Used off the shelf, Whisper produced a word error rate of 84.6% for kindergartners reading scripted prompts, against 17.5% for fourth graders on the same test (Singh, Sahidullah and Kinnunen, Causal Analysis of ASR Errors for Children 2025). Word error rate (WER) is the share of words a system substitutes, deletes or inserts compared with a human transcript, so it can even exceed 100%. The same model family scored 25.00% WER on American children talking to a science tutor and 5.40% on adult read speech (Jain et al., Interspeech 2023), and the gap tracks training data: the largest public English child corpus holds about 470 hours (Linguistic Data Consortium, LDC2021S05), while Whisper-largeV3 was trained on 1 million hours (Fan et al., Interspeech 2024). Age is only one axis of the problem; the accent and dialect gaps and hallucination rates follow similar data-coverage logic. We aggregated data from Interspeech, ICASSP and WOCCI papers, the Journal of the Acoustical Society of America, the Linguistic Data Consortium, Frontiers in Education, and the US Federal Trade Commission.
TL;DR
- Zero-shot Whisper WER on the CSLU Kids scripted test: 84.6% in kindergarten vs 17.5% in grade 4 (Singh et al., 2025).
- Whisper Large-V2 off the shelf: 25.00% WER on MyST child speech vs 5.40% on adult LibriTTS speech (Jain et al., Interspeech 2023).
- Zero-shot Whisper large: 43.9% WER for ages 4 to 7 vs 13.4% for ages 8 to 10 (Ying et al., WOCCI 2025).
- Children’s disfluency rate was 10.1% vs 6.2% in adult Switchboard conversations (Tran et al., Interspeech 2020).
- Whisper-largeV3 trained on 1 million hours vs 133 hours of filtered MyST child training data (Fan et al., Interspeech 2024).
- Only about 45% of the 227,567 MyST utterances are transcribed (Linguistic Data Consortium, LDC2021S05).
- Fine-tuning on 10 hours of child speech cut Whisper Large-V2 WER on PF-STAR from 73.68% to 2.88% (Jain et al., Interspeech 2023).
- Generative error correction cut child WER from 30.5% to 21.8%, a 28.5% relative drop (Balaji Shankar et al., CHSER, Interspeech 2025).
- A reading-screening system agreed with human raters at kappa 0.16 for phoneme blending vs 0.72 for vocabulary (Petscher et al., Frontiers in Education 2026).
- The amended COPPA Rule lists voiceprints as personal information and requires audio to be deleted right after use (FTC, COPPA Final Rule 2025).
1. The Child vs Adult Gap Off the Shelf
The headline number for any child speech product is the zero-shot score, meaning the result of a model used as shipped, with no extra training on children’s recordings. Whisper Large-V2 scored 25.00% WER on MyST children’s speech and 73.68% on British children in the PF-STAR corpus, against 5.40% on adult read speech, roughly 4.6 times and 13.6 times the adult rate (25.00 / 5.40 and 73.68 / 5.40) in the same evaluation (Jain et al., Adaptation of Whisper models to child speech recognition, Interspeech 2023). Most recent available data for that comparison: 2023.
Newer models narrow but do not close the gap. In a 2024 benchmark, NVIDIA’s Canary and Parakeet models beat every Whisper size on children’s speech despite far smaller training sets, which led the authors to conclude that data quality is sometimes more important than data size (Fan, Balaji Shankar and Alwan, Interspeech 2024). The ratio has been stubborn for two decades: an IEEE study in 2003 already reported child error rates two to five times the adult rate.
| Metric | Value | Source |
|---|---|---|
| Whisper Large-V2 WER without child training: MyST children vs LibriTTS adult read speech | 25.00% vs 5.40% | Jain et al., Interspeech 2023 |
| Whisper Large-V2 WER without child training: PF-STAR British children; CMU Kids read speech | 73.68%; 12.69% | Jain et al., Interspeech 2023 |
| Whisper-largeV3 zero-shot WER: MyST test vs OGI Kids test | 12.6% vs 19.9% | Fan et al., Interspeech 2024 |
| Canary (1B) vs Parakeet-RNNT (1.1B) zero-shot WER on MyST test; on OGI test | 9.5% vs 11.1%; 18.2% vs 16.7% | Fan et al., Interspeech 2024 |
| Whisper Medium.en zero-shot WER: CSLU Kids spontaneous speech vs LibriSpeech adult test-clean | 31.85% vs 3.02% | Attia et al., Kid-Whisper, AAAI/ACM AIES 2024 |
| Error rate of adult-trained recognizers on children vs adults | two to five times worse | Potamianos and Narayanan, IEEE Transactions on Speech and Audio Processing 2003 |
Context note: the two Whisper MyST figures (25.00% and 12.6%) are not a contradiction in model quality. The 2024 benchmark removed utterances with WER above 50% or fewer than 3 words before scoring, which alone moves results by several points (see section 5).
