KoreanTTS

Koreantts

It is a project that combines the Tacotron2 model and the Vocoder model (Griffinlim, Wavenet, Melgan) to implement Korean TTS.

Based on

• https://github.com/tensorspeech/tensorflowtts

• https://github.com/hccho2/tacotron2-korean-tts

• https://carpedm20.github.io/tacotron/

Dataset

1. Koran Single Speaker Speech

   • Professional female voice actor (12 hours, WAV, 44100KHz, 12853, 3GB)

2. Actor Yoo Inna's voice

   • Increase the volume of KBS Radio Yu Inna (3 hours, WAV, 16000KHz, 3327, 480.6MB)
   • Google Speech to text api
   • Kakao Speech API

3. Pet pet trainer Kang Hyung -wook voice

   • ETRI Korean Appraisal API

Audio data conducted on learning is not shared with copyright issues. Please check each data source.

• KSS: https://www.kaggle.com/bryanpark/korean-le-speaker-speech-dataset

• KBS Radio: http://program.kbs.co.kr/2fm/radio/Uvolum/pc/index.html

Preprocessing

1. Convert the wav file to a numpy file

2. 'Audio', 'MEL', 'LINEAR', 'TEXT', etc.

3. Data/kss/＂voice file name.npz creation

4. Mel-spectrogram, linear-spectrogram correct answer set

Project progress

There are a total of four learning.

1. Tacotron2 + GriffinLim + SingleSpeaker

2. Tacotron2 + Griffinlim + Multispeaker (Deep Voice 2)

3. Tacotron2 + Melgan + Single Speaker

4. Tacotron2 + Melgan + Multispeaker (Transfer Learning)

result

1. Tacotron2 + Griffinlim + Multispeaker (KSS + Yoo Inna) KSS data

   • Alignmnet (50000)

   

2. Tacotron2 + Griffinlim + Multispeaker (KSS + Yoo In -na)

   • Alignment (90000)

   

3. Tacotron2 + Melgan + SingleSpeaker (KSS)

   • Alignment (90000)