BEA 2026 · 2026 · Workshop

Wen Liang et al. (Columbia University / Red Hat / RoleGaku) · → Paper · → Preprint · Demo: ✗ · Code: ✗

Surveys 15 automated presentation-coaching systems through a five-dimensional taxonomy (segmental pronunciation, lexical stress, suprasegmental prosody, pacing, content faithfulness), and reviews TTS-generated practice exemplars and voice-cloned personalized references as one of the two method families these systems rely on.

Problem

Coaching for oral presentations, especially for English as a Second Language (ESL) speakers, draws on computer-assisted pronunciation training (CAPT), prosody analysis and speech synthesis. Earlier systems each covered a few facets of presentation skill. No survey had catalogued them, mapped which speech dimensions each one addresses, or set out what remains unsolved. The authors also argue that CAPT systems work on isolated words or short read sentences rather than the semi-scripted, long-form speech of a talk. They add that these systems give post-hoc error lists with no practice loop or exemplar generation (§1, §4.1).

Scope and Coverage

The survey covers speech-based coaching for presentations, both L2-specific and general public-speaking tools. It excludes visual and gesture coaching unless it is combined with speech feedback, and it excludes general-purpose ASR, TTS and spoken language assessment literature. The search covered ACL Anthology, IEEE Xplore, ISCA Archive, Google Scholar and Semantic Scholar for peer-reviewed work from 1997 to 2025. The authors added citation chaining from key papers. The result is 15 representative systems and about 50 supporting references (§1).

Organising scheme. A five-dimensional taxonomy organises the paper (§3.1, Figure 1):

  • Segmental pronunciation: phone- and word-level correctness, assessed with Goodness of Pronunciation (GOP), CTC or self-supervised detectors.
  • Lexical stress: syllable prominence in multi-syllabic words.
  • Suprasegmental prosody: F0, phrasing, rhythm and intensity, scored as F0 RMSE and Pearson r against a reference.
  • Pacing: words per minute (WPM), articulation rate and pause placement relative to punctuation and slide boundaries.
  • Content faithfulness: keyword coverage per slide, using glossary-biased ASR.

Systems also operate in three modes (§3.3): exemplar generation (controllable target reads), comparative assessment (aligning learner audio to a reference), and a mixed-initiative practice loop that turns diagnoses into drills. Surveyed systems fall into four categories (§4): CAPT-based pronunciation systems, fluency and prosody coaches, multimodal trainers, and Q&A or interaction coaches built on LLMs. Table 1 maps all 15 systems onto the five dimensions, real-time feedback, and L2 targeting.

Which parts concern speech synthesis. Most of the survey is about pronunciation assessment and multimodal trainers, not generation. The speech-synthesis content sits in four places:

  • TTS background (§2): flow-matching TTS models (the survey names F5-TTS, Voicebox and CosyVoice 2) are presented as fast enough and controllable enough (rate, pauses, emphasis) to generate coaching exemplars. Zero-shot style transfer from short enrollment clips is presented as the route to personalized references.
  • TTS-based exemplar generation (§5.1): the two controls the authors single out are tempo bands that constrain WPM and emphasis marks that elicit pitch accents on key terms. A generalized workflow renders an anchor exemplar at 120-140 WPM and an optional target at 150-170 WPM. It records the learner in 5-12 s chunks, aligns with CTC or DTW, and surfaces drills.
  • Clone-and-compare (§5.2.2): an idealized rendition is synthesized in the learner’s own voice with voice conversion or zero-shot TTS. MFCC or SSL distance curves, aligned with monotonic DTW, then flag likely mispronunciations while holding timbre constant. On the synthesis side, the cited generative spoken language model work simulates and modulates foreign-accented prosody to produce corrected exemplars that keep the learner’s vocal identity (§4.2).
  • Prosody coaching and latency: prosody exemplars feed the prosody and pacing dimensions. Appendix A sets a target of under 200 ms TTS rendering for 5-8 s of text within a sub-second per-chunk feedback budget.

Pronunciation assessment (GOP, CTC, self-supervised detectors, neural phonetic posteriorgrams), fluency scoring, content-faithfulness ASR, and multimodal gaze and gesture trainers make up the rest of the coverage.

Method

There is no original model or experiment. The contribution is the literature search, the taxonomy, a system-by-dimension comparison (Table 1), and a corpus suitability table (Table 2). Table 2 scores ten corpora, including GigaSpeech and Common Voice, for L2 speakers, long-form speech, accent labels and prosody annotations. The survey also gives a metric framework (Table 3) that pairs existing metrics with each dimension: phone/word F1 and UAR for segmental accuracy, log-F0 RMSE and Pearson r for prosody, WPM deviation and pause rate for pacing, WER and glossary hit-rate for faithfulness, and MOS and PESQ as perceptual checks. The authors state that the individual metrics are not new; only their alignment to the taxonomy is (§6). Appendix A sketches a seven-stage coaching pipeline: script and glossary ingest, exemplar synthesis, chunked recording, alignment, diagnostics, drill generation, and progress tracking.

Key Results

The findings come from the survey’s own mapping in Table 1 (§4.5):

  • Lexical stress is the least-covered dimension: only one of the 15 systems targets stress detection.
  • Real-time feedback is rare: two of 15 systems give in-session feedback; the rest assess post-hoc.
  • Content faithfulness and pacing are usually handled separately rather than with slide-aligned structure.
  • No system covers all five dimensions. PresentCoach covers four and omits lexical stress. The survey describes it as the first system to integrate slide-aware exemplar generation (with personalized voice synthesis) and multi-dimensional assessment (§4.4).