2. Error Rates by Age Band
“Children” is not one population. On the CSLU Kids scripted test, zero-shot Whisper went from 84.6% WER in kindergarten to 37.5% in grade 1 and 17.5% in grade 4, a 4.8-fold difference (84.6 / 17.5) inside one test set (Singh, Sahidullah and Kinnunen, Causal Analysis of ASR Errors for Children, 2025). Most of the improvement arrives by grade 4; after that the curve stops falling, and Whisper’s WER in grades 6 to 10 sits above its grade 4 low.
A 2025 ESPnet benchmark found the same cliff with different models. Zero-shot Whisper large scored 43.9% for ages 4 to 7, then rose again from 13.4% at ages 8 to 10 to 18.6% at ages 11 to 15, while an open model fine-tuned on child speech stayed below 10% in every band (Ying et al., Benchmarking Training Paradigms for Child ASR in ESPnet, WOCCI 2025). The same age penalty shows up in speaker verification (checking whether two recordings come from the same person), which matters for any child-account voice login.
| Metric | Value | Source |
|---|---|---|
| Zero-shot Whisper WER on CSLU Kids scripted test: kindergarten vs grade 4 | 84.6% vs 17.5% | Singh et al., 2025 |
| Same test: grade 1; grade 2; grade 3 | 37.5%; 28.0%; 23.1% | Singh et al., 2025 |
| Same test: grade 10 | 22.4% | Singh et al., 2025 |
| Zero-shot Whisper large WER by age: 4 to 7; 8 to 10; 11 to 15 | 43.9%; 13.4%; 18.6% | Ying et al., WOCCI 2025 |
| OWSM v3.1 fine-tuned on child speech, WER by age: 4 to 7; 8 to 10; 11 to 15 | 9.0%; 7.3%; 1.8% | Ying et al., WOCCI 2025 |
| Speaker verification equal error rate (share of wrong accept/reject decisions at the balance point), ages 4 to 7 vs 11 to 15 | 13.2% vs 3.73% | Ying et al., WOCCI 2025 |
| Utterances from children with a mean age of 4.9 transcribed fully correctly: Whisper large-v3 vs Google cloud ASR (2024) | 36.8% vs 9.6% | Janssens et al., Ghent University, Technological Advances in Human-Robot Interaction 2024 |
Outlier note: the MMS model scored 193.3% WER on grade 6 in the Singh et al. test, because it inserted far more words than were spoken. WER above 100% is a sign of runaway insertions, not a typo.
3. Why Child Speech Breaks Adult-Trained Models: Acoustics and Disfluency
Two mechanisms dominate. First, the voice itself is a moving target. A study of 436 children aged 5 to 17 and 56 adults found segment durations converge to adult levels only around age 12, with pitch and formant variability (the resonances that identify vowels) taking about 2 to 3 more years (Lee, Potamianos and Narayanan, JASA 1999). A model tuned to adult resonances is, in effect, listening to a different instrument. Most recent available data: Lee, Potamianos and Narayanan, 1999.
Second, children talk differently. Their spontaneous explanations showed higher disfluency and filler rates than adult conversations, including far more ‘um’ and more words cut off mid-way (Tran et al., Analysis of Disfluency in Children’s Speech, Interspeech 2020). Causal modeling in 2025 put age and the number of words in an utterance at the top of the error drivers, ahead of background noise and pronunciation ability. Speech that departs from the adult norm for clinical reasons faces the same trap, as our dysarthric and atypical speech data roundup shows.
