Help Center/ Teaching & Practice/ How does Amira handle accents or articulation challenges?

How does Amira handle accents or articulation challenges?

3 min read · Updated 2026-09-29



Amira is designed to detect reading skill, not accent, dialect, or developmental speech differences. That principle shapes the models’ architecture, the training data, the scoring rules, and how accuracy is measured. Here’s what that means in practice.

 

Purpose-built specifically to understanding children’s reading, in real classrooms
Most commercial speech recognition (the technology behind Alexa, Siri, and similar tools) was built on adult speech, and it works by mapping what it hears to the nearest word in adult language. For a child with an accent or who is disfluent at reading, those commercial systems can sometimes force their accents or disfluencies into “recognizing” the wrong word.

Amira works differently. Our speech recognition models are built entirely in house and trained on children’s voices reading in real classrooms. The primary model works at the phoneme level, meaning it transcribes the actual sounds the child produces, including sound-outs, partial attempts, and accented pronunciations, rather than guessing at the intended adult word. A pronunciation that reflects a regional dialect or a home language stays intact instead of being misheard as a different word.

Trained on a representative range of voices
Amira’s models are trained on more than 10,000 hours of audio from students reading in real, noisy classrooms across all 50 states and several countries, in both English and Spanish. Training sets are deliberately built for representation across 12 to 13 distinct accent and dialect groups, so no group’s speech patterns are underrepresented in what the models learn from.

Scoring rules that protect developmental speech
The human experts who label Amira’s training data score against a detailed written rubric, and that rubric explicitly codifies developmental speech patterns, such as substituting /w/ for /r/ or /l/, as well as speech patterns found in over a dozen common accents and dialects, that must not be counted as reading errors. Because the models learn from these labels, they learn not to penalize these patterns either.

Measured, not assumed
Fairness claims are only as good as their measurement. Amira’s accuracy is evaluated separately for each accent and dialect group and each demographic subgroup, with performance exceeding 95% per subgroup and no meaningful gaps across gender, race/ethnicity, or English learner status. At the item level, Amira’s agreement with expert human scorers meets the benchmark for how often two trained humans agree with each other (roughly 96 to 97%). A whitepaper summarizing our equity studies, including district-level subgroup analyses, is available to districts on request.

Model versions are frozen within a school year
Amira Assess models are versioned. The model version deployed at the start of a school year is the model every student is assessed by for that entire year. It does not change in response to student use, teacher corrections, or elapsed time. Improvements are released once a year, and each release is validated against human-scored benchmarks before deployment, with published norms and psychometrics tied to that specific version. Amira Tutor uses a separate set of models with its own release schedule.

Human experts are in the loop
When a student’s production is ambiguous, Amira doesn’t automatically score it as an error; the scoring model weighs the audio against what it knows about the word and the student’s skills. Sessions the model can’t score with confidence receive an “Under Review” status and are checked/rescored by a person on our internal team, typically within 72 hours. Teachers can also replay any recording and correct any score. Those corrections are collected and reviewed by our team, and are used as candidate training data for a future model version. They never alter the model a student is being assessed by.

Students receiving speech support
Amira is in use today in schools serving students with speech and language disabilities. We are transparent about the limits, though. Speech patterns that require a specialist to interpret may also be difficult for Amira, and teacher or specialist review and re-scoring matters most for these students.

If a score looks wrong
Listen to the recording and rescore it if needed. Your correction is applied to that student’s score immediately, and it is also reviewed by our team as candidate training data for the next annual model release. If scoring of a specific passage seems significantly off (errors are made on more than 5% of the words in the passage) and you’d like our team to investigate, email support@amiralearning.com with the district, school, teacher, student name, and story title(s).