When the Speech Accessibility Project began, the largest atypical speech database included 16 people with cerebral palsy; by the end of September 2025, about 2,000 participants had submitted recordings to the project (Speech Accessibility Project, 2025 and 2026). Google’s Project Euphonia, the other large corpus, reached over 1.5 million utterances from about 3,000 speakers by February 2025, but it is closed to outside teams (Martin et al., Frontiers in Language Sciences 2025; Google Project Euphonia). More data has not yet erased the gap: in a 2026 benchmark of eight AI systems, typical speech stayed below 5% word error rate (WER, the share of words a system gets wrong) while severe dysarthria exceeded 51% for every system (Alsayegh and Masood, International Journal of Intelligent Systems 2026). We aggregated data from the Speech Accessibility Project at the University of Illinois Beckman Institute, Google Research, the Linguistic Data Consortium, NIDCD, ASHA, and more than a dozen peer-reviewed papers. Our speech recognition accent bias statistics cover how these errors compare with gaps by race, dialect and accent; this roundup looks at the datasets themselves: how big they are, who is in them, and what they do to error rates.
TL;DR
- About 2,000 people had submitted recordings to the Speech Accessibility Project by the end of September 2025 (Speech Accessibility Project, 2026).
- Project Euphonia held over 1.5 million utterances from about 3,000 speakers in February 2025, and Google does not share it outside the company (Martin et al., Frontiers in Language Sciences 2025).
- TORGO, a standard benchmark since 2012, holds 8 speakers with dysarthria and about 23 hours of audio (Linguistic Data Consortium, LDC2012S02).
- UA-Speech recorded 19 speakers with cerebral palsy in 2008, and 16 remain in the current distribution (University of Illinois, UA-Speech 2008).
- Parkinson’s, cerebral palsy, ALS, Down syndrome and stroke were each planned at 400 participants in the Speech Accessibility Project (ClinicalTrials.gov NCT05889260).
- ALS speakers made up 39.2% of a 195-speaker Euphonia research subset, against 10.8% for Parkinson’s disease (Tobin and Tomanek, Google 2021).
- Every one of eight AI systems exceeded 51% WER on severe dysarthria in TORGO (Alsayegh and Masood, IJIS 2026).
- Whisper-Turbo’s zero-shot character error rate on UA-Speech was 9.3% for control speakers and 165.2% for dysarthric speakers (de Groot et al., TU Delft 2025).
- Fine-tuning Google’s USM on disordered speech cut severe-speech WER from 48.1 to 31.3 (Tobin et al., Google Research 2024).
- Personalized models improved ALS WER by 62% relative, with 71% of the gain from 5 minutes of data (Shor et al., Interspeech 2019).
- TORGO contains only 957 unique texts among 16,394 text entries (Hui et al., 2024).
- Dysarthria affects 44% to 88% of people with Parkinson’s disease (ASHA Practice Portal, Dysarthria in Adults).
1. The Two Large Corpora: Speech Accessibility Project and Project Euphonia
Disordered speech data now exists at a scale comparable to LibriSpeech, a widely used read-speech benchmark, but only one of the two large collections is open. About 2,000 participants had submitted recordings to the Speech Accessibility Project by the end of September 2025, and the project says about 65 organizations are using its data. Google’s Euphonia corpus is larger in speakers, yet its project page states that the dataset is not available to outside teams or companies (Google, Project Euphonia).
The practical meaning is that open research on disordered speech recognition runs largely through one gatekeeper. The Speech Accessibility Project releases data under a data use agreement (a signed contract that sets who may use the recordings and for what), which requires two signatures from the applying organization and lasts at most 3 years (Speech Accessibility Project, Conduct Research page). Euphonia’s growth, from 1.2 million utterances in 2024 to over 1.5 million in 2025, stays inside Google.