| Metric | Value | Source |
|---|---|---|
| Speakers in the developmental acoustics study | 436 children aged 5 to 17; 56 adults | Lee, Potamianos and Narayanan, JASA 1999 |
| Age at which segment durations reach adult levels | around 12 | Lee, Potamianos and Narayanan, JASA 1999 |
| Extra time for within-speaker pitch and formant variability to reach the adult range | about 2 to 3 years | Lee, Potamianos and Narayanan, JASA 1999 |
| Age at which boys’ and girls’ pitch and formant patterns start to split; fully established | about 11; about 15 | Lee, Potamianos and Narayanan, JASA 1999 |
| Disfluency rate: children vs adult Switchboard vs adult CallHome conversations | 10.1% vs 6.2% vs 6.3% | Tran et al., Interspeech 2020 |
| Rate of the filler ‘um’: children vs adult Switchboard | 2.3% vs 0.5% | Tran et al., Interspeech 2020 |
| Rate of word fragments (words cut off mid-way): children vs adult Switchboard | 1.8% vs 0.5% | Tran et al., Interspeech 2020 |
Context note: the disfluency corpus is small (26 children, about 1.26 hours of child speech), and the paper gives the children’s age as 4.8 to 7 in one place and 5 to 8 in another. Most recent available data: 2020.
4. Data Scarcity: Hours of Child Speech vs Foundation-Model Training Sets
The size mismatch is several orders of magnitude. Whisper-largeV3 was trained on 1 million hours of audio, while a 2024 benchmark had 133 hours of filtered MyST child speech to fine-tune on, so the child set equals about 0.0133% of that pretraining pool (133 / 1,000,000) (Fan et al., Interspeech 2024). Even the leaner Parakeet model saw 64,000 hours, about 480 times the MyST training split (64,000 / 133).
MyST itself is large only by child-speech standards. The Linguistic Data Consortium catalog lists about 470 hours from 1,371 students in grades 3 to 5, collected between 2008 and 2017, with 102,433 of 227,567 utterances transcribed (LDC2021S05, MyST Children’s Conversational Speech). Its creators describe the corpus as about 400 hours and note how hard it is to obtain children’s speech at scale (Pradhan, Cole and Ward, LREC-COLING 2024). Younger children and other languages have far less: a 2025 Mandarin corpus for ages 3 to 5 holds 41.25 hours (Zhou et al., ChildMandarin 2025).
| Metric | Value | Source |
|---|---|---|
| Pretraining data: Whisper-largeV3; Canary; Parakeet-TDT | 1,000,000; 85,000; 64,000 hours | Fan et al., Interspeech 2024 |
| MyST corpus as catalogued | about 470 hours; 1,371 students in grades 3 to 5; 10,496 tutoring sessions | Linguistic Data Consortium, LDC2021S05 (2021) |
| MyST utterances with transcripts | 102,433 of 227,567 (about 45%) | Linguistic Data Consortium, LDC2021S05 (2021) |
| ChildMandarin corpus, children aged 3 to 5 | 41.25 hours; 397 speakers | Zhou et al., ChildMandarin 2025 |
| PF-STAR British English children’s corpus, as listed in a 2025 corpus survey | 14.5 hours; 158 speakers aged 4 to 14 | Zhou et al., ChildMandarin 2025 |
| CMU Kids; OGI Kids scripted speech | 9 hours, 76 speakers; 70 hours, 1,118 speakers | Ying et al., WOCCI 2025 |
| Child vs adult training data in a 2018 transfer-learning study | 91.61 hours (ages 6 to 14) vs 205.82 hours (TED-LIUM adults) | Shivakumar and Georgiou, arXiv 2018 |
Context note: MyST hour counts differ by source (about 470 in the LDC catalog, 393 or about 400 in the corpus paper, about 240 transcribed hours in the 2024 benchmark). We report each figure with its own source.
5. Fine-Tuning and Adaptation Gains
Fine-tuning means continuing to train a model on child recordings. It is the most reliable fix, and the gains are largest where the zero-shot model was worst. Whisper Large-V2 fell from 73.68% to 2.88% WER on PF-STAR British children after fine-tuning on just 10 hours of PF-STAR speech (Jain et al., Interspeech 2023). On MyST conversational speech, Whisper-small went from 13.4% to 9.3%, a 30.6% relative reduction ((13.4 - 9.3) / 13.4), and a small adapter module updating 1.29 million parameters matched full fine-tuning of 242 million.
