artcraft/old_code/trumpet-trumped.com/TODO.md

Trumped TODO

Improve synthesis system

• Speech modes:
  • "SHOUT" words for emphasis.
  • Question intonation in interrogative sentences.
  • Clear conjection words: "But I think..." vs. "Blah blah, BUT blah blah blah."
• Arpabet where the numbers are parameters: AW(0), AW(2), etc.
• Multiple word samples with unit selection
• Pluralization

Words to Fix

• you're / your is 'yer'. too short.
• want is 'wan'. too short.
• that's
• 'what' is 'wet'. bad.

Needed words

• need
• shut up
• suck/sucks
• tell
• idea

Shout words

• die
• favorite
• hands
• tiny

Phonemes to fix

• BReak
• badASS
• deMAND
• naME
• thinKING
• unfortuNate

OLD Last Minute Burndown

• Disable volume and speed parameters.

• In-Memory LRU Cache

• Better audio samples

  • Multiple "umm" / "uhh" samples
  • Fix words that sound bad

• User input filtering

  • Strip spaces in frontend.
  • Make sure there's no XSS injection.
  • Remove speed query parameter (for now).

• Frontend filesize

  • Downscale the animation. Maybe make a sprite sheet?

OLD Super Low Priority

• Frontend filesize:
  • Progressive loading of the animation (involved).
  • Zepto.js
  • Remove pixi.js dependency

OLDER Old Goals

HIGH LEVEL GOALS

1. Sound better
2. Handle load
3. Recoup costs
4. Look good (UI, speaking capability) - good reflection
5. Good code

Audio Samples

• Remove extra space between words (using parsing to insert spacing)
• Phonemes
  • Real versions of ['AO0', 'AW0', 'AW2', 'AY0', 'EY0', 'OY0', 'UH0'], which were backfilled with fakes.

Scaling

1. Load test
2. Read wave files into memory
3. Multi-server, with nginx as load balancer
4. Prevent others from calling it
5. Caching strategy - Caching headers (make sure they work) - Filter strings clientside to boost savings.

Tokenizing

• Plurals
• Times (9 p.m. ET)

Parsing

• (Everything in tokenizing)
• Decimal numbers
• Collapse pauses, eg: "Foo,,," -> "Foo (breath)"
• Numbers that are years (2016)!
• Twitter hashtags, usernames - conjoined word parsing (difficult)

Synthesis

• Get rid of static with "Uhhh", "Ummm" of variable length (option).
• Syllable System:
  • http://cdmckay.org/blog/2012/08/15/counting-syllables-and-detecting-rhyme-in-php/
  • http://stackoverflow.com/questions/33111685/convert-arpabet-to-ipa-with-stress
  • Or scrape something:
    • http://www.syllablecount.com/syllables/biologically

Cleanup

• Refactor into a very well-defined "SpeakRequest" with all parameters, including new ones like "is_twitter", "missing_word_mode = slience, record_pop, umm", etc.
• Remove old query parameters
• Set default volume

Server

1. Favicon, .ico and apple icon
2. robots.txt

Hardening

1. Reduce API surface
2. Limit length of input sentences
3. Rate limiting

Presentation

• Better frontend UI
• Testing page (current frontend)

Monetization

• Can I get sued?
• Ads
• Native Apps
  • (I don't have time to learn this.)
  • Contact other developers?
  • PhoneGap? (Blah)

Extra

• Clientside speed changing
• Let people type nonsense syllables

OLD Parsing Examples

• Camel case even outside of hashtags:
  • "Google acquires FameBit", StateOfW, AdaLovelaceDay
• Hyphenated: 2-party
• They will get horrible sentences, violently.We have the support.#RepealThe1st
• #100YearsStrong