What is Phoneme?

Phonemic annotation is finer-grained than transcription: it records the sounds actually produced rather than the spelling. It is especially useful for languages with wide pronunciation variation, such as the Arabic dialects.

Phonemic annotation costs noticeably more than transcription, so it is usually only done when the client has a specific need for it.

Levels of phonetic annotation

Phonetic annotation comes in two depths. Broad transcription writes one symbol per phoneme and ignores the differences between sounds that do not change a word's meaning. Narrow transcription records how the sound was actually produced: the burst of air on a consonant, the length of a vowel, the way one sound shifts towards its neighbour. Broad is faster and cheaper; narrow captures the detail accent work depends on.

The second split is how the labels are produced. Manual annotation has a trained person listen and write the symbols, which is slow and is the only way to handle unusual pronunciation or speech the dictionary does not cover. Alignment takes an existing transcript and a pronunciation dictionary and matches the two, placing boundaries automatically. Most phoneme-level labels in large corpora come from alignment.

Behind both sit two artefacts that have to be chosen before work starts. The phoneme inventory is the symbol set, and the International Phonetic Alphabet is the standard, though many speech tools use their own ASCII-friendly sets such as ARPAbet for English. The pronunciation lexicon maps each word to its sequence of phonemes, and it has to match the language and the variety being recorded.

Where it shows up in a quote

Phoneme-level annotation is bought on top of a transcript, not instead of one, and it is priced per audio hour at a rate well above plain transcription, because it takes a trained annotator rather than a fast typist. A quote that offers phonetic labelling at transcription prices is worth questioning, since the two jobs take different amounts of time and skill.

Four things belong in the specification: broad or narrow; which symbol set, with a copy of the inventory if the supplier uses their own; the language and variety; and whether the output is time-aligned to the audio. One more decision is easy to miss, which is what happens to words the lexicon does not contain. Names and new coinages have no entry.

The use case decides whether it is worth the multiple. Pronunciation scoring and language learning depend on phoneme labels directly, as do speech synthesis front ends, which need to know how a word is said before they can say it. Keyword spotting and accented-speech work benefit from them too. For a general recognition model, the transcript already carries what the model needs.