Table 2 finds no corpus that combines L2 speech, long-form presentation structure, accent labels and prosodic annotation (§6). The authors propose “presentation mini-sets” of 5-10 minute talks from 20-30 speakers across 5+ L1 backgrounds, with slide boundaries, phone-level transcriptions and expert prosody ratings (§7).

For speech generation specifically, the survey reports no measurement. The statements that modern TTS achieves sub-second latency and fine-grained rate, pause and emphasis control are asserted from the cited model papers. The survey does not test them in a coaching setting (§2, §5.1.1).

Novelty Assessment

The new element is the five-dimensional taxonomy and its explicit mapping of systems to dimensions, which exposes coverage gaps. The method reviews are short. The exemplar-generation workflow is presented as a generalization of patterns in prior systems and shadowing pedagogy, not as a new design. The metric framework reuses established metrics. The speech-synthesis material is a brief statement of how existing TTS and voice-cloning capabilities fit into a coaching loop; it does not compare TTS systems.

Field Significance

low — For speech generation, the survey gives an application framing: controllable and voice-cloned TTS as a source of practice targets and personalized references in language learning. It also lists the controls (tempo bands, emphasis) and latency budgets such an application would need. It provides no comparative evidence about synthesis quality or controllability. Its main value lies in pronunciation assessment and educational technology rather than generation.

Claims

  • supports: Controllable TTS can supply reference exemplars for pronunciation and prosody practice, with speaking rate and emphasis as the controls that matter most for this use.

    Evidence: The survey identifies tempo bands constraining WPM and emphasis marks eliciting pitch accents as the two most relevant controls. Its generalized workflow renders an anchor exemplar at 120-140 WPM and a faster target at 150-170 WPM for shadowing practice. (§5.1.1, §5.1.2)

  • complicates: A single synthesized reading used as the correct target can over-constrain learner style, and synthetic emphasis may not match domain conventions.

    Evidence: The survey lists both as known limitations of TTS-based coaching pipelines. It also recommends that exemplars be short (under 12 s) and offered in multiple speed bands. (§5.1.2)

  • complicates: Using voice-cloned speech in the learner’s own voice as a pronunciation reference has not been validated against expert references.

    Evidence: Clone-and-compare synthesizes an idealized rendition via voice conversion or zero-shot TTS and flags mispronunciations from DTW-aligned MFCC/SSL distance peaks. The authors state that they provide no empirical comparison of clone-and-compare against expert references, and they propose such a comparison against expert annotations as a needed validation protocol. (§5.2.2, §6, §Limitations)

  • refines: Interactive speech-coaching applications set a tight per-utterance latency budget for synthesis, within a sub-second end-to-end feedback loop.

    Evidence: The appendix targets TTS rendering under 200 ms for 5-8 s of text, alignment and metric computation within 300 ms, and UI updates under 200 ms. Voice cloning is deferred to an offline enrollment step. These are design targets, not measured results. (§A)

Limitations and Open Questions

  • The speech-generation coverage is thin and descriptive. TTS capabilities such as latency and rate, pause and emphasis control are taken from the model papers without evaluation in a coaching setting. The survey also groups CosyVoice 2 with non-autoregressive flow-matching models, although CosyVoice 2 pairs an autoregressive LM with a flow-matching decoder (§2).
  • The authors note that the survey is not exhaustive, focuses on English, derives its taxonomy and metrics from existing literature, and gives no empirical comparison of clone-and-compare against expert references (§Limitations).
  • The proposed shadowing-efficacy test, a randomized crossover of TTS-guided against unguided rehearsal, remains a recommendation. Whether TTS exemplars actually improve learner outcomes on the five dimensions is not established (§6).
  • Open problems listed: lexical stress and discourse-level prosody planning conditioned on slide structure, few-shot personalization of pacing and prosody targets from a 2-3 minute calibration recording, accent-fair thresholds, grounding LLM feedback in acoustic evidence, and languages beyond English (§7).
  • Ethics: voice cloning for personalized references needs explicit consent, encrypted embeddings, deletion mechanisms and a non-clone baseline for users who opt out (§Ethics Statement).

Wiki Connections

  • Zero-Shot TTS — Surveys clone-and-compare coaching, where a learner hears a reference rendered in their own cloned voice from a short enrollment clip; the survey describes the idea but reviews or compares no zero-shot cloning methods.
  • F5-TTS — Cited as the example flow-matching TTS model for exemplar generation and as the basis for the appendix’s sub-200 ms rendering target.
  • CosyVoice 2 — Cited as a streaming TTS architecture with sub-second latency for on-the-fly exemplar rendering during rehearsal.
  • FastSpeech 2 — Cited as part of the neural TTS progress that motivates automated presentation coaching.
  • Whisper — Cited as an end-to-end ASR source of word durations and confidence for diagnostics, and for WER-based content-faithfulness scoring.
  • EnCodec — Cited as a building block for the sub-second diagnostics that real-time coaching would need.
  • GigaSpeech — Assessed in Table 2 as long-form but not L2-specific, so unsuitable on its own for presentation coaching.
  • Common Voice — Assessed in Table 2 as having accent labels but only short read sentences.