How to Conduct a Language Sample: A Step-by-Step Protocol for SLPs

Every SLP who has ever sat down with a laptop, a preschooler, and a ten-minute slot between IEPs knows that "go conduct a language sample" is not a single skill — it is five distinct skills stacked on top of each other, and losing any one of them contaminates the whole sample. Elicitation that produces only two-word utterances gives you nothing to score. Recording equipment that drops out halfway through ruins an hour of clinic time. Segmentation done on the fly, in your head, while the child is still talking, reliably produces a different MLU than segmentation done from a clean transcript three days later. A written protocol is the only realistic way to make the whole workflow reproducible across children, clinicians, and caseload sizes.

The SUGAR (Sampling Utterances and Grammatical Analysis Revised) project at the University of Central Florida (Pavelko & Owens, 2017, 2019) was designed specifically around this reproducibility problem. Their published sampling protocol — 50-utterance conversational sample, elicited with a fixed set of prompts, transcribed into a standard format, segmented with clear rules — is the nearest thing the field has in 2026 to a clinical consensus on "what a language sample should look like". This pillar adapts that protocol for a school or outpatient clinician and integrates it with the free browser-based LSA tools on this site so that every step has a concrete tool attached to it.

The guiding principle throughout is clinical realism. A 50-utterance sample is not the research-grade ideal (Heilmann et al. 2010 show that longer samples are more stable), but it is the length that most school clinicians can actually collect and transcribe in a single IEP window. The protocol below is calibrated so that a full workflow — from consent to scored transcript — fits in roughly 60 to 90 minutes of total clinician time, which is the point at which language sampling starts being practical at scale rather than theoretical.

The ten minutes before the child walks into the room is where every avoidable language sample failure gets prevented. Write down the specific clinical question the sample is meant to answer before you plan the session — "does this 4-year-old’s MLU match age expectations?", "is this 6-year-old’s percent grammatical utterances in the clinical range?", or "is this 8-year-old’s narrative grammar structured enough for a school-age IEP?". The question determines the elicitation approach. A child whose question is MLU-focused needs a conversational sample; a child whose question is narrative structure needs a story-retell or story-generation task; a child whose question is grammaticality needs a sample long enough to give stable PGU.

Consent is the second step and it is the step that most commonly blocks the workflow three weeks later when you realise you never got written permission to audio-record. Every district has its own consent form; if yours does not, the minimum language to include is: (1) the purpose of the recording (clinical assessment), (2) who will hear the recording (the clinician and supervising SLP), (3) how long it will be retained, and (4) the parent’s right to withdraw consent and have the recording deleted. Get the signature before the session, not after.

Room setup is quick but non-negotiable. The room should be as quiet as you can realistically get — hallway door shut, HVAC vent not blasting directly over the microphone, no overlapping conversations at the other end of the SLP office. Sit at right angles to the child rather than directly across (right-angle seating keeps the mic away from direct blasts of air from the child’s mouth and reduces the "plosive pop" that makes disordered consonants hard to transcribe). A small table with age-appropriate materials in arm’s reach — picture books, a small set of figurines, a barrier game, a wordless picture book — is enough to carry a 20-minute session without the clinician having to reach for new props.

The single biggest determinant of sample quality is prompt choice. A 3-year-old asked "tell me about your weekend" produces one-word answers; the same child shown a wordless picture book produces ten-word stretches of connected speech. Prompt-to-age matching is the difference between a 50-utterance sample collected in 20 minutes and a 50-utterance sample that never materialises. The prompts below come from the SUGAR protocol, the Heilmann lab recommendations (Heilmann, 2010), and decades of clinical-room practice — they are calibrated to each developmental band to maximise both quantity and grammatical variety.

For children aged 3 to 4, concrete-object play and simple picture-book prompts outperform open questions. Favourites include the "What’s in the bag?" barrier game with 5-8 small objects, a wordless picture book like Mercer Mayer’s Frog Where Are You?, and a simple pretend-play setup with a dollhouse or farm animals. Avoid "tell me about..." prompts at this age because they consistently fail to elicit multi-word utterances in typically developing children, let alone children with language delay.