| Metric | Value | Source |
|---|---|---|
| Participants who had submitted recordings to the Speech Accessibility Project, end of September 2025 | About 2,000 | Speech Accessibility Project, homepage 2026 |
| Speech Accessibility Project data package, catalogue description | About 1,090 hours of speech from about 1,130 participants; five conditions | University of York YorVoice catalogue, Speech Accessibility Project record |
| Speech Accessibility Project research access page | 1,500 hours of recorded speech; about 999 participants; about 65 organizations using the data | Speech Accessibility Project, Conduct Research page 2026 |
| Project Euphonia corpus, February 2025 | Over 1.5 million utterances from about 3,000 speakers | Google (Martin et al.), Frontiers in Language Sciences 2025 |
| Project Euphonia corpus, 2024 | 1.2 million utterances from over 2,000 speakers | Google (Jiang et al.), Interspeech 2024 |
| Euphonia recordings with human-reviewed, corrected transcripts | 352,130 (29% of the corpus) | Google (Jiang et al.), Interspeech 2024 |
| Euphonia speakers with speech characteristic labels | Over 75% of speakers, more than 40 labels | Google (Jiang et al.), Interspeech 2024 |
Context: for scale, LibriSpeech, built from read audiobooks, is a corpus of about 1,000 hours (OpenSLR, LibriSpeech). The two Speech Accessibility Project descriptions do not match (more hours but fewer participants on the project’s own page), which we record in the Data watch note. The Euphonia figures come from Jiang et al. and Martin et al..
2. The Older Corpora: TORGO, UA-Speech, and the Small-Data Era
For more than a decade, dysarthric speech recognition was benchmarked on a handful of speakers. TORGO contains 8 speakers with dysarthria (5 male, 3 female) and 7 control speakers, about 23 hours of English recorded in Toronto between 2008 and 2010. UA-Speech, recorded at the University of Illinois, covered 19 speakers with cerebral palsy, each reading 765 isolated words (Kim et al., Interspeech 2008). A model reported as state of the art on either corpus has, in effect, been tested on fewer people than a typical classroom.
The corpora also shrink and fragment over time. UA-Speech now distributes 16 of its original 19 speakers because three withdrew permission or had corrupted data, and its license bars commercial use (University of Illinois, UA-Speech). Outside English, the picture is similar: an Italian command corpus and a Mandarin corpus are each built on a few dozen speakers.
| Metric | Value | Source |
|---|---|---|
| TORGO speakers | 8 with dysarthria (cerebral palsy or ALS) and 7 controls | Linguistic Data Consortium, TORGO LDC2012S02 (2012) |
| TORGO audio and release | About 23 hours, recorded 2008 to 2010; released January 19, 2012 | Linguistic Data Consortium, TORGO LDC2012S02 (2012) |
| TORGO reading material | 162 sentences from Yorkston and Beukelman, 460 sentences derived from TIMIT, plus tongue and lip movement recordings (electromagnetic articulography) | Linguistic Data Consortium, TORGO LDC2012S02 (2012) |
| UA-Speech speakers with cerebral palsy, original vs currently distributed | 19 vs 16 | Kim et al., Interspeech 2008; University of Illinois, UA-Speech page |
| UA-Speech vocabulary and recording setup | 765 isolated words per speaker, including 300 distinct uncommon words; 8-microphone array | Kim et al., Interspeech 2008 |
| Largest atypical speech database when the Speech Accessibility Project began | 16 people with cerebral palsy | Speech Accessibility Project, Microsoft Azure article 2025 |
| EasyCall (Italian) | 21,386 recordings from 31 dysarthric and 24 healthy speakers | Turrisi et al., Interspeech 2021 |
| CDSD (Mandarin) | 133 hours from 44 speakers | Wang et al., Chinese Dysarthria Speech Database (arXiv v2, 2025) |
Most recent available data: TORGO (LDC, 2012), UA-Speech (Kim et al., 2008) and EasyCall (Interspeech 2021) are all older than three years, and they remain the corpora most papers still report on. The Mandarin CDSD paper gives the 133 hours as 124 hours of audio plus 9 hours of video, with 80 of the audio hours coming from just 8 speakers recorded for 10 hours each.
3. Who Is in the Data: Etiology, Severity, and Language
Every large corpus is skewed toward whoever was easiest to recruit first. The Speech Accessibility Project was designed for 2,000 people, 400 for each of five etiologies (the underlying medical cause of the speech difference): Parkinson’s disease, ALS, cerebral palsy, Down syndrome and stroke (ClinicalTrials.gov NCT05889260). Its first data package contained only Parkinson’s speakers, and the challenge release of April 2024, recorded mainly in 2023, was still dominated by them (Zheng et al., Interspeech 2025 Challenge). Our Parkinson’s speech and voice treatment statistics cover the clinical side of that population.
