This is the official PyTorch implementation of the following paper:

Automatic Prosody Annotation with Pre-Trained Text-Speech Model
Ziqian Dai, Jianwei Yu, Yan Wang, Nuo Chen, Yanyao Bian, Guangzhi Li, Deng Cai, Dong Yu

Abstract: Prosodic boundary plays an important role in text-to-speech synthesis (TTS) in terms of naturalness and readability. However, the acquisition of prosodic boundary labels relies on manual annotation, which is costly and time-consuming. In this paper, we propose to automatically extract prosodic boundary labels from text-audio data via a neural text-speech model with pre-trained audio encoders. This model is pre-trained on text and speech data separately and jointly fine-tuned on TTS data in a triplet format: {speech, text, prosody}. The experimental results on both automatic evaluation and human evaluation demonstrate that: 1) the proposed text-speech prosody annotation framework significantly outperforms text-only baselines; 2) the quality of automatic prosodic boundary annotations is comparable to human annotations; 3) TTS systems trained with model-annotated boundaries are slightly better than systems that use manual ones.

Visit our demo page for audio samples.

This implementation supports our prosody estimation model (Conformer-Char as the audio encoder) and the code for inference. Note that the training code was not released due to the company's confidentiality policy, please refer to our paper for technical details.

For data preparation in STEP I, you need to install kaldi for feature extraction. For inference in STEP II, a couple of packages are required with Python 3.6. In stall them use

pip install -r requirements.txt

The audio data needs to be preprocessed before fed into the model. Firstly, install kaldi tookit. Then extract audio features with kaldi using the following command. 80 dimension FBank and 3 dimension pitch are concatenated as the input feature.

echo "input raw_audio.wav" > tmp.scp
compute-fbank-feats --num-mel-bins=80 scp:tmp.scp ark:fbk.ark
compute-kaldi-pitch-feats scp:tmp.scp ark:- | process-kaldi-pitch-feats ark:- ark:pitch.ark
paste-feats --length-tolerance=3 ark:fbk.ark ark:pitch.ark ark,scp:feature.ark,feature.scp

This command saves the extracted features in feature.ark, which can be read through feature.scp.

Download the project, put the features files mentioned above in folder “data”， then run the inference code by

python code/main.py

The result will be stored in "prediction_save/test.txt". Note that the label 0-5 corresponds to CC, LW, PW, PPH, IPH in the paper respectively.