The newer techniques chase the last few points. Merging only the attention weights of related models set a MyST record of 8.69% WER for Whisper-small (Balaji Shankar et al., Selective Attention Merging, ICASSP 2025), and a text-correction model trained on 200K pairs of ASR output and human transcript cut errors across four child corpora. How a test set is cleaned matters as much as method: dropping files with WER above 50% or fewer than 3 words moved Whisper-Large-v1 test WER on MyST from 26.2% to 14.2%, with no change to the model (Attia et al., Kid-Whisper, AIES 2024).
| Metric | Value | Source |
|---|---|---|
| Whisper Large-V2 WER on PF-STAR: before vs after fine-tuning on 10 hours of child speech | 73.68% vs 2.88% | Jain et al., Interspeech 2023 |
| Whisper-small WER, zero-shot vs fine-tuned: MyST test; OGI Kids test | 13.4% vs 9.3%; 25.4% vs 1.8% | Fan et al., Interspeech 2024 |
| Parameters trained, adapter tuning vs full fine-tuning (both 9.3% WER on MyST) | 1.29 million vs 242 million | Fan et al., Interspeech 2024 |
| Whisper Medium.en WER on CSLU Kids spontaneous speech: zero-shot vs fine-tuned on MyST plus CSLU | 31.85% vs 16.53% | Attia et al., Kid-Whisper, AIES 2024 |
| Whisper-small with Selective Attention merging plus data augmentation, MyST WER | 8.69% (up to 14% relative reduction) | Balaji Shankar et al., ICASSP 2025 |
| Generative error correction, WER across four child corpora: before vs after | 30.5% vs 21.8% (28.5% relative) | Balaji Shankar et al., CHSER, Interspeech 2025 |
| Mixed child and adult fine-tuning vs the pre-trained model | up to 5% absolute (21% relative) lower WER | Graave et al., Interspeech 2024 |
| Whisper-small character error rate on Mandarin speech of children aged 3 to 5: zero-shot vs fine-tuned | 37.99% vs 17.45% | Zhou et al., ChildMandarin 2025 |
Outlier note: gains do not transfer cleanly across accents, ages and speaking styles. In the same 2023 study, Whisper Medium scored 19.76% on PF-STAR children after fine-tuning on American MyST speech, against 3.15% after fine-tuning on PF-STAR itself.
6. Classroom Reading Assessment: Where Error Rates Meet Decisions
In education, WER is not the end metric. A reading tutor has to decide whether the child read each word correctly, so a transcript can post a respectable WER and still miss the errors a teacher needs to see. In Dutch oral-reading tests with 71 children aged 7 to 11, Whisper large-v2 with no fine-tuning reached a Matthews correlation coefficient (MCC, an agreement score where 0 means chance and 1 means perfect) of .01 for detecting reading errors, against .63 for a tailored Kaldi system (Molenaar et al., Automatic Assessment of Oral Reading Accuracy for Reading Diagnostics, Interspeech 2023). Most recent available data: 2023.
A 2026 US study of a commercial speech verification system found the same pattern by task (Petscher et al., Frontiers in Education 2026). Agreement with human raters, measured by Cohen’s kappa (agreement beyond chance, from 0 to 1), was highest for expressive vocabulary and lowest for phoneme blending, the most phonologically complex of the three tasks. The fit also differed by student group on word reading, a finding that matters for any district using automated screening to decide which children need phonics support.
| Metric | Value | Source |
|---|---|---|
| Children in the Dutch oral-reading study; age range | 71; 7 to 11 | Molenaar et al., Interspeech 2023 |
| Agreement with human reading-error judgments (MCC): Kaldi-PM system vs Whisper large-v2 | .63 vs .01 | Molenaar et al., Interspeech 2023 |
| WER on children’s reading against manual transcripts: Kaldi-PM system vs Whisper large-v2 | 5.2% vs 13.4% | Molenaar et al., Interspeech 2023 |
| Kindergartners in the US speech verification study | 429 from 20 schools in three states | Petscher et al., Frontiers in Education 2026 |
| Kappa between system and human raters at the 50 threshold: phoneme blending; expressive vocabulary; word reading | 0.16; 0.72; 0.51 | Petscher et al., Frontiers in Education 2026 |
| Share of responses classified the same as human raters at the 50 threshold, same three tasks | 0.52; 0.87; 0.80 | Petscher et al., Frontiers in Education 2026 |
| Word reading kappa at the 80 threshold: Black vs White students | 0.25 vs 0.47 | Petscher et al., Frontiers in Education 2026 |
Context note: Kaldi-PM is not a deployable shortcut, because its vocabulary list combines the reading prompts with words from the manual transcripts of the test data. Treat it as a measure of what a tightly constrained system can do, and read the 5.2% against that setup.
7. Privacy Rules That Limit Child Speech Collection
The data gap in section 4 is partly a policy outcome. The amended COPPA Rule, published April 22, 2025, added biometric identifiers to the definition of children’s personal information and names voiceprints explicitly (FTC, COPPA Final Rule Amendments 2025). Its parental-consent exception for audio covers a file holding a child’s voice and no other personal information, used only to answer the child’s request and deleted immediately afterwards, which leaves no room under that exception for keeping the file as training data.
