Steven Hé (Sīchàng)

speech transcription and speakers (authored by agents unless marked 🧑)

takeaway

• recommendation: build a local server baseline, then study how to spend extra computation on the parts most likely to be wrong
  • output the words, their times, speaker labels, and unresolved regions
  • measure completed recordings and human correction effort
  • an attractive average error rate can hide a recording that repeats forever or loses a speaker
• plausible systems question: can selective reprocessing improve complete-recording accuracy under a fixed GPU-time budget?
  • novelty is unconfirmed
  • chunk-size tuning, speaker memory, faster decoding, and diarization-based recognition already have close prior work
• evidence checked 7 Oct 2026
  • primary paper abstracts, selected full-paper methods/results/limitations, and official implementation documentation
  • recent arXiv results below are authors’ claims, not independently reproduced measurements
  • no transcription service implemented or audio benchmark run for this review

human scope 🧑

• “processes MP3 or similar files into speakers + accurate transcripts, designed to be run on the server”
  • supplied in the task delegation as the human’s exact request

first principles

• speech recognition answers what was said
• speaker diarization assigns anonymous speaker labels to time intervals
  • speaker identification additionally maps those labels to known people
  • inference: anonymous labels should remain anonymous without enrollment or human confirmation
• forced alignment places an existing transcript on the audio timeline
  • inference: a precise timestamp does not prove that a word was actually spoken
• voice activity detection finds speech intervals
  • inference: removing silence may reduce invented text but can also remove quiet speech
• overlapping speakers require more than one active label
  • inference: a single ordered text stream can lose one speaker’s words
• server requirements differ from a paper’s isolated inference measurement
  • decoding the MP3, queueing, loading models, reprocessing, and writing results all consume time
  • retain original audio and intermediate results so a failed stage can resume

