Skip to lesson

Path: Bharat language & voice

Evaluate speech and transcripts

Distinguish what was spoken, what is audible, and what a transcriber has inferred.

Starting level
Some basics
Time
22 minutes
Public lesson guide
EN + हिंदी

You will learn to

Mark audible speech, uncertainty, and transcription inference separately.

Why this matters

Guessing an unclear word turns uncertainty into false training data.

Three things to remember

  • Use timestamps for every issue.
  • Keep spoken words separate from reviewer inference.
  • Use the project's uncertainty convention consistently.

See one example

Task

Traffic noise masks one word in a sentence.

Weak approach

Guess the likely word from context.

Stronger approach

Mark the span unclear at the timestamp and preserve the surrounding words.

Why the stronger approach works

Faithful uncertainty is higher-quality data than a plausible guess.

Make something yourself

Create a three-line transcript review with timestamp, observed issue, and confidence.

A useful starting pointDo not invent missing speech. Use an explicit unclear marker.
Practise this lesson

Read further

The source is optional. The complete teaching and practice are available inside AI2Bharat.

Read the primary or official source: BhashaDaan