This example contains code used to train a [Tacotron2 ](https://arxiv.org/abs/1712.05884) model with [AISHELL-3](http://www.aishelltech.com/aishell_3). The trained model can be used in Voice Cloning Task, We refer to the model structure of [Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis](https://arxiv.org/pdf/1806.04558.pdf). The general steps are as follows:
1. Speaker Encoder: We use Speaker Verification to train a speaker encoder. Datasets used in this task are different from those used in Tacotron2 because the transcriptions are not needed, we use more datasets, refer to [ge2e](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/ge2e).
2. Synthesizer: We use the trained speaker encoder to generate speaker embedding for each sentence in AISHELL-3. This embedding is an extra input of Tacotron2 which will be concated with encoder outputs.
3. Vocoder: We use WaveFlow as the neural Vocoder, refer to [waveflow](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/ljspeech/voc0).
You can choose a range of stages you want to run, or set `stage` equal to `stop-stage` to use only one stage, for example, run the following command will only preprocess the dataset.
Use pretrained GE2E (speaker encoder) to generate speaker embedding for each sentence in AISHELL-3, which has the same file structure with wav files and the format is `.npy`.
There is silence in the edge of AISHELL-3's wavs, and the audio amplitude is very small, so, we need to remove the silence and normalize the audio. You can the silence remove method based on volume or energy, but the effect is not very good, We use [MFA](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get the alignment of text and speech, then utilize the alignment results to remove the silence.
We use Montreal Force Aligner 1.0. The label in aishell3 includes pinyin，so the lexicon we provided to MFA is pinyin rather than Chinese characters. And the prosody marks(`$` and `%`) need to be removed. You should preprocess the dataset into the format which MFA needs, the texts have the same name with wavs and have the suffix `.lab`.
We use [lexicon.txt](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/paddlespeech/t2s/exps/voice_cloning/tacotron2_ge2e/lexicon.txt) as the lexicon.
You can download the alignment results from here [alignment_aishell3.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/AISHELL-3/alignment_aishell3.tar.gz), or train your MFA model reference to [mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) (use MFA1.x now) of our repo.
We revert the transcription into `phones` and `tones`. It is worth noting that our processing here is different from that used for MFA, we separated the tones. This is a processing method, of course, you can only segment initials and vowels.
```bash
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
python3 ${BIN_DIR}/preprocess_transcription.py \
--input=${input} \
--output=${preprocess_path}
fi
```
The default input is `~/datasets/data_aishell3/train`，which contains `label_train-set.txt`, the processed results are `metadata.yaml` and `metadata.pickle`. the former is a text format for easy viewing, and the latter is a binary format for direct reading.
Our model removes stop token prediction in Tacotron2, because of the problem of the extremely unbalanced proportion of positive and negative samples of stop token prediction, and it's very sensitive to the clip of audio silence. We use the last symbol from the highest point of attention to the encoder side as the termination condition.
In addition, to accelerate the convergence of the model, we add `guided attention loss` to induce the alignment between encoder and decoder to show diagonal lines faster.