Euphonia’s skew runs the other way, toward ALS, and toward milder speech. In a 195-speaker research subset, ALS accounted for 39.2% of speakers and a quarter were rated severe. Most rated recordings in the Speech Accessibility Project’s first package sat at the least impaired end of the intelligibility rating, which matters because a dataset of mostly mild speech will flatter a model’s average.
| Metric | Value | Source |
|---|---|---|
| Speech Accessibility Project planned design | 2,000 people, 400 per etiology, 1.2 million speech samples (600 per person) | ClinicalTrials.gov NCT05889260, Speech Accessibility Project record |
| Actual enrollment in the registry record (which lists Parkinson’s disease and related parkinsonian conditions) | 485 participants; study completed July 31, 2025 | ClinicalTrials.gov NCT05889260, Speech Accessibility Project record |
| SAP-240430 challenge release, training split | 369 speakers, 131,420 utterances, 290.35 hours | Zheng et al., Interspeech 2025 Speech Accessibility Project Challenge |
| SAP-240430 development and test splits | Dev 55 speakers (43.56 hours); Test1 50 speakers (42.16 hours); Test2 50 speakers (38.77 hours) | Zheng et al., Interspeech 2025 Speech Accessibility Project Challenge |
| First data package recruitment, people registered vs approved vs recorded | 746 vs 283 vs 253 | Speech Accessibility Project (Hasegawa-Johnson et al.), JSLHR 2024 |
| First data package, intelligibility ratings on a 1 to 7 scale (1 = least impaired) | 76% of rated recordings at 1; 1.5% at 5 to 7 | Speech Accessibility Project (Hasegawa-Johnson et al.), JSLHR 2024 |
| Euphonia 195-speaker subset, by etiology | ALS 39.2%; cerebral palsy 13.4%; Parkinson’s disease 10.8%; hearing impairment 5.2%; ataxia 3.1% | Google (Tobin and Tomanek), Personalized ASR on Small Disordered Speech Datasets 2021 |
| Euphonia 195-speaker subset, by severity | Typical 7.7%; mild 35.4%; moderate 31.8%; severe 25.1% | Google (Tobin and Tomanek), Personalized ASR on Small Disordered Speech Datasets 2021 |
Context: language coverage is the thinnest dimension. Every participant in the Speech Accessibility Project’s first package reported US English as a native language (Hasegawa-Johnson et al., JSLHR 2024), and Google’s Euphonia page notes that the Speech Accessibility Project program is currently available in English only. Euphonia’s expansion to four more languages had reached 132 speakers by 2025: 76 in Japanese, 38 in Spanish, 14 in French and 4 in Hindi (Martin et al., Frontiers in Language Sciences 2025). People with ALS face the same gap as their speech changes, which our ALS bulbar speech decline statistics track over time.
4. The Word Error Rate Gap for Dysarthric Speech
Off-the-shelf recognizers are close to solved for mild dysarthria and still fail at the severe end. In a 2026 benchmark of eight systems on TORGO, including Whisper large-v3, Deepgram Nova-3, GPT-4o and Gemini 2.5 Pro, severe dysarthria exceeded 51% WER for every system, while leading systems scored roughly 1% to 2% on mild dysarthria (Alsayegh and Masood, IJIS 2026). Multimodal large language models showed no advantage over conventional recognizers under default settings.
