How to synthesize long sentences?

keithito / tacotron

A TensorFlow implementation of Google's Tacotron speech synthesis with pre-trained model (unofficial)

MIT License

2.96k stars 957 forks source link

How to synthesize long sentences? #220

Open hadiasheri opened 6 years ago

hadiasheri commented 6 years ago

I've trained tacotron with Nancy corpus for about 1M iterations. The synthesized speech is great for sentences with maximum length of about 25 words. However, when you feed a long sentence ( e.g length~50), it fails to produce acceptable speech because some parts of the sentences are missing or pronounced awfully. Its seems that there might be some alignment failures. I'd be grateful if you could share your ideas about this problem.

maozhiqiang commented 6 years ago

@hadiasheri you can using batch to synthesis long sentence

erogol commented 5 years ago

You can try https://github.com/mozilla/TTS , if you burden the framework change but on Nancy TTS works very good even for very long sentences.

ArashHosseini commented 5 years ago

@hadiasheri Is it possible to provide your 1M model? thx