established modular baselines

• Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision, 2022
  • full ICML 2023 paper, task description and appendix A
  • authors: “inaccurate transcription in one window may negatively impact transcription in the subsequent windows”
  • trains on thirty-second chunks; predicted times advance long-form windows
  • beam decoding and temperature fallback address repetitive or low-score output
    • token-score and compression thresholds trigger retries
    • preceding text conditions later windows under selected temperatures
    • silence filtering combines no-speech and token probabilities
  • these heuristics are not calibrated probabilities of transcript correctness
  • long-form evaluation covers seven datasets and historical commercial defaults
    • word errors do not measure speaker assignment or total server cost
  • OpenAI’s model card, performance and limitations: “predictions may include texts that are not actually spoken in the audio input”
  • the same card reports uneven performance across languages and accents
  • implication: fluent output needs recording-specific evaluation
    • charge fallback retries to the proposed repair budget
  • reading limit: selected full task and long-form methods inspected
    • model training and benchmark results not independently reproduced
• Bain et al., WhisperX: Time-Accurate Speech Transcription of Long-Form Audio, Interspeech 2023
  • full primary paper, §§2–3
  • authors: “without conditioning on previous text”
  • cuts at low voice activity between fifteen and thirty seconds, merges chunks up to thirty seconds, and batches recognition independently
  • phoneme alignment places recognized words in time
    • unsupported transcript phonemes inherit neighboring timestamps
    • translation cannot use this phonetic alignment
  • evaluates meetings, telephone conversations, TED talks, and videos
    • word timing scores require text agreement and timestamp overlap
    • they do not score speaker assignment
  • 11.8× transcription speed uses batch 32 on one A40 against full-audio batch 1
    • not a complete diarization or arbitrary-hardware speed guarantee
  • implication: batching and timing repair are established contributions
  • current repository, limitations: “Overlapping speech is not handled particularly well”
    • investigate overlap separately from ordinary recognition
    • repository performance claims depend on models, hardware, batching, and versions
  • reading limit: selected full methods and tables inspected
    • multilingual alignment examples lack quantitative evaluation
• Bredin, pyannote.audio 2.1 speaker diarization pipeline: principle, benchmark, and recipe, Interspeech 2023
  • abstract: “speaker segmentation applied to a short sliding window”
  • extracts a representation of each local voice, then groups similar voices across the recording
  • includes adaptation using manually annotated target data
  • implication: local detection plus global speaker grouping is an established baseline
• Plaquet and Bredin, Powerset multi-class cross entropy loss for neural speaker diarization, Interspeech 2023
  • full primary paper, §§1–5
  • authors: “overlapping speech regions inferred automatically from the first step are discarded before computing the embeddings”
  • five-second windows every 500 ms use three local speaker slots and seven silence/single/pair classes
    • excludes three-or-more simultaneous speakers
    • local slots do not identify people across windows
  • voice embeddings and clustering join slots across a recording
    • activity threshold is removed; clustering threshold remains tuned
  • segmentation selection and whole-recording tuning use development data
    • domain adaptation additionally uses target-domain training and development data
    • ordinary results use no boundary tolerance; shaded comparisons allow 250 ms
  • implication: overlap segmentation differs from recognizing simultaneous words and assigning each to a speaker
  • reading limit: selected full pipeline and evaluation methods inspected
    • artifacts and long-recording continuity not independently reproduced
• WhisperX speaker assignment, inspected revision
  • assigns at most one speaker to a timed word by maximum summed temporal overlap
    • words can remain unassigned
  • optional nearest-speaker fallback handles no overlap
  • inference: this join does not recover both speakers’ words during simultaneous speech
  • current code inspection, not an independently reproduced 2023 result
• pyannote.audio current repository, community-1 example: “runs locally”
  • model download requires accepting conditions and obtaining a Hugging Face token
  • its benchmark, marked September 2025, reports community-1 DER of 17.0% on AMI headset audio and 19.9% on AMI single distant microphone
  • Ego4D development DER is 46.8%
  • these are vendor-maintained comparisons, not a universal accuracy promise
  • premium precision-2 API and local community-1 have different execution locations
• SYSTRAN faster-whisper, performance-comparison instructions: “make sure to run the comparison with similar settings”
  • uses CTranslate2 to execute Whisper with batching and reduced numeric precision
  • documentation asks for matching decoding settings, transcript error rate, and CPU thread count
  • implication: a shorter, incomplete transcript can appear faster
• NVIDIA Canary-Qwen-2.5B model card, limitations: “The maximum audio duration in training was 40s”
  • English-only recognition alternative using a speech encoder and language-model decoder
  • model card warns that longer input may degrade accuracy
  • implication: a short-input leaderboard result does not establish whole-meeting reliability

