I know it becomes harder to find the start and end of phonemes due to co-articulation. I want to know about other problems that co-articulation causes in speech recognition.

  • Multiply-articulated sounds like labiovelars or aspirated stops may well be heard as different sounds. They are also subject to great individual variation because of the independent articulations, which must be timed precisely and usually aren't -- voice onset time is never exactly at the beginning of a voiced segment, nor is lip-rounding always timed exactly at the beginning of a labiovelar. So we say, and hear, different things all the time. All we have to depend on is our habits, and our impression of others' habits.
    – jlawler
    Jan 12, 2019 at 18:23

1 Answer 1


Modern data-driven approaches to speech recognition are ignorant about but robust to this one of many aspects of orthography that are not fully phonetic but occur relatively consistently in the training data.

For one, it is more an issue of mapping n:1 (different realisations of one set of characters), whereas challenges for speech recognition generally revolve around:

  1. 1:n (ambiguity):

    • homophones like doe vs dough, see vs sea, wreck a nice beach vs recognise speech

    • meta language like contact at priceline.com vs [email protected], anneal- vs Anil Dash, He told me quote unquote I will kill you vs He told me "I will kill you.", acronyms

  2. shift vis-a-vis the training data:

    • out-of-domain and domain-specific words

    • variations in accents and dialects

    • a foreign language or mixed language

    • background noise

  3. the fact that users don't really want a literal transcription:

    • I'ma, um, I'ma call you on, um, Saturday vs I'm going to call you on Saturday

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.