Thorsten-Voice/README.md
2020-04-18 19:34:21 +02:00

4.6 KiB

Introduction

Many smart assistants like Amazon Alexa, Google Home, Apple Siri and Microsoft Cortana need an internet connection to offer the functions STT (speech in text) and TTS (text in speech) in decent quality. But there are also open source projects that develop alternative wizards, some of which work offline. Personally i'm playing currently around with MyCroft AI which has a great community.

For the area "STT / TTS", however, good training test data (eg for deep learning) are required. This is where the Mozilla Common Voice project comes into play.

And?!

I want to make my small modest contribution and contribute the model (tacotron version 1 and 2) of my personal voice (german) to the community for free to use.

Samples of my original voice

To get an impression of what my voice sounds to decide if it fits to your project i published some sample recordings. As soon the final model is available i'll upload some generated TTS samples too.

Dataset information

I recorded my voice using mimic-recording-studio by mycroft and a german corpus provided by @gras64. After recording and removing bad recordings, @domcross optimized the dataset concerning random noise, echo and background beeps. Finally the dataset used for training is:

  • 20.711 recorded phrases (wav files)
  • more than 20 hours of pure audio
  • samplerate 22.050Hz
  • mono
  • avg sentence length: 47 chars
  • avg spoken chars per second: 14
  • sentences with question mark: 2.759
  • sentences with exclamation mark: 1.869
  • ljspeech-1.1 structure

TODO: Add bokeh graph. Changing recording location and equipment duing the recording session leads to two clusters in recording bokeh.

Training tacotron version 1 (for mimic2 usage)

Graphs

analyze output

char_len_vs_avg_secs char_len_vs_med_secs char_len_vs_mode_secs char_len_vs_num_samples char_len_vs_std

alignment

Training in progress - step 81k

Tensorboard

Training in progress - step 81k

TTS samples

All samples are "training in progress"

Model release (tacotron 1)

Will be released as soon training process is finished

Mozilla TTS training/model (tacotron 2)

Training info

Model release (tacotron2)

Will be released as soon training process is finished

Special thanks

I want to thank the whole community for providing great open source products.

Special thanks go to

for their support or this.

A very special thank you for @domcross for his time, support, audio optimzing knowhow and finally his gpu computing power. Thank you :-)

Finally

If you use my (concrete) TTS voice I would be grateful for an info about the project and a demo.

Please do not use my voice for evil.

TODO

  • add training plots
  • add bokeh graph
  • google drive download urls for model download

Links