joint and long-recording approaches

• Park et al., Sortformer, ICML 2025, full author manuscript, §5 and appendices A/D
  • authors: “we include only sessions with four or fewer speakers”
  • orders speaker channels by first arrival and combines ordering with best-permutation supervision
  • ninety-second diarizer training samples truncate long sessions and exclude more than four speakers
    • DIHARD evaluation also excludes those sessions
    • overlap is scored, but boundary tolerances differ by dataset
  • speaker activity supervises Canary-based recognition on ten-to-twenty-second training samples
    • some word times are approximated; boundary overlap exclusions apply
  • added runtime measurement uses synthetic three-speaker mixtures on one Ada GPU at batch 100
    • not a low-batch or arbitrary-length server guarantee
  • implication: local arrival ordering does not ensure speaker continuity across independently processed chunks
  • reading limit: selected full loss, integration, and evaluation methods inspected
    • later streaming releases and implementation not reproduced
• DiaPer: End-to-End Neural Diarization with Perceiver-Based Attractors, 2023
  • abstract: “finding the quantity of speakers in a conversation more accurately”
  • compares a compact end-to-end diarization model with cascaded methods across more than ten public wide-band datasets
  • speaker counting belongs in the evaluation, alongside total diarization error
• Wang et al., DiarizationLM: Speaker Diarization Post-Processing with Large Language Models, Interspeech 2024
  • abstract: “reduce the WDER by rel. 55.5% on the Fisher telephone conversation dataset”
  • combines recognition and diarization output, then uses a finetuned language model to revise speaker labels
  • WDER means word-level speaker-label error rate
  • claim is relative improvement on that dataset, not a 55.5 percentage-point gain
  • implication: text-based repair is prior work and may need domain-specific training
• Train Short, Infer Long, 2025 preprint
  • full primary methods, §§2.2, 3.3–3.4 and algorithm1
  • authors: “storing one utterance (audio clip and its transcription) for each previously observed speaker”
  • JEDIS-LLM prepends cached voice clips and speaker-attributed text to each new chunk
  • forced alignment selects long nonoverlapping clips up to five seconds per speaker
    • updates depend on predicted labels/text, timing, and a separate speaker verifier
    • incorrect cached labels can propagate
  • global evaluation uses telephone subsets and chunks up to ten seconds
    • compares VAD chunks with reference-derived sentence boundaries
    • excludes edge regions without reference transcripts
  • enrollment uses manually labeled clips from otherwise untranscribed audio
    • supplied identity evidence differs from anonymous speaker discovery
  • implication: cross-chunk speaker memory and enrollment already have direct prior work
  • reading limit: selected full cache-update and evaluation methods inspected
    • arbitrary multihour meeting continuity and prompt-growth cost not established
• Streaming Sortformer, 2025 primary paper, §§3–4 and tables1–2
  • authors: “declared latency values refer to the input buffer delay”
  • orders cached acoustic frames by speaker and joins them with recent context and incoming audio
    • compression favors recent, confidently single-speaker frames
  • fine-tunes sequential windows within ninety-second training samples
    • inference-only cache addition performs worse than cache-aware training
  • maximum four output speakers
    • evaluating five-plus-speaker subsets does not remove that architectural limit
  • declared buffer delays exclude computation
    • separate runtime factor uses batch1 on one RTX6000 Ada
  • implication: speaker-cache tracking is an existing baseline
  • reading limit: selected full cache, training, and evaluation methods inspected
    • later releases, artifact execution, and arbitrary-length speaker continuity not verified
• VIBEVOICE-ASR Technical Report, January 2026 preprint
  • abstract: “single-pass processing for up to 60 minutes of audio”
  • jointly produces words, speaker labels, and times
  • reports over 50 languages and context prompts for names and technical terms
  • Microsoft’s implementation supplies file inference and vLLM execution paths
  • inference: compare terminology prompts against both rare-word recovery and invented prompted words
• Lee, Kamahori, and Kasikci, MURMUR: An Efficient Inference System for Long-Form ASR, May 2026 preprint
  • limitation: all experiments are in English
  • combines intermediate-size chunks with dropping old cached model state
  • abstract reports 4.2× lower latency while matching single-pass AMI-IHM accuracy
  • table 1: “Failed clips due to repetition loops are excluded”
    • two of fourteen VibeVoice-ASR clips are excluded from that table
    • inference: reproduce failures as part of the service result, rather than comparing only successful clips
  • figure 3: “repetition loop failures counted as 100% error”
    • the paper also reports failure-inclusive comparisons
    • distinguish its fixed failure penalty from scoring actual repeated output and service completion
  • experiments use VibeVoice-ASR, AMI, TED-LIUM3, and Earnings21
  • implication: another chunk-size sweep alone is unlikely to establish novelty
• Grounding Spoken LLMs in Multi-Speaker Audio via Diarization Conditioning, June 2026 preprint
  • abstract: “condition the acoustic encoder on diarization masks”
  • Dixtral uses speaker-time information to focus recognition on one speaker
  • table 2 reports average cpWER of 15.4% for Dixtral and 14.0% for specialized DiCoW v3.3 across its listed conditions
  • benchmark includes distant AMI and NOTSOFAR-1, mixed LibriSpeech, and Mixer6
  • implication: supplying speaker intervals to recognition is an important stronger baseline
  • downstream questions partly use machine-generated labels
    • transcription evidence and downstream-question evidence need separate treatment

