastSpeech2借（根據MIT許可證）
https://github.com/ming024/fastspeech2
@misc{ren2022fastspeech2fasthighquality,
title={FastSpeech 2: Fast and High-Quality End-to-End Text to Speech},
author={Yi Ren and Chenxu Hu and Xu Tan and Tao Qin and Sheng Zhao and Zhou Zhao and Tie-Yan Liu},
year={2022},
eprint={2006.04558},
archivePrefix={arXiv},
primaryClass={eess.AS},
url={https://arxiv.org/abs/2006.04558},
}
MEL解碼器的實施是（根據MIT許可證）從Tomoki Hayashi的Parallelwavegan項目中藉出來的：
https://github.com/kan-bayashi/parallelwavegan G2P變壓器模型基於Axel Springer News News Media＆Tech Gmbh＆Co。 KG基於Deepphonemizer -deepphonemizer-創意工程（MIT許可證）
https://github.com/as-ideas/deepphonemizer
@inproceedings{Yolchuyeva_2019, series={interspeech_2019},
title={Transformer Based Grapheme-to-Phoneme Conversion},
url={http://dx.doi.org/10.21437/Interspeech.2019-1954},
DOI={10.21437/interspeech.2019-1954},
booktitle={Interspeech 2019},
publisher={ISCA},
author={Yolchuyeva, Sevinj and Németh, Géza and Gyires-Tóth, Bálint},
year={2019},
month=sep, pages={2095–2099},
collection={interspeech_2019} }
Clova AI Research從Voxceleb_trainer那裡借了基於Zeroshot Resnet的揚聲器編碼（根據MIT許可）
https://github.com/clovaai/voxceleb_trainer
@inproceedings{chung2020in,
title={In defence of metric learning for speaker recognition},
author={Chung, Joon Son and Huh, Jaesung and Mun, Seongkyu and Lee, Minjae and Heo, Hee Soo and Choe, Soyeon and Ham, Chiheon and Jung, Sunghwan and Lee, Bong-Jin and Han, Icksang},
booktitle={Proc. Interspeech},
year={2020}
}
@inproceedings{he2016deep,
title={Deep residual learning for image recognition},
author={He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian},
booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
pages={770--778},
year={2016}
}
基於Zeroshot全球標記的揚聲器嵌入基於Chengqi Deng（MIT許可證）的GST-TACOTRON
https://github.com/kinglittleq/gst-tacotron
這是實施
@misc{wang2018style,
title={Style Tokens: Unsupervised Style Modeling, Control and Transfer in End-to-End Speech Synthesis},
author={Yuxuan Wang and Daisy Stanton and Yu Zhang and RJ Skerry-Ryan and Eric Battenberg and Joel Shor and Ying Xiao and Fei Ren and Ye Jia and Rif A. Saurous},
year={2018},
eprint={1803.09017},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
揚聲器條件層歸一化（SCLN），該層（根據MIT許可）從
https://github.com/keonlee9420/cross-speaker-emotion-transfer by keon lee
@misc{wu2021crossspeakeremotiontransferbased,
title={Cross-speaker Emotion Transfer Based on Speaker Condition Layer Normalization and Semi-Supervised Training in Text-To-Speech},
author={Pengfei Wu and Junjie Pan and Chenchang Xu and Junhui Zhang and Lin Wu and Xiang Yin and Zejun Ma},
year={2021},
eprint={2110.04153},
archivePrefix={arXiv},
primaryClass={eess.AS},
url={https://arxiv.org/abs/2110.04153},
}
