Resources · Stage 3 of 5
Transcribe a language sample
Transcription records what the speaker said using the conventions for the intended analysis. Those conventions determine utterance boundaries and how repetitions, revisions, and partially understood speech are represented. You can begin manually or review a draft produced by speech recognition.
On this page
Choose the conventions for your analysis
Use the transcription guidance associated with your chosen approach. A transcript prepared for one system may need changes before it can be analyzed in another, even when both report measures with familiar names.
Useful starting points include:
- SALT transcription resources, including its conventions, segmentation rules, and language-specific guidance.
- SUGAR training and procedures, including its transcription module, procedure handout, and FAQs.
- The CHAT manual, which documents the transcript format used by TalkBank and CLAN.
If you collected the sample using SLAM materials, its Guidelines for Analysis support recording and examining the responses. To add a measure such as MLU, choose a specific calculation procedure before preparing the transcript for it.
“MLU in words” or “MLU in morphemes” identifies what you count, but leaves the utterance boundaries and inclusion rules unspecified. Dividing the same speech into more units changes the denominator; treating repetitions differently can change the numerator. Use the segmentation and counting rules together from the selected procedure. For a reference comparison, follow the conventions used to produce those reference values, along with its collection requirements. A sample elicited with SLAM materials does not automatically qualify for a SALT or SUGAR reference comparison because you calculated a measure with the same name.
Record the conventions with the transcript so they are available when you analyze it or compare it with another sample.
Divide speech into utterances
Utterance boundaries affect both the number of units and the amount of language in each. Speech-recognition punctuation may differ from the boundaries required by the analysis method.
SALT’s C-unit segmentation rules consider grammatical relationships alongside pauses and intonation. The guide covers clauses, fragments, and responses to questions under the conventions used in its English reference databases.
For example, a pause in “She stayed inside … because it was raining” falls between a main clause and its subordinate clause. Under the C-unit guidance above, that pause alone does not create a new unit. The grammatical relationship and surrounding intonation inform the boundary decision.
Annotate for the analysis you are using
In “He went to the—she went to the park,” the speaker revises the subject after beginning the sentence. Using SALT’s maze notation, the transcript line would be:
C (He went to the) she went to the park.
Here, C identifies the child. Parentheses mark the abandoned wording as a maze; the completed production remains outside them. SALT excludes maze words from the main body used for measures such as MLU, while retaining them in the transcript. This represents both the revision and the language used for the calculation. See sections 10 and 15 of the SALT conventions.
Methods differ in their treatment of repetitions, revisions, fillers, and incomplete productions. SALT’s transcription conventions explain its notation and counting rules; SUGAR and CHAT provide their own guidance in the resources above.
For speech-sound analysis, IPA transcription can accompany the word transcript. The intended analysis determines the level of phonetic detail and whether you need both target and actual productions represented.
Use speech recognition as a starting point
The AusKidTalk transcription study describes a workflow using speech recognition followed by manual correction for a child-speech corpus.
Compare the draft with the recording at the level needed for your analysis. The following checklist identifies discrepancies to look for:
| What can happen | What to watch for in the draft |
|---|---|
| Grammatical forms differ | A grammatical ending or word in the draft differs from what is audible in the recording. |
| Repetitions or revisions are omitted | Repeated words, abandoned starts, or an earlier part of a self-correction are missing. |
| Speech is assigned to the wrong person | A clinician’s prompt appears under the student’s name, or a brief response is attached to the preceding speaker’s turn. |
| Words are missed or inserted | Audible words are absent, or text appears where the recording does not support it. |
| Automatic segmentation differs from the analysis conventions | One analysis unit is split across lines, or separate units are joined together. |
The checklist guides review of an individual transcript; it does not indicate how often a particular system makes these errors.
For an unresolved passage, retain a timestamp or audio link so you can replay it with the preceding and following turns. Those turns may help identify the speaker, clarify a response, or distinguish a revision from a new utterance.
Mark partially understood speech
Suppose you hear “She put the,” followed by an unclear word, then “on the table.” Most of the utterance is understood. The method’s notation represents the unclear portion alongside the words you could identify, preserving the available language for analysis.
The CHAT manual distinguishes levels of unintelligibility and provides notation for uncertain transcriptions. An unclear word, several unclear words, and an entirely unintelligible utterance may be treated differently in analysis. The selected method specifies which material can enter each measure.
Analyze and interpret a language sample covers calculation, reference comparisons, and descriptions of language patterns in the completed transcript.
Sources
- LEADERSproject. (n.d.). SLAM guidelines for analysis. https://www.leadersproject.org/page/slam-guidelines-for-analysis
- MacWhinney, B. (n.d.). Tools for analyzing talk, Part 1: The CHAT transcription format. TalkBank, Carnegie Mellon University. https://talkbank.org/0info/manuals/CHAT.html
- Owens, R. E., Jr., & Pavelko, S. L. (n.d.). Downloads. SUGAR Language. https://www.sugarlanguage.org/downloads
- SALT Software. (n.d.). C-unit segmentation rules [PDF]. https://www.saltsoftware.com/media/wysiwyg/tranaids/CunitSummary.pdf
- SALT Software. (n.d.). Transcription aids. https://www.saltsoftware.com/resources/tranaids
- SALT Software. (n.d.). Transcription conventions: Standard SALT conventions [PDF]. https://www.saltsoftware.com/media/wysiwyg/tranaids/StandardConventions.pdf
- Szalay, T., Nan, Z., Huang, R., Shahin, M., Tharmakulasingam, S., Ballard, K., & Ahmed, B. (2026). AusKidTalk: Developing transcription guidelines for continuous Australian English child speech. In Proceedings of the Fifteenth Language Resources and Evaluation Conference (pp. 5794–5804). https://aclanthology.org/2026.lrec-1.458/
