alignment-from-stt

Fork of BBC/alignment-from-stt

Work in progress, originally developed by Mathieu Triay, in BBC R&D as part of the New News prototype "fast forward". Being refactored into a reusable module by News Labs.

Setup

stack - optional

How to build and run the code/app

git clone [email protected]:pietrop/alignment-from-stt.git

cd alignment-from-stt

npm install 

Usage

npm install alignment-from-stt

const fs = require('fs');
const align = require('alignment-from-stt');
const automatedTranscription = require('../sample-data/kaldi-alignments.json').words;
// array of words, each .end, .start., .transcript_text;
const humanCorrectedTranscription = fs.readFileSync('./sample-data/human-transcription-on-one-line.txt').toString();

const reAlignedTranscription = align(humanCorrectedTranscription, automatedTranscription);

console.log(reAlignedTranscription)

See examples in src/example-usage.js.

Optional attributes

var reAlignedTranscription = align(
  humanCorrectedTranscription,
  automatedTranscription,
  'word', // or punct, - optional textAttributeName
  true, // defaults to true - optional segmentationBasedOnSBD
  });

textAttributeName attribute for text field in word attribtue

{start: 0.2, end: 3, word: 'hello'} 

or

{start: 0.3, end: 6, punct: 'world'} 

by default is set to text.

segmentationBasedOnSBD by default uses sbd for sentence boundary detection. if set to false, then it spltis on \n. Helpful if you need more fine graine control on the lines you want to align.

System Architecture

Input is

• a json, array of words objects, with start, end, and transcript_text.
• plain text, humanan accurate, correct text.

Then lib/text-segmentation uses a sentence boundary detection module to split the human accurate plain text into an array of sentences. This avoid threating honorifics like Mr. as if they were sentence boundaries.

lib/alignement uses the segmented sentences and transposes the timecodes from the json (from STT) onto the accurate text.

...
  { start: 578.43,
    end: 582.52,
    text: '"The only thing I\'m here to discuss is your resignation," said Pashinyan.',
    words: [] },
  { start: 582.92,
    end: 594.12,
    text: 'Mr. Sargsyan walkedout, saying that Pashinyan had no right to speak on behalf of the people when his party had only received 7% of votes in the last elections.',
    words: [] },
  { start: 595.1,
    end: 600.86,
    text: 'Shortly afterwards, Nikol Pashinyan and other protest leaders were detained by the police.',
    words: [] },
...

At the moment it does so at sentence, level. Ideally will get to a point where it also reaches word level.

Development env

• node > 8
• npm > 5.6.0

Build

No build step for now

Tests

Unit tests written with jest.

npm run test

or

npm run test:watch

If you want to run tests while developing - recommended.

See docs on sample material

Deployment

publish to npm

npm run publish:public

TODO:

• [ ] there is an issue with this char ’ that is being escaped as \\\' in final result. it comes from .txt input file. should add check to replace as '.