The research, published in Discover Education, was conducted by Yang Zhang and Ying Zhou of Kaifeng Vocational College of Culture and Arts in Henan Province. The pair set out to test whether AI-assisted speaking practice could improve oral English among first-year students in higher vocational education, a sector in China where English instruction has traditionally emphasized reading, writing, and examination performance over genuine communicative ability. Their quasi-experimental study involved 92 students aged 18 to 20, all enrolled in a compulsory College English course and all with at least six years of prior English study behind them.

The design was straightforward but carefully controlled. Two intact classes of 46 students each were compared. Both groups received the standard four hours per week of conventional College English instruction from the same teacher, ensuring that any difference in outcomes could not be attributed to differences in classroom teaching. The experimental group, however, added three thirty-minute sessions per week of AI-assisted speaking practice using iFLYTEK Spark, version 3.5, a generative AI platform developed by one of China’s leading speech technology companies. The control group spent the equivalent additional time on conventional pair-work and group speaking activities covering the same textbook topics at comparable task complexity.

The technology at the heart of the intervention is more sophisticated than a simple chatbot. iFLYTEK Spark integrates automatic speech recognition, natural language processing, and generative dialogue capabilities. For speaking practice, it offers phoneme-level pronunciation assessment with visual feedback on individual sounds, intonation, and stress; real-time scoring aligned with IELTS-style criteria on a zero-to-nine scale covering pronunciation, vocabulary, grammar, fluency, and coherence; interactive AI-generated dialogue for open-ended conversation, role-play, and topic-based discussion; and detailed feedback reports pinpointing specific areas for improvement. Each practice session followed a three-phase cycle: a five-minute warm-up dialogue with the AI on topics related to the current unit, a fifteen-minute block of guided role-play and one-minute monologues with immediate feedback, and a ten-minute self-reflection phase in which students reviewed their AI-generated reports and re-recorded weak segments.

At the end of the twelve weeks, students’ speaking was assessed through a standardized task consisting of a two-minute monologue on a randomly assigned topic and a one-minute interactive exchange with an examiner. Three trained English teachers, each with at least five years of experience, rated the audio recordings independently and blind to group assignment, using an analytic rubric covering pronunciation, language use, fluency and coherence, and interaction. Inter-rater reliability, measured with intraclass correlation coefficients, ranged from 0.810 to 0.926, indicating good to excellent agreement. The results were striking: the experimental group achieved a significantly higher mean overall posttest score of 3.69 compared with 3.19 for the control group, a difference that was statistically significant with a medium-to-large effect size of Hedges’ g equal to 0.63.

Perhaps the most intriguing finding lay in the pattern across subdimensions. The largest effect appeared in interaction, the ability to engage in meaningful conversational exchange, where the effect size reached 0.75 and remained significant even after conservative Bonferroni correction for multiple comparisons. Differences in language use and in fluency and coherence showed medium effect sizes of 0.49 and 0.50 respectively, though these did not survive the strictest statistical threshold and should be treated as suggestive. Pronunciation, by contrast, showed only a smaller and non-significant difference of 0.38. The authors suggest that segmental pronunciation accuracy may require more specialized and sustained phonological training than general-purpose AI conversation tools typically provide, or that twelve weeks was simply too short for such gains to emerge, or that students prioritized communicative fluency over phonological precision during practice.

The theoretical scaffolding behind the study helps explain why the intervention worked. Swain’s Output Hypothesis holds that language production, not merely comprehension, drives second language acquisition by forcing learners to process language deeply and notice gaps in their own competence. The AI sessions demanded substantial pushed output across varied tasks. Krashen’s Affective Filter Hypothesis proposes that anxiety can block acquisition, and the AI’s non-judgmental, private practice environment appears to have lowered that barrier. Long’s Interaction Hypothesis emphasizes negotiation of meaning, a function the AI’s conversational role-play partially fulfilled by prompting students to clarify, rephrase, and respond in real time. The pronounced interaction advantage observed in the data aligns closely with this third mechanism.

Interviews conducted after the intervention with one teacher and three students added a human dimension to the statistics. All three students reported that the AI dramatically increased their opportunities for spoken practice, offering unlimited attempts without the time constraints and turn-taking limits of classroom pair work. Two students said they felt less embarrassed making mistakes in front of the machine because there was no fear of social judgment, with one describing the AI as a patient practice partner that never showed frustration. The teacher observed that students who were typically reluctant to speak in class became more willing participants after gaining confidence through AI sessions. Yet the participants also flagged limitations: the AI’s conversational responses sometimes felt mechanical or repetitive, one student preferred combining AI practice with human interaction, and the teacher stressed that continued teacher involvement remained essential for guiding reflection and correcting persistent errors the AI failed to address.

The authors are candid about the study’s constraints. Because intact classes rather than randomly assigned individuals were used, definitive causal claims are not possible. Baseline comparability relied on a coarse institutional proficiency classification rather than a dedicated standardized speaking pretest, so pre-existing differences in oral skill cannot be fully ruled out. The sample came from a single college in Henan Province, the qualitative sample was small, and a novelty effect from using exciting new technology cannot be excluded. The posttest-only design means the findings should be read as group differences under approximate baseline comparability rather than demonstrated learning gains.

Even with those caveats, the study offers a replicable template for educators: a structured cycle of warm-up dialogue, guided practice, and self-reflection, layered on top of conventional instruction rather than replacing it. As generative AI tools with speech recognition spread through education systems worldwide, the evidence suggests their greatest value may lie not in replacing teachers but in multiplying opportunities for the one activity that language learning cannot do without: speaking, freely and often, without fear.

Subject of Research: Effects of AI-assisted speaking practice on spoken English performance in higher vocational education

Article Title: Effects of AI-assisted speaking practice on spoken English performance in higher vocational education

Article References: Effects of AI-assisted speaking practice on spoken English performance in higher vocational education. (n.d.). https://doi.org/10.1007/s44217-026-02260-9

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02260-9

Keywords: artificial intelligence, spoken English, higher vocational education, iFLYTEK Spark, generative AI, speech recognition, language learning, quasi-experimental study, speaking anxiety, second language acquisition, China, educational technology