📄 Issues in the Transcription of English Conversational Grunts

Summary

Short non-lexical sounds such as uh-huh, um, mm and oh turn up constantly in spoken English, yet no transcription convention captures them both reliably and usefully. Ward reviews three existing traditions. The first uses fully general phonetic notations (Trager, Poyatos), which are too heavy for everyday work. The second is function-driven corpus labeling (the Switchboard standard, Lander), which forces each sound into a fixed category such as filler or yes/no. The third is naive comic-book spelling, which is ambiguous. He then builds a "regularized" ASCII orthography: familiar spellings are kept, but each letter is given a defined phonetic value, and annotations are added for creaky voice and pitch. On his own conversation corpus of 317 grunts, the scheme handles about 97% of tokens. It still needs fallbacks into ordinary spelling, IPA or ad hoc tags. The work matters because these sounds signal attitude (enthusiasm, reluctance, boredom) and turn-taking, which affects dialogue systems, transcription and learners of English.

Key points

• Frequency: in the author's American English data, a grunt occurs about once every 5 seconds. In 79 Switchboard conversations, um ranked 6th among all items, and uh, uh-huh, um and um-hum together accounted for 4% of tokens.
• Function-based schemes fail for two reasons: the same sound appears in several roles (backchannel, filler, response, etc.), and the inventory is large and open-ended. Only 1% of corpus grunts were negative, so a yes/no top-level split is poorly motivated.
• Letter values in the scheme:
  • h marks breathiness, but a single syllable-final h is silent.
  • n marks nasalization.
  • u marks schwa.
  • uu marks a creaky schwa.
  • tsk marks an alveolar click.
  • A hyphen marks a strong syllable break.
  • A repeated letter marks length or several weakly separated syllables.
• Annotations follow a colon. :cr1–:cr3 marks degrees of creakiness. Pitch anchors use a 1–5 scale; for example, uh-huh:32-34 is the polite backchannel.
• Coverage was 306/317 (≈97%) of grunts. The scheme has not been tested with other labelers, and inter-annotator agreement has not been measured.

Interesting ideas

• I1. Transcribe the form compositionally (breathiness, nasality, boundary, pitch) and leave the function to context, instead of baking a fixed meaning into the label.
• I2. Since a single grunt can be a backchannel, a filler, an answer or a confirmation, its discourse role should be decided from its position in the exchange, not looked up from the token.
• I3. Start from a notation that people already half-know and tighten its ambiguous parts, rather than inventing a new expert notation. This keeps the scheme learnable without training.
• I4. Report explicit coverage of the inventory, with sanctioned escape hatches for leftovers, plus tie-breaking rules of thumb for borderline items (e.g. exclude laughter, include anything possibly communicative).
• I5. The author flags that cross-labeler agreement must be measured before a scheme is trusted.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): Interjection tokens like uh-huh or mm-hm could carry decomposed feature attributes rather than one opaque lemma, so later stages can reason about what the sound signals.
• I2 (frames): The speech-act layer could classify grunts in dialogue (assent, continuation, hesitation, bid for the turn) from neighbouring turns and keep that decision explainable.
• I3 (back): When generating dialogue text, Mova could map attitude frames (reluctant, enthusiastic) onto a small, consistent set of grunt spellings instead of arbitrary ones.
• I4 (learn): Annotation guidelines for silver data could state coverage rates and fallback categories openly, so that disagreements on rare cases are traceable rather than hidden.
• I5 (learn): Any new tag set Mova introduces should be checked by agreement between independent witnesses before it is used as training signal.