Enforcement backs the text. In 2023 the FTC and Department of Justice charged Amazon with keeping children’s Alexa recordings indefinitely; the proposed order carried a 25 million dollar civil penalty and barred using the covered voice data to build or improve data products (FTC press release, May 31, 2023). Researchers are responding by moving recognition onto the device: a fine-tuned Whisper tiny.en model ran on a Raspberry Pi at 15.9% WER on MyST, citing privacy rules as the motivation (Dutta, Chandupatla and Hansen, WOCCI 2025).
| Metric | Value | Source |
|---|---|---|
| Age covered by COPPA | children under 13 | FTC press release, 2023 |
| Amended COPPA Rule: Commission vote; public comments reviewed | 5-0; nearly 300 | FTC press release, January 16, 2025 |
| Amended rule: Federal Register publication; full compliance window | April 22, 2025; one year from publication | FTC, COPPA Final Rule Amendments 2025 |
| Biometric identifiers named as personal information | voiceprints, alongside fingerprints, facial templates and gait patterns | FTC, COPPA Final Rule 2025 |
| Amazon Alexa case: civil penalty; children using Alexa through their own profile | 25 million dollars; more than 800,000 | FTC, Amazon Alexa press release and business blog, May 31, 2023 |
| On-device Whisper tiny.en on a Raspberry Pi: MyST WER; with data filtering; real-time factor (processing time divided by audio length) | 15.9%; 11.8%; 0.23 to 0.41 | Dutta et al., WOCCI 2025 |
Context note: the FTC’s 2017 enforcement policy on voice recordings set the same short-retention logic, allowing audio collected only as a replacement for written words and deleted promptly (FTC policy statement, 2017). Most recent available data on the Amazon case: 2023.
Summary: Children’s Speech Recognition Word Error Rate by the Numbers
| Metric | Value | Source |
|---|---|---|
| Zero-shot Whisper WER on CSLU Kids scripted test: kindergarten vs grade 4 | 84.6% vs 17.5% | Singh et al., 2025 |
| Whisper Large-V2 WER without child training: MyST children vs adult read speech | 25.00% vs 5.40% | Jain et al., Interspeech 2023 |
| Zero-shot Whisper large WER: ages 4 to 7 vs 8 to 10 | 43.9% vs 13.4% | Ying et al., WOCCI 2025 |
| Speaker verification error rate: ages 4 to 7 vs 11 to 15 | 13.2% vs 3.73% | Ying et al., WOCCI 2025 |
| Children with mean age 4.9: fully correct transcriptions, Whisper large-v3 vs Google | 36.8% vs 9.6% | Janssens et al., Technological Advances in Human-Robot Interaction 2024 |
| Adult-trained recognizers on children vs adults | two to five times worse | Potamianos and Narayanan, IEEE TSAP 2003 |
| Age at which children’s segment durations reach adult levels | around 12 | Lee, Potamianos and Narayanan, JASA 1999 |
| Disfluency rate: children vs adult Switchboard | 10.1% vs 6.2% | Tran et al., Interspeech 2020 |
| Pretraining data, Whisper-largeV3 vs filtered MyST child training set | 1,000,000 vs 133 hours | Fan et al., Interspeech 2024 |
| MyST corpus size | about 470 hours; 1,371 students | Linguistic Data Consortium, LDC2021S05 |
| Mandarin child corpus, ages 3 to 5 | 41.25 hours; 397 speakers | Zhou et al., ChildMandarin 2025 |
| Whisper Large-V2 on PF-STAR: before vs after 10 hours of fine-tuning | 73.68% vs 2.88% | Jain et al., Interspeech 2023 |
| Best reported Whisper-small WER on MyST (attention merging plus augmentation) | 8.69% | Balaji Shankar et al., ICASSP 2025 |
| Generative error correction across four child corpora | 30.5% to 21.8% WER | Balaji Shankar et al., CHSER, Interspeech 2025 |
| Reading-error detection agreement (MCC): Kaldi-PM system vs Whisper large-v2 | .63 vs .01 | Molenaar et al., Interspeech 2023 |
| Kindergarten screening kappa: phoneme blending vs expressive vocabulary | 0.16 vs 0.72 | Petscher et al., Frontiers in Education 2026 |
| Amazon Alexa children’s voice data case, civil penalty | 25 million dollars | FTC, 2023 |
Methodology and Sources
We ran 21 web searches and read the primary paper, catalog entry, journal article or regulatory text behind every figure above. WER values are reported exactly as each paper prints them; test sets, text normalization and filtering differ between studies, so compare numbers within a row or a single study rather than across studies. Figures older than three years are flagged in their sections as most recent available data.