invented text and compute controls

• Calm-Whisper, May 2025 preprint
  • abstract: “over 80% reduction in non-speech hallucination”
  • selectively finetunes three decoder attention heads using non-speech examples
  • reported preservation of LibriSpeech accuracy does not establish preservation across languages or noisy meetings
  • inference: retain a separate non-speech test rather than hiding insertions in speech-heavy averages
• Whisper-CD, March 2026 preprint
  • abstract: “contrasts clean-audio logits against negative logits”
  • compares each proposed next word against predictions made from noise, silence, and shifted audio
  • reports improved WER across five English long-form datasets
  • table 1 contains baseline WER above 100%
    • full results section attributes this to repetition that generates more words than the reference
  • method adds three decoding paths
    • fewer repeated words can offset added work
    • turbo results still show extra runtime compared with its own unmodified baseline
  • implication: measure total GPU work and completion time, not token generation speed alone
• Efficient and Robust Speaker Diarization via Structured Pruning of Self-Supervised Models, June 2025 preprint
  • abstract: “up to 80% model size reduction and 4x faster inference”
  • reports tests across eight diarization datasets and CHiME-6 transfer
  • implication: model compression is an existing compute-saving alternative to selective reprocessing

evaluation that matches the requested output

• Open ASR Leaderboard, paper checked with results dated March 2026
  • abstract: “English short- and long-form and multilingual short-form tracks”
  • measures word error rate and audio duration divided by processing duration
  • long-form datasets include CORAAL interviews, Earnings21/22 calls, and TED-LIUM talks
  • short-form evaluation largely cuts audio to at most thirty seconds
  • implication: choose the matching track and preserve full recordings
• CHiME-8 DASR organizers
  • task description: “high variability in the duration: from few minutes to hours”
  • combines dinner parties, interviews, and office meetings
  • ranking uses scenario-wise mean tcpWER with a five-second time tolerance
  • its multichannel setup is richer than one uploaded MP3
    • compare the same input channels when claiming improvements
• DIHARD III, challenge paper
  • abstract: “two speech activity conditions”
  • contrasts supplied reference speech intervals with detecting speech from scratch across eleven domains
  • implication: measure failures in speech detection separately from voice grouping
• MeetEval toolkit, features: “Time-Constrained minimum-Permutation Word Error Rate (tcpWER)”
  • WER counts substituted, deleted, and inserted words divided by reference word count
  • cpWER groups words by speaker and chooses the best mapping between anonymous labels
  • tcpWER also restricts matching to nearby times
  • DER measures missed speech, false speech, and wrong-speaker time
  • pyannote.metrics reference: “Defaults to False (i.e. keep overlap regions)”
    • explicitly report excluded boundary intervals and whether overlapping speech is scored
    • metrics with different tolerance or overlap rules are not directly comparable
• recommendation: report recording-weighted completion and failure counts alongside aggregate WER, tcpWER, and DER
  • retain per-recording errors and separate overlap, silence, terminology, language, and speaker-count breakdowns
  • use one fixed text normalization and tokenization policy for every model
    • report Chinese character error rate separately from English WER
    • document how mixed-language words and characters are counted
  • count invented words per silent audio hour
    • ordinary WER has a zero denominator when the reference contains no words
  • report correction minutes per audio hour and name/number errors
  • fixed timeout prevents a repetition loop from consuming an unlimited service budget