For children aged 5 to 6, narrative tasks become productive and are worth prioritising because they stress the syntactic complexity metrics (DSS, IPSyn) that are the dominant clinical targets at this age. Story-retell from a wordless picture book works well: read the book through once silently pointing at the pictures, then hand it to the child and ask them to "tell the story back". Story-generation from a single stimulus picture (for example, the SBS test stimuli or a National Geographic Kids magazine cover) also works but produces shorter samples. Conversational prompts ("what did you do at recess?", "tell me about your pet") start being productive at this age but are still not as reliable as structured tasks.

For children aged 7 to 11, conversational sampling is the default. School-age children produce long stretches of connected speech on topics they care about, and the clinician’s job shifts from "prompt harder" to "get out of the way". High-value prompts include recent movies or games the child likes, how to play a favourite video game (this is a goldmine for procedural discourse and complex syntax), and "tell me the plot of..." prompts on books or shows the child has recently watched. Narrative generation tasks still work at this age and are the right choice if your clinical question is specifically about story grammar.

For children aged 12 and up, expository tasks become the best source of complex syntax. Ask a teenager to explain how to play a sport, how their favourite video game works, or how to do something they are good at (bake a cake, skateboard, solve a Rubik’s cube). Expository sampling produces the longest T-units and the highest subordination indices of any elicitation task in this age range, which is precisely what you want for older-student LSA. Story-retell from a single short film or a short story still works but produces shorter samples than expository tasks for most teens.

A clean recording is the cheapest insurance a clinician can buy against wasted clinic time. The recording does not need to be studio-grade — a modern smartphone or a $30 USB lavalier microphone clipped to the child’s shirt produces audio that transcribes cleanly in 2026, as long as the room is quiet and the mic is within 12 inches of the child’s mouth. The specific device matters less than three simple rules: proximity, redundancy, and pre-session test.

Proximity means the microphone should be close enough that the child’s voice is louder than the room noise on the recording. If you play the first ten seconds back and the HVAC is competing with the child, move the mic closer or turn the HVAC off. Redundancy means you should record on two devices in parallel whenever possible — phone plus laptop microphone, phone plus a cheap USB recorder, whatever you have. Audio files get corrupted, phones run out of storage, and a dropped recording after a 30-minute session is a cost you cannot afford on a school-clinician schedule. Pre-session test means you press record, say "test one two three" in the child’s seat, play it back, and confirm you can hear clearly before the child walks in. Every experienced clinician has a story about the recording that looked fine on the meter and turned out to be 30 minutes of silence.

For the actual file format, a lossless or near-lossless recording (WAV, M4A, MP3 at 192kbps or above) is what you want. Voice memo apps on iOS and Android default to M4A at sufficient quality for LSA transcription. Do not use speech-to-text-to-file pipelines during the session itself — you want the raw audio, and any transcription step runs after the session when you have the time to edit it. Save the file with a child identifier that matches your district’s naming convention (never the child’s full name in the filename) and move it to your encrypted clinical drive within 24 hours.

The session itself is where clinician experience compounds. The target is 50 scorable utterances from the child, which typically takes 15 to 25 minutes of clock time for a conversational sample and slightly longer for narrative or expository tasks (because the clinician has to set up the stimulus first). The clinician’s job during the session is to keep the child talking without contaminating the sample with adult-scaffolding utterances that the child then parrots back.

A few concrete rules help. First, minimise yes/no questions — they elicit yes/no answers, and yes/no answers are not scorable for MLU, DSS, or most grammatical-accuracy metrics. Open-ended prompts ("tell me more", "what happened next", "and then?") consistently outperform closed questions for sample length. Second, resist the urge to expand or recast the child’s utterances during the sample. Recasting is a valuable therapy technique but it contaminates the assessment sample because the child may repeat your scaffold verbatim. Third, avoid giving the child multiple-choice prompts ("was it red or blue?") because they elicit one-word answers. Fourth, allow silences. Children fill conversational silences with their own speech — an adult who rushes to fill every pause will produce a sample that is mostly adult turns, which is exactly the opposite of what you want.

As the session runs, the clinician should keep a loose tally of how many child utterances have been produced so that you do not run long. Fifty utterances is the floor; if you are at 20 utterances after ten minutes and the child is giving you useful material, keep going for another ten. If you are at 10 utterances after 15 minutes because the child is closing down, switch prompts or change the stimulus object before forcing more — a sample collected from an increasingly reluctant child will have MLU and PGU values that reflect the child’s reluctance rather than their underlying competence.

