What Separates TOEFL Speaking Score Bands: Delivery

Methodology

Sample

55,614 answer records from 14,884 TOEFL 2026 Take an Interview practice sessions on LingoLeap; the headline analysis reports the 54,247 answers in the 2.0–4.5 AI score bands. The number of distinct learners is not recorded.

Period

From 1 January 2026. The end of the extraction window is not recorded in the data file.

Source

A read-only, aggregate-only analysis of LingoLeap's own practice records. No transcripts, recordings, user IDs or individual answers are published.

How scores were produced

Delivery timing and the delivery sub-scores (0–100) come from automatic speech assessment (Azure). The answer score (0–5) and the Language Use and Topic Development signals come from LingoLeap's AI rubric grader, mapped to the public TOEFL Speaking dimension names. None of this is ETS scoring.

Exclusions

Bands below 2.0 are excluded: very short or truncated answers produce speech-rate readings of 200–380 wpm from one or two words. The 5.0 band (n = 29) is too sparse to report. The speech-rate view keeps buckets with at least 100 answers, up to 180 wpm. 3,922 records are marked unattempted; whether any of them fall inside the reported bands is not recorded.

Limitations

Observational and descriptive: no significance tests or causal claims, and repeated learners and prompts are not independent. AI scores may differ systematically from human rating, and the three dimensions may correlate with each other.

Pattern 1: Every delivery signal rises with the score

Fluency, prosody, pronunciation, and accuracy all climb monotonically across bands — and fluency moves the most.

<figure style="margin:0;max-width:660px">
  <img src="https://lingoleap.ai/research/charts/toefl-speaking-fluency-delivery.svg" alt="Figure 1. Delivery sub-scores by AI score band" width="660" height="360" style="max-width:100%;height:auto" loading="lazy">
  <figcaption>Source: <a href="https://lingoleap.ai/research/toefl-speaking-fluency">LingoLeap, What Separates TOEFL Speaking Score Bands: Delivery</a></figcaption>
</figure>

Of the four, fluency has the widest spread across the reported bands (about 50→81). By contrast, prosody moves from 67→81 and accuracy from 70→91. This compares band averages; it does not show that fluency has a stronger causal effect on the score.

Pattern 2: Pace and length rise with the score too

<figure style="margin:0;max-width:660px">
  <img src="https://lingoleap.ai/research/charts/toefl-speaking-fluency-speech-rate.svg" alt="Figure 2. Speech rate by AI score band" width="660" height="320" style="max-width:100%;height:auto" loading="lazy">
  <figcaption>Source: <a href="https://lingoleap.ai/research/toefl-speaking-fluency">LingoLeap, What Separates TOEFL Speaking Score Bands: Delivery</a></figcaption>
</figure>

<figure style="margin:0;max-width:660px">
  <img src="https://lingoleap.ai/research/charts/toefl-speaking-fluency-word-count.svg" alt="Figure 3. Answer length by AI score band" width="660" height="320" style="max-width:100%;height:auto" loading="lazy">
  <figcaption>Source: <a href="https://lingoleap.ai/research/toefl-speaking-fluency">LingoLeap, What Separates TOEFL Speaking Score Bands: Delivery</a></figcaption>
</figure>

Speech rate and average score are not linearly related. After bucketing answers by speech rate, the 140–159 wpm bucket has the highest average (about 3.5), with lower averages in faster buckets. Length, clarity, proficiency, and recognition quality may confound this pattern, so it is not a recommended target speed.

<figure style="margin:0;max-width:660px">
  <img src="https://lingoleap.ai/research/charts/toefl-speaking-fluency-speech-rate-score.svg" alt="Figure 4. Score by speech-rate bucket" width="660" height="300" style="max-width:100%;height:auto" loading="lazy">
  <figcaption>Source: <a href="https://lingoleap.ai/research/toefl-speaking-fluency">LingoLeap, What Separates TOEFL Speaking Score Bands: Delivery</a></figcaption>
</figure>

Longer-response buckets also have higher average scores, but this does not establish an ideal length or causal effect; completeness, proficiency, prompt difficulty, and other factors may affect both length and score.

Pattern 3: The pause myth — zero pauses is the lowest, not the highest

Pause count is not “fewer is better”: the zero-pause bucket has the lowest average score, but it is substantially confounded by short and truncated responses.

<figure style="margin:0;max-width:660px">
  <img src="https://lingoleap.ai/research/charts/toefl-speaking-fluency-pauses.svg" alt="Figure 5. Score by pause count" width="660" height="300" style="max-width:100%;height:auto" loading="lazy">
  <figcaption>Source: <a href="https://lingoleap.ai/research/toefl-speaking-fluency">LingoLeap, What Separates TOEFL Speaking Score Bands: Delivery</a></figcaption>
</figure>

This is a textbook confound. Raw pause count is entangled with length — no pauses usually means barely any speech. So we don’t frame “pause less” as advice; we describe delivery with fluency, speech rate, and length, and keep pause count as its own honest sub-finding.

Turn the finding into practice

This observational dataset cannot prescribe one practice strategy, but it does not support treating ‘go faster, don’t pause’ as a scoring shortcut. Practice should consider completeness, clarity, and individual feedback together.

What this study cannot show

• Scores are LingoLeap AI scores, not official ETS results; dimension names borrow the public framework without reproducing ETS rubric text.
• Observational, descriptive data: denominators and differences, no significance tests or causal claims.
• Repeated learners and prompts are not independent observations.
• Bands below 2.0 are excluded for speech-rate artefacts; speech rate uses a minimum-timed-words guard.

Data

How to cite

Questions we hear from learners