closest work: confidence is not expected repair benefit

• Identifying and Calibrating Overconfidence in Noisy Speech Recognition, 2025 preprint, §§III–V
  • authors: “temperature scaling selectively”
  • learns a token-level overconfidence selector and a confidence adjustment
  • training uses short LibriTTS utterances with controlled speech-shaped noise
    • evaluation uses 400 R-SPIN sentences at multiple noise levels
    • calibration module is trained for the low-signal range
  • inputs include prediction probabilities, token features, and an acoustic summary
  • reports better confidence calibration under severe noise
    • this is not measured accuracy improvement from a second recognition pass
    • controlled noise is not overlapping speakers, long silence, or whole meetings
  • implication: probability calibration needs separate validation on the intended recording domain
    • a low average calibration error does not establish useful ranking of repair candidates
• Jitkrittum and colleagues, When Does Confidence-Based Cascade Deferral Suffice?, NeurIPS 2023, §§3–5
  • authors: “not modelling the errors of downstream models”
  • optimal deferral depends on the difference between the two models’ error probabilities
    • uncertainty of the first model alone is insufficient
  • experiments include specialist models, noisy labels, and distribution shifts
    • primary settings are classification, not full-recording speech repair
  • implication: identify regions the second pass can actually fix
    • equally bad models can agree or disagree without offering repair benefit
    • correcting confidence cannot create recoverable acoustic information
• A Unified Cascaded Encoder ASR Model for Dynamic Model Sizes, 2022, §§2–4
  • authors: “jointly training ASR models with different sizes”
  • shares encoder computation across differently sized recognition paths
  • uses separate decoders and joint training to trade quality against memory and computation
  • selects deployment configurations rather than demonstrating a per-region repair controller
  • direct baseline for adaptive resource use
    • do not assume independent small/large models are the cheapest way to construct a cascade

closest work: diarization confidence and active correction

• Plaquet and Bredin, On the calibration of powerset speaker diarization models, Interspeech 2024, §§2–3
  • authors: “oracle labeler”
  • examines confidence across training-like domains and eleven DIHARD domains
  • low-confidence five-second chunks have higher diarization error
  • simulates annotation by revealing withheld reference labels
    • compares low-confidence and random selection for training and validation
    • equal annotation budgets do not necessarily improve DER more with low-confidence training
    • confidence calibration and annotation-efficient validation can improve
  • important scope limit
    • local segmentation confidence is not confidence that global speaker clusters are correct
    • reference-label duration does not measure actual human correction time
  • direct prior work for choosing diarization regions using confidence
• Prokopalo and colleagues, Active correction for speaker diarization with human in the loop, IberSPEECH 2021, §§3–4
  • authors: “User simulation and correction module”
  • asks whether two clusters contain the same speaker
  • answers merge or split a hierarchical clustering tree
  • questions are ordered and stopped to limit intervention
  • ALLIES experiments simulate correct human answers using reference labels
    • includes an ideal-correction comparison
    • human interaction cost uses an estimated per-question time
  • limitation: perfect answers and estimated times do not establish real-user efficiency
  • direct prior work for selective speaker repair and workload-aware stopping
  • global cluster edits can improve many regions at once
    • count affected words and intervals, not only queried duration

closest work: real-user transcript correction under a time budget

• Sperber and colleagues, Transcribing Against Time, Speech Communication 2017, §§4–8
  • authors: “transcription is performed from scratch, but with the ASR hypothesis visible”
  • chooses segment locations and lengths using predicted errors and human time
  • updates a Gaussian-process time model and segment choices as the person works
    • accounts for the overhead of starting another segment
  • user study: twelve participants, four TED talks, twenty minutes per talk
    • alternates dynamic and static methods and shuffles talk assignments
    • static comparator uses the same initial time model without updates
    • cost-insensitive comparators appear in simulations, not this user comparison
  • reported mean productivity gain is 15% relative
    • the slower six-person subgroup gains 42%, with a significant subgroup difference
    • the overall simple-model difference is not statistically significant
  • directly establishes budgeted transcript selection and adaptive human-cost modeling as prior work
    • does not establish gains for automatic GPU reprocessing, overlapping meetings, or modern ASR post-editing
    • proposed human comparison should measure actual correction time and include a time-adaptive selector