Mark the end of the session clearly for the recording: say something like "okay, thank you, we’re all done" at a normal volume. This gives you a clear audio landmark when you transcribe later and it signals to the child that the assessment portion is over. If you are continuing with other clinical activities afterwards (play, games, standardised testing), stop the recording first so you do not contaminate the file with non-sample audio.

Segmentation is the least-glamorous and most error-prone step in LSA. Every MLU, DSS, IPSyn, and PGU calculation depends on where the clinician draws the boundaries between utterances, and small segmentation changes can shift MLU by half a morpheme in either direction on a 50-utterance sample. The good news is that the rules are deterministic once you pick a framework; the bad news is that every framework picks slightly different rules, so you have to commit to one and use it consistently.

The C-unit (Communication Unit) framework is the default for SUGAR-style school-age sampling. A C-unit is an independent clause plus any modifiers and subordinate clauses attached to it. The rules: a C-unit ends at a main-clause boundary, a coordinating conjunction that starts a new main clause, a clear intonational fall, a long pause, or a speaker change. Run-on sentences connected with "and then... and then... and then..." are segmented into separate C-units at each "and then". The T-unit (Terminable Unit) framework, developed by Hunt for school-age writing analysis, is similar but slightly more permissive about embedded coordination. Brown’s original framework for preschool data (Brown, 1973) is still the right choice for ages 3-5 and uses utterance boundaries defined primarily by pauses and intonation rather than grammatical completeness.

A small number of conventions apply to all three frameworks and are worth committing to memory. First, repetitions and stutters inside an utterance are excluded from the word count but the utterance itself is still scored — a child who says "I-I-I went to the store" has an utterance of length four ("I went to the store"), not six. Second, mazes (the child’s false starts, fillers, and self-corrections) are excluded from MLU counts but tracked separately as a fluency measure. Third, unintelligible utterances are excluded from MLU and NDW but kept for the intelligibility calculation — you cannot score what you cannot transcribe. Fourth, imitations of the clinician’s immediately preceding utterance are excluded from the sample entirely because they reflect the clinician’s grammar, not the child’s.

The single most important habit is consistency. Pick one framework (SUGAR’s C-unit rules are the easiest to explain to new clinicians and match the published normative data on this site), write the framework name on the transcript header, and use it the same way every time. Clinicians who switch frameworks between children introduce scoring noise that swamps any real change in the child’s grammar.

Once the session is recorded, you have to turn the audio into a typed transcript that a calculator can score. This is the single slowest step in LSA and it is the reason most caseloads run on standardised tests instead of language samples. Three transcription approaches are realistic in 2026: manual, AI-assisted, and fully-automated with clinician review. Each has different time costs, different error profiles, and a different place in a high-caseload clinician’s workflow.

Manual transcription is the gold standard for accuracy and the hardest to justify on a school caseload. A trained clinician transcribing a 50-utterance preschool sample at normal listening speed takes 45 to 60 minutes; a sample from a child with disordered speech can take 90 minutes or longer. Manual transcription captures the clinician’s ear for speech errors, mazes, and pragmatic context in a way that no automated tool currently matches. It is the right choice when you have a complex differential diagnosis question, when the recording is poor quality, or when the child’s speech is significantly disordered and an automated transcript would miss the clinically relevant errors.

AI-assisted transcription is the most common approach in 2026 for typical-speech samples. The workflow is: run the audio through a modern speech-to-text engine (OpenAI Whisper, Google Chirp, or a clinical-SLP tool like ConductSpeech that wraps a speech recogniser with LSA-specific post-processing), get a rough transcript in under two minutes, and then clean it manually — fix misrecognised words, mark the speaker turns, apply your segmentation framework, and exclude mazes. The total clinician time drops to 15-25 minutes per sample, which is the threshold at which language sampling starts being viable at scale. Accuracy on typical preschool speech is roughly 90-95% and the clinician’s edit pass closes the remaining gap.

