What is Prosody?
The intonation, rhythm, stress and pauses in speech, as distinct from the words themselves.
The same sentence spoken with different prosody can mean the opposite. Emotional TTS and natural dialogue systems both depend on prosody annotation.
Prosody annotation is far more subjective than transcription, agreement between annotators is correspondingly lower, and it needs a tighter guideline and more review.
Types of prosodic features
Prosody is a bundle rather than one measurement. The features that get annotated are intonation (the rise and fall of pitch), stress (which syllable or word is emphasised), rhythm, tempo, pauses and loudness. A brief that says to annotate prosody has not yet said which of these the project needs. Most datasets mark two or three of them, not all six.
The features also sit at different levels. Sentence-level prosody marks where a question ends or a phrase boundary falls; word-level prosody marks emphasis inside an utterance. Some labels are categories, others are continuous tracks such as a pitch contour. In tonal languages pitch is part of the word itself rather than a layer above it, so those datasets are annotated differently.
What the prosody carries is a second split: linguistic meaning (statement versus question, new information versus old) against emotional state. Emotional TTS needs the second; a dictation system mostly needs the first. The two use different label sets and different guidelines, and a dataset labelled for one is rarely usable for the other without re-annotation.
What to specify when buying
State which features are labelled, at what level, and with what label set. A per-utterance emotion tag and word-level stress marking are different jobs at different costs. Because prosody is subjective, the guideline has to define each label with examples, including the boundary cases: how flat is flat, and when an utterance counts as neutral rather than amused.
Ask for inter-annotator agreement on a pilot batch before the full run. Prosody agreement runs lower than transcription agreement, and the number tells you whether the label set is workable before the money is spent. If two annotators disagree on a third of the batch, the labels will teach the model contradictions rather than a pattern, and no amount of model tuning repairs that.
Check how labels are delivered. Emotional TTS usually needs labels aligned to segments of the recording, not one verdict per file; a single file-level tag cannot drive synthesis at the phrase level, and a model trained on file-level tags learns one mood per clip. Aligned annotation costs more, and it is the difference between usable labels and decorative ones.
Related terms
-
TTS
Text-to-Speech, the task of generating spoken audio from written text.
-
Singing Voice
Vocal recordings with melody, in both a cappella and accompanied form, used for singing voice synthesis, music information retrieval, and lyric alignment.
Keep reading
-
Full glossary
Every term we explain, in one list.
-
How to buy training data
Where these terms actually show up, and which ones change a quote.
-
Help center
How a project runs, from specification to delivery.
Submit a sourcing request
Tell us the language, the hours, and what the data needs to look like. You will get a real number and a real timeline, not a range. If we cannot source it well, we will tell you that instead.
- Pilot batch before the full run, so problems surface early.
- Consent documentation delivered with the data.
- No medical or clinical data. No recorded telephone calls.
- You send the spec Language, hours, recording conditions, and the deadline.
- We reply within two business days A real number and a real timeline, not a range.
- You approve a pilot batch A small run first, so problems surface before the full order.
- Full run, then delivery with documents Audio, transcripts, and the consent record for every speaker.