How AI Scores Poetry Recitation: Inside Real-Time Speech Evaluation Tech

Hello, this is the Zuoyebang technical team.

We have launched a new column “Tech Talk” that uses MG animation to explain technical knowledge.

Intelligent speech assessment relies on computer‑assisted language learning (CALL); the same principle powers English word‑reading scoring.

Evaluation dimensions include accuracy, fluency, completeness, prosody, and intonation, while human experts score subjectively based on experience; AI systems aim to learn continuously to approach expert scores.

Reliability is measured by Pearson correlation coefficient and kappa coefficient—higher values indicate closer alignment with human scoring.

Zuoyebang’s system extracts features such as Goodness of Pronunciation (GOP) scores, consonant quality, part‑of‑speech, tone, duration, and fluency, and uses a neural network model to predict scores with high consistency to human judges.

For poetry recitation, the text is processed as whole sentences rather than isolated words. A multi‑branch evaluation supports repeated reading, jump‑reading, and offers flexibility.

To provide real‑time feedback, the model is deployed cross‑platform via ONNX, pruned to under 10 MB, enabling smooth performance on mobile devices.

With insufficient data even the best model cannot learn enough; a patented technology (CN 113901992 A) enables automatic data feedback training for continual model iteration.

Zuoyebang – technology for people, achieving greatness.