Fully-automated transcription with clinician review is the newest approach and the one that ConductSpeech is built for. The workflow is: upload audio, the tool produces a transcript, segments it into utterances using the Brown/SUGAR rules, tags morphemes in obligatory contexts, computes MLU-m, MLU-w, NDW, and PGU, and drafts a present-levels paragraph. The clinician’s job shifts from typing to reviewing: skim the transcript for misrecognised words, confirm the segmentation, spot-check the morpheme tags, and edit the draft paragraph. Total clinician time on a typical-speech sample drops to 5-10 minutes. The approach does not replace clinical judgement — the transcript, the segmentation, and the scores are all editable — but it changes the calculus of "should I do a language sample this week?" from "no, I don’t have 45 minutes" to "yes, I have 10".

Once the transcript is segmented and cleaned, scoring is the fastest step of the whole workflow. The free browser-based LSA calculators on this site are designed specifically for the paste-transcript-click-score pattern. The MLU Calculator takes a plain-text transcript (one utterance per line), applies Brown’s morpheme rules in code, and returns MLU-m, MLU-w, total morphemes, total utterances, and the matching Brown’s stage in one pass. The Lexical Diversity Calculator returns NDW, TTR, MATTR, and vocd-D from the same paste. The DSS Calculator implements the Lee (1974) eight-category syntactic scoring. The IPSyn Calculator implements the 60-item Index of Productive Syntax. The PGU Calculator returns percent grammatical utterances using the SUGAR rule.

Which calculator to run depends on the clinical question. For ages 3-5 with an MLU-focused question, run the MLU Calculator and cross-reference the Brown’s Stages Lookup to get the stage and expected morpheme mastery. For school-age children with a syntactic-complexity question, run the DSS Calculator or the IPSyn Calculator depending on which metric your district report template expects. For older students with a grammaticality-focused question, run the PGU Calculator to get percent grammatical utterances and cross-reference the SUGAR Norms Lookup. For any child, the Lexical Diversity Calculator gives you NDW in under a second.

The scoring step should take no more than 5 minutes of clinician time once the transcript is clean, because every calculator runs in the browser and requires only a paste-and-click. The clinical interpretation takes longer — matching the scores against published norms, writing the present-levels paragraph, and drafting the IEP goals — but the arithmetic itself is effectively free. This is the single biggest reason the free-tools approach is viable for a high-caseload school SLP: it is not that the scoring is faster than SALT (it is comparable), it is that there is no licence cost, no installation step, and no transcription-format learning curve. You paste, you click, you get the numbers.

The five most common failure modes in clinical language sampling are not clinical mistakes — they are workflow mistakes, and they are all preventable with a written protocol. They come up at every district training and they are worth spelling out so that new clinicians do not have to discover them by burning clinic time.

The workflow recommendation that comes out of this protocol is calibrated for a school or outpatient clinician with a caseload of 20 to 50 students and a district that expects ecologically valid LSA at IEP reviews. It combines the free tools on this site, a modern recording approach, and either AI-assisted or fully-automated transcription depending on caseload size. The recommendation is not the research-grade ideal — longer samples, manual transcription, and multiple coders would all be better in isolation — but it is the stack that actually fits the clinical week.

The recommended workflow, start to finish, is: (1) write the clinical question down, (2) get parent consent to audio-record, (3) print the Language Sample Worksheet on this site for the age band, (4) set up a quiet room with props in arm’s reach, (5) record the session on two devices in parallel, (6) target 50 scorable utterances in 15-25 minutes, (7) transcribe with AI assistance (Whisper, Chirp, or ConductSpeech) and clean manually, (8) segment using the C-unit framework, (9) paste into the MLU Calculator and the metric calculator matching your clinical question, (10) cross-reference the Brown’s Stages Lookup or SUGAR Norms Lookup for the age-banded norm, and (11) use the IEP Goal Generator to draft the present-levels paragraph and annual goals. The whole pipeline fits in 60 to 90 minutes of clinician time per child, which is the threshold where LSA stops being a luxury and starts being the default assessment.

For clinicians whose caseload or time budget makes even that pipeline tight, ConductSpeech is built specifically to automate steps 7 through 11 — audio in, scored transcript and draft IEP goals out — so the clinician’s role is to record, review, and edit. The honest framing is not that ConductSpeech replaces clinical judgement but that it reallocates clinician time from the slow transcription step to the fast interpretation and goal-writing step. For every caseload size, though, the protocol above is the frame: the specific transcription approach changes with caseload, but the elicitation, segmentation, and scoring steps stay the same.