[DRAFT] Workflow of Common Voice

This page is unofficial. Please refer to the forum and official website for accurate information.

1. Sentence
   • Contributor: Who knows grammar
   • Tool: Sentence Collector
   • To do: Find, add and review public domain sentences
   • Goal: 1,800,000 sentences
   • Prepare: Request a language to be added in Common Voice category of Mozilla Discourse
     • Collecting voices requires 5,000 sentences
2. Voice
   • Contributor: Who can read a sentence
   • Tool: Common Voice
   • To do: Record and validate the voices
   • Goal: 2,000 hours of validated voice
   • Prepare: Translate the site with Common Voice's Pontoon
3. Training
   • Contributor: Engineer
   • Tool: Datasets
   • To do: Find the right data for the purpose, train the system
   • Goal: Create a system to understand speech
   • Prepare: Download datasets
4. Working
   • User: Everyone
   • Tool: Application
   • To do: Enter a voice, or a sentence for the app to read
   • Goal: Depending on the user
   • Prepare: Choose an app

• The end user of the dataset is the engineers who train the speech recognition system, although the final end user is everyone (or the system's target population).
• The origin of the dataset - voice recording - is that of sentence collection.
• ref:
  • Mozilla Voice Community Playbook
  • Readme: How to see my language on Common Voice
  • Common Voice languages and accent strategy v5

Sentence flow

• You can do both collecting and reviewing, or just one of them.

1. Find public domain texts
   • ref. Find (Task Setting)
2. Editing a sentence
   • ref. How-to on Sentence Collector
   • The easier to read and the shorter the better.
   • Remember, some people are not native speakers or have trouble reading out loud.
3. Add to Sentence Collector
   • ref. Add (Task Setting)
4. Review a sentence
   • ref. Review (Task Setting)
5. Approved sentences are added to Common Voice
6. Volunteer reads a sentence

Voice flow

• Details: How does it work? - Common Voice
• You can do both recording and validation, or just one of them.

1. Record a voice
   • ref. Recording (Task Setting)
2. Validate an audio clip
   • ref. Validate (Task Setting)
3. Datasets with additional clips are released
4. User downloads the dataset
5. Engineer uses dataset to train speech recognition system (or use it for other purposes)