A Survey of Automated Presentation Coaching: Systems, Methods, and Open Challenges

Technical Summary: A Survey of Automated Presentation Coaching

Problem Statement
Automated coaching for oral presentations operates at the intersection of Computer-Assisted Pronunciation Training (CAPT), prosody modeling, and speech synthesis. Despite significant progress in these underlying technologies over the last two decades, the field lacks a unified framework to systematically survey, categorize, and compare existing systems. Current research is fragmented, often addressing isolated aspects of speaking (e.g., segmental accuracy or general fluency) without a holistic view of the specific demands of technical and academic presentations. Furthermore, there is a disconnect between research prototypes and industry deployment, and a scarcity of annotated corpora that capture the long-form, slide-aligned nature of presentations with diverse L1 backgrounds.

Methodology
The authors conducted a systematic literature search across major databases (ACL Anthology, IEEE Xplore, ISCA Archive, Google Scholar, Semantic Scholar) covering publications from 1997 to 2025. The search utilized keywords such as "presentation coaching," "pronunciation training," "CAPT," and "prosody assessment." The review focused on 15 representative systems and approximately 50 supporting references, excluding studies that addressed only visual/gestural coaching without speech components or general-purpose ASR/TTS papers.

The core methodological contribution is the construction of a five-dimensional task taxonomy designed to organize presentation skills based on the nature of feedback required and the methods available for assessment:

1. Pronunciation (Segmental): Phone- and word-level correctness, including technical terminology.
2. Lexical Stress: Syllable prominence in multi-syllabic words.
3. Prosody (Suprasegmental): Intonation (F0), phrasing, rhythm, and intensity contours.
4. Pacing: Words-per-minute (WPM), articulation rate, and pause placement relative to punctuation and slide boundaries.
5. Content Faithfulness: Coverage of key content, flagging missing keywords or omissions per slide.

The survey maps existing systems onto this taxonomy, analyzes their input/output requirements (e.g., need for slide-aligned scripts, enrollment clips for voice cloning), and reviews the core technical methods they employ, specifically TTS-based exemplar generation and diagnostic approaches.

Key Contributions

1. Taxonomy and Mapping: The paper introduces the five-dimensional taxonomy and explicitly maps surveyed systems to it. This reveals significant coverage gaps: lexical stress and content faithfulness are the least-addressed dimensions. While many systems cover pronunciation and prosody, few address the specific structural needs of presentations (e.g., slide-aligned pacing).
2. Systematic Categorization: Existing systems are categorized into four groups:
   • CAPT-based pronunciation systems (focused on isolated words/short sentences).
   • Fluency and prosody coaches (focused on suprasegmental quality).
   • Multimodal trainers (integrating gaze, gesture, and speech).
   • Q&A and interaction coaches (leveraging LLMs for interactive practice).
     The analysis shows that no single existing system spans all five taxonomy dimensions. The most recent system, PresentCoach (Chen et al., 2025), covers four dimensions but still omits lexical stress.
3. Technical Review: The paper details the core methods underpinning these systems:
   • Exemplar Generation: Utilizing modern neural TTS (e.g., flow-matching models like F5-TTS, CosyVoice 2) for controllable rate, pause insertion, and zero-shot voice transfer to generate personalized reference audio.
   • Diagnostics: Reviewing methods including Goodness of Pronunciation (GOP) scores, CTC-based alignment, self-supervised representations (wav2vec 2.0, HuBERT), and "clone-and-compare" techniques that synthesize an idealized version of the learner's voice to isolate prosodic errors.
4. Evaluation Framework: The authors consolidate evaluation metrics aligned with the five dimensions (e.g., Phone/Word F1 for segmental, Log-F0 RMSE for prosody, WER for faithfulness) and identify the lack of a standardized benchmark. They note that no existing corpus simultaneously provides L2 speech, long-form presentation structure, accent labels, and prosodic annotations.

Results and Findings

• Coverage Gaps: Lexical stress is almost entirely neglected by existing systems despite its documented impact on L2 comprehensibility. Content faithfulness is also under-addressed, with few systems checking for missing domain-specific keywords against slide content.
• Real-Time Limitations: Real-time feedback is rare; only two of the 15 surveyed systems provide in-session feedback. Most systems operate post-hoc.
• Data Scarcity: Major corpora (e.g., LibriSpeech, TED-LIUM) are not designed for L2 coaching or lack slide structure. Newer resources like Speak & Improve offer spontaneous L2 speech but lack presentation-specific annotations.
• Research-Industry Gap: Commercial platforms (e.g., ELSA Speak, Yoodli) typically address only one or two dimensions and rarely publish technical details, making systematic comparison difficult. Academic research often fails to address deployment constraints like device heterogeneity and noise.

Significance and Claims
The paper claims that its primary significance lies in providing the first unified survey that catalogs and compares automated presentation coaching systems along a structured, multi-dimensional framework. By mapping systems to the taxonomy, the authors expose critical gaps in the current state of the art, specifically the lack of integrated systems that handle lexical stress and content faithfulness simultaneously with pronunciation and prosody.

The authors modestly claim that their work serves as a reference for researchers developing the next generation of integrated, evidence-based coaching systems. They emphasize that while individual components (diagnostics, TTS, prosody modeling) have matured, the challenge remains in integrating them into deployable, real-time systems that are fair across diverse L1 backgrounds and grounded in the specific structure of academic presentations. The paper calls for community efforts to create standardized benchmarks (e.g., presentation mini-sets with slide markers and diverse accents) and for closer collaboration between research and industry to bridge the gap between theoretical methods and practical deployment.