TOEFL speaking practice: train for the machine that scores you

TOEFL speaking has a property no other part of your test prep does: part of your score comes from software. ETS rates the speaking section with human raters plus an automated engine (SpeechRater), which means your delivery is literally being measured by a machine — pace, pauses, vowel quality, rhythm. You can practice for that at home more directly than for any human-judged exam, if you practice the right things. Here's the full picture.

The 4 TOEFL iBT speaking tasks

The section takes about 16 minutes. You speak into a microphone; nobody speaks back.

1. 1.Task 1 — Independent. An opinion question ("Do you agree that…?"). 15 seconds prep, 45 seconds to answer.
2. 2.Task 2 — Integrated, campus situation. Read a short campus announcement, listen to two students discuss it, then summarize one speaker's opinion and reasons. 30 seconds prep, 60 seconds.
3. 3.Task 3 — Integrated, academic concept. Read a short academic passage, listen to a lecture excerpt that illustrates it, explain the concept using the example. 30 seconds prep, 60 seconds.
4. 4.Task 4 — Integrated, lecture only. Listen to a lecture, summarize its main points. 20 seconds prep, 60 seconds.

Each response is scored 0–4 on general description, delivery, language use, and topic development, then the section converts to 0–30.

How automated scoring changes your prep

SpeechRater-style engines extract measurable features from your audio: speaking rate, length and placement of pauses, pronunciation clarity, vowel quality, stress and rhythm patterns, and how varied your vocabulary and grammar are. Human raters still judge content, but delivery is machine-measured — so machine-readable clarity pays directly.

Practical consequences:

• Mumbled-but-clever loses to clear-and-simple. A plainly-worded answer delivered with clean sounds and steady pace outscores a sophisticated answer full of hesitations and swallowed vowels.
• Fillers are measured. Every "um" and dead pause is a data point. So is a pause in the wrong place — mid-clause instead of between ideas.
• Individual sounds matter. If your /ɪ/ vs /iː/ or /θ/ collapse, the acoustic features degrade even when a patient human would understand you fine.

This is also why practicing into a normal transcription app misleads you: standard speech recognition auto-corrects your errors into real words, so it hides exactly what an acoustic scoring engine will penalize. sayit's recognizer is language-model-free — it scores the raw sounds you produced, per phoneme, which makes it an unusually honest proxy for how a delivery-scoring machine hears you. The mechanics are in how AI pronunciation scoring works.

Templates: decide the structure now, spend the 15 seconds on content

You cannot invent a structure in 15 seconds of prep. Memorize one skeleton per task type, and use prep time only to slot in content.

Independent (Task 1):

I believe that [position], for two main reasons. First, [reason one] — for example, [10 seconds of example]. Second, [reason two] — in my experience, [example]. That's why I think [restate position].

Integrated campus (Task 2):

According to the announcement, the university plans to [change] because [reason]. The man/woman disagrees/agrees for two reasons. First, [reason + detail]. Second, [reason + detail].

Integrated academic (Tasks 3–4):

The reading/lecture introduces the concept of [term], which means [one-sentence definition]. The professor illustrates this with [example]. First, [step one]. Then, [step two], which shows [link back to the concept].

Say each template out loud until the connectors (first, second, for example, that's why) come out automatically with natural stress — connectors are where rhythm falls apart under time pressure. A template you've said fifty times also buys you something the scoring engine can measure: pauses that land between ideas instead of inside them, because half of every response is already wired into muscle memory. Our guide to word stress and intonation covers why stressing content words and gliding over function words instantly sounds more fluent.

The daily drill plan (40 minutes)

1. Timed response reps (15 min)

Take sayit's TOEFL-style speaking exams: prompts with real prep/response timers. Do three Task 1s back-to-back — 15 seconds prep, 45 seconds speaking, stop exactly on time. The skill you're building is finishing your second reason at 0:42, and it only comes from reps against a clock. Review the fluency and completeness scores after each rep; hesitation you can't hear in yourself shows up in the numbers.

2. Delivery work: read-aloud + shadowing (15 min)

• Read-aloud mode on a leveled passage gives you the per-phoneme heatmap: every word color-coded, each miss shown against its target IPA with a concrete articulation tip. Drill your two weakest sounds — the Magic Wand coach will find them across your history and build the plan for you. For stubborn sounds, start from the hardest English sounds and look words up on the pronunciation hub.
• Shadowing — speaking along with a model at natural speed — is the fastest way to fix rhythm and pause placement, which are precisely the features automated scoring measures.

3. Dictation for the listening half (10 min)

Tasks 2–4 are listening tests wearing a speaking costume: if you miss the professor's example, no delivery saves the response. sayit's dictation mode — listen, type what you heard, get scored against the source — trains the note-worthy-detail ear directly. Ten minutes a day compounds fast, and it doubles as vocabulary work.

All four modes are on the free tier — browser-based, no card; see features for the full list and pricing if you outgrow the free minutes.

A one-week checkpoint

After a week of the plan, run a full four-task exam in one sitting and check three trends in your progress analytics:

1. 1.Pronunciation score trend on read-aloud — your machine-readable clarity.
2. 2.Fluency score on timed tasks — are pauses shrinking and landing between ideas?
3. 3.Dictation accuracy — the integrated tasks' ceiling.

If pronunciation is flat while fluency climbs, you're rehearsing templates on top of the same sound errors; go back to phoneme drills. If dictation is the laggard, shift ten minutes from speaking reps to listening.

One honest note: sayit's scores are a training signal, not an ETS prediction. Before test day, do at least one official ETS practice set (TOEFL Practice Online) so the interface and pacing hold no surprises, and if you can, have a teacher or strong speaker sanity-check one full set of responses for topic development — the content half that machines still judge worst.

Start with one timed task today

You'll learn more from one honestly-scored 45-second response than from another hour of tips. Open sayit free in your browser — no card — do a single TOEFL-style Task 1, and look at the heatmap. The sounds it flags are your prep plan.