research candidate: selective repair under a fixed budget

• hypothesis: combine cheap error indicators to select regions for a stronger second pass
  • indicators: disagreement between recognizers, low speech evidence, repeated text, speaker changes, overlap, and implausible times
  • target: expected error reduction per additional GPU second
    • predict second-pass benefit using separately labeled development recordings
    • compare with merely ranking first-pass error probability
  • these are candidate signals, not calibrated probabilities
• proposed controller chooses between accepting, retranscribing, expanding surrounding audio, and requesting correction
  • preserve accepted words and speaker assignments unless new acoustic evidence supports a change
  • compare against random selection, fixed overlap repair, all-region second pass, and a larger single model
    • enable and charge Whisper’s built-in fallback in every appropriate baseline
  • add confidence-only deferral, learned benefit prediction, and shared-computation cascades where available
  • label reference-informed best-region selection as an oracle upper bound
    • it is unavailable to the deployed controller
  • include Whisper-CD, speaker-conditioned recognition, JEDIS-style memory, and MURMUR-style chunking where implementable
• main result to seek: accuracy and correction effort versus total GPU seconds on held-out full recordings
  • set equal budgets including the first pass, detection, and failed retries
  • split by recording and speaker to reduce leakage
  • validate chosen thresholds on data separate from the final test
  • keep calibration, selection ranking, and actual repair benefit as separate measurements
  • score regressions introduced by the second pass
    • preserve speaker identities and timing outside the repaired region
    • local retranscription can break cross-chunk consistency
  • compare automatic repair and human correction using separate resource budgets
    • GPU seconds and person-minutes are different costs
• reject the research claim if a fixed rule or larger baseline achieves the same improvement within measurement uncertainty
  • useful engineering can remain even if novelty disappears

informative null results

• no selector beats random selection
  • likely explanation to test: available signals do not predict which errors the stronger pass can fix
  • report the oracle’s improvement to distinguish selector weakness from absent repair opportunity
• oracle has little advantage under the same budget
  • stronger pass offers too little recoverable accuracy or costs too much
  • a larger first pass or improved audio acquisition may be the better baseline
• confidence calibration improves while repair quality stays flat
  • calibrated uncertainty is useful for reporting uncertainty but not necessarily allocating computation
• DER improves without speaker-attributed word accuracy improving
  • corrected time intervals may contain recognition errors or few words
  • report both metrics rather than treating them as interchangeable
• simulated correction wins but human correction does not
  • investigate listening time, ambiguity, incorrect answers, and interface cost
  • a perfect-label simulation overestimates deployment benefit
• these results can narrow an engineering design
  • publication value still depends on systematic evidence across models and recording domains

small server pilot

• recommendation: begin with one queued worker using WhisperX/faster-whisper and local pyannote community-1
  • pin library versions and model revisions
  • decode uploaded files once and keep channel information
  • write structured segments with words, start/end seconds, anonymous speaker, model revision, and review flags
  • provide readable text and subtitles as derived outputs
  • retain explicit unknown-speaker and unintelligible regions
• evaluate a small collection of full lectures, interviews, and meetings before adding a second model
  • include long silence, interruptions, several similar voices, technical names, English/Chinese switching, and files longer than an hour
  • use human reference transcripts only for evaluation
  • compare long-context VibeVoice-ASR when hardware permits
• measure cold model loading, warm throughput, peak memory, queue delay, failure rate, and interrupted-job recovery
  • multi-GPU service scheduling becomes a research question only after real concurrent jobs reveal a bottleneck

adjacent question: listening to long documents

• Taming Long-form Text-to-Speech, September 2026 preprint

  • authors: “roll back to the error onset and regenerate with temporary guardrails”
  • reports a long-prompt reliability gap hidden by strong short-prompt scores
  • proposed monitoring checks omitted/repeated words and voice drift
  • recommendation: preserve document-to-audio section positions and verify content coverage
  • this is a separate study direction from recognizing uploaded speech
  • full methods and results, §§3–5
    • 144 call-center prompts, ten random seeds, three TTS variants
    • word-error scoring uses an external speech recognizer
    • monitors selected attention heads and retries only until a fixed cap
    • detector latency is measured in generated-audio seconds
      • this differs from wall-clock latency and required playback buffering
    • authors: “neither repairs the tail” for VoxCPM2 beyond 1,300 words
    • possible experiment: compare document section boundaries, coverage errors, and playback stalls at equal total compute
      • include ordinary chunking, always-on attention constraints, and retry controls
      • manually check content errors that the scoring recognizer may miss
    • read-depth limit: full primary methods and results read; code promised on publication and not reproduced

