Skip to content

Latest commit

 

History

History
166 lines (131 loc) · 4.05 KB

segmentation.md

File metadata and controls

166 lines (131 loc) · 4.05 KB
Raw

Text segmentation

Input texts could be "segmented" (and configured) in different parts: characters, words, phrases.

  • Static phrases
  • Template literals
  • Word-by-word concatenation
  • Character-by-character spelling

Static phrases

It's the simplest scenario: you have a list of static, immutable, ready done, phrases. By example:

Looks like her company has three containers set to sail for tonight.

The above sentence corresponds to:

  • an audio (speech) file with some naming convention

    by example: your/path/speech/en/looks_like_her_company_has_three_containers_set_to_sail_for_tonight.ogg

  • or with a obscure name/UUID (maybe made with nanonid)

    by example: your/path/speech/en/V1StGXR8_Z5jdHi6B-myT.ogg.

const {ttsfile} = require('jointts')
const fileName = ttsfile('Looks like her company has three containers set to sail for tonight', 'en')
// -> 'your/path/speech/en/looks_like_her_company_has_three_containers_set_to_sail_for_tonight.ogg'

Template literals

Template literals are string literals allowing embedded expressions. They are static phrases containing also variable parts (entities) to be resolved at run-time. By example:

Container JL1349-76 has been cleared for pick-up.

in the sentence up here, the entity JL1349-76 is a domain specific code (a shipping container code), to be spelled as a {alphanumericCode}. At the configuration level, a template literal could have a syntax like:

Container {alphanumericCode} has been cleared for pick-up.

That "template literal" be a concatenation of 3 strings component parts:

  • Container , a static string
  • {alphanumericCode}, an alphanumeric code to be spelled char-by-char
  • has been cleared for pick-up., a static string

At run-time, the TTS translation function must recognize the template literal, concatenating the sequences.

A possible API could be somthing like:

const {ttsfile} = require('jointts')
const fileName = ttsfile(
  'Container {alphanumericCode} has been cleared for pick-up.', 
  'en', 
  { alphanumericCode: 'JL1349-76' } 
)
// -> 'Container JL1349-76 has been cleared for pick-up..ogg'

Words concatenation

Phrases are built concatenating words and or letters (that is the general case of concatenative text-to-speech). Apparently it's a "worst case" because the system has to preset all the words of the grammar.

In the example:

Container JL1349-76 has been cleared for pick-up.

The sentence could split in blank separated word tokens:

Container
JL1349-76
has
been
cleared
for
pick-up.

The downside of this approach is that having a file for each word, implies a run time concatenation, with possible unnatural sounding play, due to the static file sequencing.

Character-by-character spelling

joinTTS foresee a ready done letters/symbols set of audio files, for spoken spelling of single chars, acronyms, numbers, or alphanumeric codes, unknown words (not included in custom grammar):

Examples (alphanumeric codes):

RAIU 690011 4 25 U1
CSQU3054383
1006.760

BTW, the two initial lines are container codes with ISO6346 format (see my npm package iso6347).

In the case of an alphanumeric sequence of chars, the required spoken spelling is the concatenation of letter-by-letter speech. The original input text:

CSQU3054383

is spelled as sequence of single characters (separated by brief pauses):

C S Q U 3 0 5 4 3 8 3

Where the spelling of each character corresponds to the pseudo-phonetic spelling audio, depending on

A basic TTS for English language could be

sii es kiu iu three zero five four three eight three

Whereas, using the NATO phonetic alphabet:

Charli for sii 
Sierra for es
Quebec for kiu
Uniform for iu
three
zero
five
four
three
eight
three

or simply:

Charli
Sierra
Quebec
Uniform
three
zero
five
four
three
eight
three

top | home