This is a REST API server that speech-synthesizes articles (stripping boilerplate, doing sentence segmentation, etc.) into a podcast feed of MP3s which you can consume with your favorite podcasting app. You can feed in articles via bookmarklet or PWA.

The PWA is available at https://github.com/yang/audiolizard-pwa.

Install prerequisites on Ubuntu:

sudo apt-get install \
  postgresql-10 \
  ffmpeg \
  ubuntu-restricted-extras \
  postgresql-server-dev-10 \
  build-essential \
  python-dev \
  postfix \
  openjdk-11-jdk # Needed for boilerpipe
sudo snap install google-cloud-sdk  --classic

Recommended: set up postfix to send via gmail:

https://www.howtoforge.com/tutorial/configure-postfix-to-use-gmail-as-a-mail-relay/

Install prerequisites on OS X:

sudo port install postgresql95-server postgresql95 openjdk10

In this source dir, install the application (e.g. into a virtualenv):

pip install --process-dependency-links -e .

Download nltk data:

python -c 'import nltk; nltk.download()'

Create the necessary postgresql user and DB (and store the password in your pgpass file):

sudo -u postgres createuser -P webreader
sudo -u postgres createdb -O webreader webreader
touch ~/.pgpass
chmod 600 ~/.pgpass

cat >> ~/.pgpass
localhost:5432:webreader:webreader:PASSWORD
^D

psql -h localhost -U webreader webreader
create table articles (id serial primary key, url text, created timestamp not null, title text, body text, converted timestamp);
^D

Follow https://cloud.google.com/text-to-speech/docs/quickstart-client-libraries.

Configure GOOGLE_APPLICATION_CREDENTIALS as described.

Make sure the gcloud command works!

Run:

web-reader convert URL OUTMP3

For example:

web-reader convert https://medium.com/@AnandWrites/209ffc24ab90 aspen.mp3

You can also select a local plain-text file:

web-reader convert-file ~/Documents/article.txt article.mp3

One-time: run web-reader init to set up the DB and MP3 dir.

Run the web server with just web-reader webserver.

Run the converter daemon with web-reader converter.

To set up Google TTS API auth, run with the appropriate environment, e.g.:

GOOGLE_APPLICATION_CREDENTIALS=... web-reader converter

Try submitting a web page with http://localhost:5000/api/v1/enqueue?url=SOMEURL.

For normal on-going use, you can use a handy bookmarklet for one-click submission of your current page. It tries to extract the main body content by default, but you can also just have some text on the page already selected when you press the bookmarklet to process just that selection:

javascript:var r=new XMLHttpRequest();try{r.open('POST','http://localhost:5000/api/v1/enqueue',false);r.setRequestHeader("Content-Type", "application/json;charset=UTF-8");r.send(JSON.stringify({url:document.location.href,body:window.getSelection?window.getSelection().toString():document.selection.createRange().text}));alert('done');}catch(e){alert('failed');}

Finally, you can reap the fruits by subscribing to the output Podcast RSS feed with http://localhost:5000/feed.

In all of the above, you should replace localhost:5000 with whatever final server you're hosting on.

To make your app server run at system startup, you can use something like the following Upstart scripts:

https://github.com/yang/personal-cm/tree/master/roles/webreader/etc/init

To make your web server accessible outside your firewall, localtunnel is a quick solution.

You can use the reconvert sub-command to force certain article IDs to get re-processed. This will just produce new MP3s in place. No new feed entries are created. Text won't be re-extracted - only the audio synthesis is rerun.

You can also use the resubmit sub-command to retry submitting old articles that failed. This will create new feed entries (and won't affect old entries). For instance,

web-reader resubmit http://localhost:5000 -d 2015-11-01 -o oldest -n 20

will resubmit to the AudioLizard web server running on port 5000 all the URLs that were ever only failures (i.e., ignore failures that got converted successfully another time), limited to a batch of 20 (starting with the oldest first).

This logic only considers distinct URLs as candidates for resubmission, which is usually the correct behavior.

1. Enqueue the web article URL into a PQ queue.
2. Fetch the page.
3. Extract the main body content (remove boilerplate) using both goose and boilerpipe, whichever recalls more.
4. Clean up the text with ftfy.
5. Split up text into sentences (iSpeech API only takes small chunks of text) using nltk sentence segmentation.
6. Submit each sentence to the iSpeech API to get back many MP3s.
7. Combine the MP3s (with 500ms of silence in between sentences) using pydub.
8. Generate the podcast feed with feedgen.

(Note: This is very outdated now!)

The best open source TTS implementation I'm aware of:

https://github.com/r9y9/wavenet_vocoder

Samples: https://r9y9.github.io/wavenet_vocoder/

However, it's extremely slow to synthesize (~20 minutes per second of speech on Colaboratory!). An open source implementation of parallel-wavenet is not there yet.

More resources:

• https://github.com/keithito/tacotron
• https://www.reddit.com/r/MachineLearning/comments/8utx7o/d_best_open_source_text_to_speech_networks/
• https://www.reddit.com/r/MachineLearning/comments/845uji/d_are_the_hyperrealistic_results_of_tacotron2_and/