• Ghosh et al, MagpieTTS-LF, full methods and evaluation, June 2026

  • authors: “without model retraining”
  • splits text into sentences while retaining previous text tokens, encoder representations, and attention position
    • soft attention guidance favors forward progress without discarding all earlier context
    • this is a direct baseline for carrying context between document sections
  • evaluates 20 English long texts on one A6000 GPU
    • compares XTTS, Qwen3-TTS, and VibeVoice
    • competitor sentence splitting or unsplit input uses whichever performs better
    • Whisper-Large measures word errors; pitch and energy changes assess sentence boundaries
  • inference: compare state carry against independent chunks using the same model
    • cross-model differences do not isolate the benefit of state carry
    • these scores do not measure section navigation, formula comprehension, or playback stalls

• Pan et al, Comprehensive Benchmarking of Long-Form Speech Generation in Diverse Scenarios, full paper, ACL Findings 2026

  • authors: “1,101 samples spanning 17 common speech scenarios”
  • retrieved paper calls the benchmark SwanBench-Speech; Anthology abstract calls it LFSBench
  • separates acoustic consistency, content accuracy, and expressive delivery
    • word errors use speech recognition; some delivery scores use learned audio evaluators
    • prosody validation compares 50 matched-text audio pairs rated by ten evaluators
    • reported rank correlation with that human comparison is 0.82
  • inference: content coverage and natural delivery require separate outcomes
    • limited listening validation does not establish reliable scoring of technical-document comprehension

• MGM-Omni, full Long-TTS-Eval methods, appendix A.2, September 2025

  • authors: “formulas, URLs, or classical Chinese poetry”
  • long-text set includes 341 Chinese and 353 English samples across six domains, including academic papers
    • separate hard set contains URLs, email addresses, numbers, and formulas
    • this already covers technical-text rendering beyond ordinary prose
  • recognizers transcribe independent 28-second audio chunks
    • computes English word errors and Chinese character errors
    • takes the smaller error against written text or a GPT-5-generated spoken normalization
  • inference: normalization avoids some false penalties but introduces another model’s possible mistakes
    • manually validate alternative readings of equations, identifiers, and URLs
    • semantic correctness and usable navigation are not established by the smaller error score
  • artifact: evaluation scripts

• refined document-playback hypothesis

  • preserving checked section positions and recovering locally from failed synthesis reduces missing content and playback stalls
    • long-form synthesis, state carry, and technical-text benchmarks are established prior work
    • compare independent chunks, state-carry chunks, and monitored retries at equal model and compute budget
    • measure first-audio delay, stall duration, duplicate or missing sections, and successful navigation to requested content
    • manually score technical meanings rather than relying solely on speech-recognizer agreement
  • read-depth limit: selected full methods and evaluation read for all three sources; artifacts not executed

remaining uncertainty

• targeted full-method review now covers confidence calibration, cascade deferral, dynamic ASR configurations, diarization confidence, and active speaker correction
  • selective repair itself is established prior work
  • proposed distinction is validated whole-recording repair benefit under equal GPU budgets
• real-user timing is reviewed for one older single-speaker transcript interface
  • transfer to modern post-editing, multilingual meetings, and speaker corrections remains untested
• artifact-level reproduction, confidence transfer across languages, and selector training remain untested
• recent joint-model claims need reproduction on the user’s actual server and recording types
• model cards and rolling repositories can change
  • record exact revisions when running the pilot
• central study coordinator handles the requested ChatGPT opinion consultation
  • this page does not claim that consultation has completed

Last edited: 2026-10-08