Turning speech into text.
WHAT THE TASK IS
Converting an audio signal into the words spoken.
WHAT MAKES IT DIFFICULT
Variation between speakers Accents and dialects Background noise Overlapping speech Words that sound identical Continuous speech without clear boundaries
WHAT TRADITIONAL SYSTEMS COMBINED
An acoustic model A pronunciation dictionary A language model
WHAT MODERN SYSTEMS DO
Learn the mapping directly, from audio to text.
WHAT THAT REQUIRED
Very large quantities of transcribed audio.
WHAT SELF-SUPERVISED APPROACHES CHANGED
Learning from untranscribed audio, then adapting with a smaller transcribed set.
WHY THAT MATTERS ENORMOUSLY HERE
It makes systems for languages with little transcribed data feasible.
WHAT ACCURACY DEPENDS ON
How closely the audio resembles training data.
WHAT THAT MEANS PRACTICALLY
Systems trained mainly on certain accents perform worse on others.
WHAT TO TEST WITH
Recordings from your actual users, in their actual conditions.
WHY
Published accuracy figures rarely reflect local speech.
WHAT TO PROVIDE
A correction route, since errors are certain.
