AI & Data

Real-time speech feedback, strict privacy, minimal retention, measurable accuracy.

Models and signals

• Automatic Speech Recognition (ASR): high-accuracy speech-to-text tuned for Spanish and English first.

• Pronunciation: phoneme alignment and GOP-style scoring, prosody checks for stress and rhythm.

• Phrasing: natural rephrases with one correction per turn, then a follow-up question to keep you speaking.

• Mastery tracking: spaced-repetition prompts and adaptive difficulty tied to CEFR outcomes.

Data handling

• Minimize: store only what we need. Raw audio is short-lived, derived features persist.

• Encrypt: data in transit and at rest. Secrets managed in Key Management Service (KMS).

• Separate: PII and learning data live in different stores with separate access paths.

• Opt-in: users can allow anonymized samples to improve models. Default is off.

• Retention: clear timelines for deletion, with self-serve wipe on request.

Quality and fairness

• Targets: publish per-accent word error rate targets and review drift monthly.

• Bias checks: evaluate across genders, accents, and device types, then correct defects.

• Human review: high-stakes assessments can trigger human-in-the-loop verification.