- Singh, Sahidullah and Kinnunen, Causal Analysis of ASR Errors for Children, arXiv 2025 (submitted to Computer Speech and Language)
- Jain, Barcovschi, Yiwere, Corcoran and Cucu, Adaptation of Whisper models to child speech recognition, Interspeech 2023
- Fan, Balaji Shankar and Alwan, Benchmarking Children’s ASR with Supervised and Self-supervised Speech Foundation Models, Interspeech 2024
- Ying et al., Benchmarking Training Paradigms, Dataset Composition, and Model Scaling for Child ASR in ESPnet, WOCCI 2025
- Attia et al., Kid-Whisper: Towards Bridging the Performance Gap in Automatic Speech Recognition for Children vs Adults, AAAI/ACM AIES 2024
- Janssens et al., Child Speech Recognition in Human-Robot Interaction: Problem Solved?, 2024
- Potamianos and Narayanan, Robust Recognition of Children’s Speech, IEEE Transactions on Speech and Audio Processing 2003
- Lee, Potamianos and Narayanan, Acoustics of children’s speech: developmental changes of temporal and spectral parameters, JASA 1999
- Tran et al., Analysis of Disfluency in Children’s Speech, Interspeech 2020
- Linguistic Data Consortium, MyST Children’s Conversational Speech, LDC2021S05
- Pradhan, Cole and Ward, My Science Tutor (MyST): A Large Corpus of Children’s Conversational Speech, LREC-COLING 2024
- Zhou et al., ChildMandarin: A Comprehensive Mandarin Speech Dataset for Young Children Aged 3-5, arXiv 2025
- Shivakumar and Georgiou, Transfer Learning from Adult to Children for Speech Recognition, arXiv 2018
- Balaji Shankar et al., Selective Attention Merging for low resource tasks: A case study of Child ASR, ICASSP 2025
- Balaji Shankar et al., CHSER: A Dataset and Case Study on Generative Speech Error Correction for Child ASR, Interspeech 2025
- Graave et al., Mixed Children/Adult/Childrenized Fine-Tuning for Children’s ASR, Interspeech 2024
- Molenaar et al., Automatic Assessment of Oral Reading Accuracy for Reading Diagnostics, Interspeech 2023
- Petscher et al., Evaluation of the Consistency of a Speech Verification System with Human Raters in Early Literacy Screening Assessments, Frontiers in Education 2026
- Dutta, Chandupatla and Hansen, Adapting Whisper for Lightweight and Efficient ASR of Children for On-device Edge Applications, WOCCI 2025
- FTC, FTC Finalizes Changes to Children’s Privacy Rule, January 16, 2025 and 16 CFR Part 312: COPPA Final Rule Amendments
- FTC and DOJ, Amazon Alexa press release, May 31, 2023 and FTC business blog on the Alexa case
- Data watch: (1) zero-shot Whisper results on MyST range from about 12.5% to 25% depending on whether the test set was filtered, so the 2023 and 2024 benchmarks are not directly comparable. (2) MyST size is reported as about 470 hours (LDC), 393 or about 400 hours (corpus paper) and about 240 transcribed hours (2024 benchmark). (3) The Singh et al. grade-level results are a preprint, and the paper describes MyST as grades 3 to 5 in one place and grades 5 to 8 in another. (4) The ESPnet benchmark attributes its best average WER to Whisper-small in one section and Whisper medium in another. (5) The CHSER abstract’s 13.3% reduction for already fine-tuned systems applies to Whisper Tiny only; the text gives a 9% to 13.3% range. (6) The Molenaar et al. paper prints Kaldi-PM’s MCC as .62 in its results text and .63 elsewhere (we use .63), and its table shows Kaldi-MO, whose vocabulary comes only from the test transcripts, at .65. (7) The Frontiers study evaluated one commercial system, so its agreement figures should not be generalized to all reading tools. (8) Several foundational figures (1999, 2003, 2020, 2023) are flagged as most recent available data; ICASSP 2027 papers are expected to update the benchmark numbers.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when the ICASSP 2027 proceedings are published.