SpeechRater and e-rater score your TOEFL before a human ever sees it.
ETS made SpeechRater and e-rater the first-pass scorer for every 2026 TOEFL Speaking and Writing response. Knowing what each engine listens for, and what it ignores, is the difference between a 22 and a 26.
By Michael Buckhoff · 30 years grading TOEFL Speaking and Writing · Cross-checked against ETS research on SpeechRater ›
Last verified: 22 May 2026 · Updated when ETS publishes a new technical briefSpeechRater
SpeechRater is an acoustic and linguistic engine. It listens to your Speaking response as audio, runs speech recognition to produce a transcript, then extracts hundreds of features across delivery, language use, and topic development. The features feed a model trained against thousands of human-rated samples.
SpeechRater does not understand the meaning of your words the way a human does. It scores the signals that statistically correlate with human scores: how often you pause, how varied your vocabulary is, whether your stress patterns match native rhythm, whether your grammar holds together.
What SpeechRater rewards
- Steady speech rate (140 to 170 words per minute is the sweet spot)
- Pauses between thought groups, not mid-phrase
- Diverse vocabulary, including academic word families
- Clear consonant articulation and final sounds
- Natural prosody (rising/falling intonation matching content)
- Specific examples and concrete nouns over abstractions
What SpeechRater penalizes
- Long mid-sentence pauses or filler words ("uh," "um," "you know")
- Memorized template phrases ("There are several reasons")
- Monotone or flat intonation
- Repetition of the same vocabulary throughout
- Slurred consonants, dropped final sounds
- Speaking too fast (over 200 wpm) or too slow (under 100 wpm)
e-rater
E-rater is a natural language processing engine. It parses your Writing response sentence by sentence, identifies grammatical structures, scores word choice and academic register, measures organization through paragraph cohesion, and evaluates how well you develop your ideas.
Like SpeechRater, e-rater does not "understand" the topic. It scores how the writing looks structurally and statistically against the patterns of high-scoring essays.
What e-rater rewards
- Varied sentence structures (mix simple, compound, complex)
- Strong topic sentences with clear thesis
- Explicit transitions between paragraphs
- Specific examples and concrete details
- Correct subject-verb agreement and tense consistency
- Academic vocabulary used naturally (not forced)
What e-rater penalizes
- Template phrases ("In my opinion," "To begin with")
- Misspelled common words, dropped articles
- Sentence fragments and run-ons
- Repetitive sentence structure
- One-paragraph responses without paragraph breaks
- Verbatim copying from the source text in Integrated Writing
The 2026 shift: engine-first scoring
2 of 2Both SpeechRater and e-rater are now the first-pass scorer for every Speaking and Writing response. A human rater only reviews the response if the engine flags it as unusual (high confidence rating below threshold, mismatch with task prompt, suspected memorized response). The change took effect with the January 21, 2026 test format update.
Before 2026, human raters scored every response and AI was a secondary check. After 2026, the engine has first authority. The human rater sees only flagged responses. If your sample never gets flagged, your score is whatever the engine assigned, and there is no second opinion.
What this means for your prep
- Stop relying on memorized templates. Both engines were trained on patterns of high-scoring essays. The template phrases that get recommended on YouTube were the first signals ETS taught the engine to flag. The Free Speaking Course drills original phrasing instead.
- Practice with a recording, not just notes. SpeechRater listens to delivery, not just content. You cannot tell from notes whether your pauses sound natural or whether you sound monotone. The 5 free practice test banks include recording prompts for every Speaking task.
- Build vocabulary diversity, not vocabulary volume. The engines reward using a wider range of words across your response, not stuffing in obscure ones. A natural mix of academic and everyday vocabulary scores higher than a thesaurus-loaded response.
- Write in real paragraphs. E-rater rewards paragraph structure. Single-block responses lose organization points even if the content is solid. See the Free Writing Course for paragraph structure templates that score high.
Related reading: The full 2026 TOEFL rubric breakdown and which UK universities stopped accepting the new TOEFL.
Frequently asked
Does AI grade the TOEFL Speaking section?
Yes. SpeechRater, an automated speech scoring engine developed by ETS, listens to every TOEFL Speaking response. It assigns a score based on delivery, language use, and topic development. A human rater only reviews the response if SpeechRater flags it as unusual.
Does AI grade the TOEFL Writing section?
Yes. E-rater, an automated essay scoring engine developed by ETS, reads every TOEFL Writing response. It scores grammar, usage, mechanics, style, organization, and topic development. Human raters review only the responses e-rater flags.
Can I trick SpeechRater or e-rater with memorized templates?
No. Both engines flag template phrases like "There are several reasons," "In my opinion," and "To begin with." In the 2026 rubric, templated language costs you points rather than helping you. The engines were specifically trained to detect repeated patterns from test-prep materials.
Will SpeechRater penalize my accent?
SpeechRater is calibrated on a wide range of L2 English speakers and does not penalize accent itself. It penalizes intelligibility problems, slurred speech, and pronunciation errors that interfere with comprehension. A heavy accent that is still clear and accurate scores well. A mild accent with dropped consonants and unclear vowels does not.
What is the optimal speaking rate for SpeechRater?
Between 140 and 170 words per minute. Slower than 100 wpm signals hesitation. Faster than 200 wpm signals memorization or panic, and the engine struggles to extract clear features from rushed speech.
How does the engine handle Integrated Writing where I paraphrase the source?
E-rater compares your text against the source passage and penalizes verbatim copying. Light paraphrasing (changing word order without changing vocabulary) also gets flagged. Strong paraphrasing replaces both the words and the syntax of the original.
If I disagree with my score, can I request a human review?
Yes. ETS offers a Score Review service. You submit a request within 30 days of receiving your score, pay a fee, and ETS has a human rater re-score the response. The new score may go up, stay the same, or go down. You commit to whatever the human rater assigns.
Want to know how SpeechRater would score your Speaking sample?
Send Michael a 60-second sample ›