Severity, not the brand of model, is the variable that predicts failure. Google’s own baseline climbs from 10.7 WER on mild speech to 48.1 on severe speech on its prompted test set (Tobin et al., Google Research 2024). On UA-Speech, zero-shot Whisper-Turbo (zero-shot meaning used with no adaptation to the speakers) produced a character error rate above 100%, which happens when a model outputs more wrong characters, including invented words, than the reference contains.
| Metric | Value | Source |
|---|---|---|
| TORGO, eight commercial and multimodal AI systems, typical speech vs severe dysarthria | Below 5% WER vs above 51% WER for every system | Alsayegh and Masood, International Journal of Intelligent Systems 2026 |
| TORGO, mild dysarthria, leading systems | Roughly 1% to 2% WER | Alsayegh and Masood, International Journal of Intelligent Systems 2026 |
| Severe dysarthria, default prompt vs verbatim-transcription prompt | GPT-4o 60.1% to 52.9%; GPT-4o Mini 66.0% to about 55% | Alsayegh and Masood, International Journal of Intelligent Systems 2026 |
| UA-Speech, Whisper-Turbo zero-shot character error rate, control speakers vs dysarthric speakers | 9.3% vs 165.2% | de Groot et al. (TU Delft), Diffusion-Based Speech Enhancement for Dysarthric Speech 2025 |
| Euphonia prompted test set, USM baseline WER by severity | Mild 10.7; moderate 29.5; severe 48.1 | Google (Tobin et al.), Towards a Single ASR Model That Generalizes to Disordered Speech 2024 |
| Parkinson’s test set (all prompts), models trained only on typical read speech | LibriSpeech 100 hours: 45.15% WER; LibriSpeech 960 hours: 38.76% WER | Speech Accessibility Project (Hasegawa-Johnson et al.), JSLHR 2024 |
| Speakers whose out-of-the-box WER met a 15% home automation target, all speakers vs severe speakers | 35% vs 10% | Google (Tobin and Tomanek), Personalized ASR on Small Disordered Speech Datasets 2021 |
Outlier: GPT-4o’s improvement with a verbatim prompt shows that some of the multimodal models’ errors on severe speech come from rewriting what was said, not only from mishearing it; the authors report that the Gemini models gained nothing consistent from the same prompts. The UA-Speech figure is from de Groot et al., who report the dysarthric result only as an aggregate, not by intelligibility group, for the zero-shot condition.
5. Gains from Personalized and Fine-Tuned Models
Small amounts of the right data move error rates more than any architecture change. Personalized models improved WER by 62% relative for speakers with ALS, and 71% of that improvement came from just 5 minutes of training data (Shor et al., Interspeech 2019). Those models were trained on 67 ALS speakers, about 36 hours of audio (Google Research blog, 2019).
The newer result is that a single shared model can absorb much of that benefit. Google’s USM, fine-tuned on 1,158 hours of disordered speech, closed 64% of the gap between the baseline and per-speaker models, and Microsoft reports accuracy gains of 18% to 60% in Azure AI Speech from Speech Accessibility Project data, depending on the speaker’s disability. The ceiling is still severity: the fine-tuned model’s severe-speech WER of 31.3 is more than four times its mild-speech figure of 7.3 in the same table (31.3 divided by 7.3 is 4.3).
| Metric | Value | Source |
|---|---|---|
| Relative WER improvement from personalized models, ALS speech vs accented speech | 62% vs 35% | Google (Shor et al.), Interspeech 2019 |
| ALS speakers’ WER after personalization, message bank phrases, mild vs more serious dysarthria | 10% vs 20% | Google (Shor et al.), Interspeech 2019 |
| Speakers reaching a 15% WER home automation target with 50 utterances (3 to 4 minutes) vs 250 utterances (18 to 20 minutes) | 63% vs 79% | Google (Tobin and Tomanek), Personalized ASR on Small Disordered Speech Datasets 2021 |
| USM fine-tuned on disordered speech, WER before vs after, by severity | Mild 10.7 to 7.3; moderate 29.5 to 19.6; severe 48.1 to 31.3 | Google (Tobin et al.), Towards a Single ASR Model That Generalizes to Disordered Speech 2024 |
| USM fine-tuning, relative improvement on prompted vs spontaneous conversational speech | 33% vs 26% | Google (Tobin et al.), Towards a Single ASR Model That Generalizes to Disordered Speech 2024 |
| Microsoft Azure AI Speech gains using Speech Accessibility Project data | 18% to 60%, depending on the speaker’s disability | Speech Accessibility Project, Microsoft Azure article 2025 |
| Mandarin CDSD test character error rate, trained only on CDSD vs pretrained on WenetSpeech and then fine-tuned | 24.9% vs 16.4% | Wang et al., Chinese Dysarthria Speech Database (arXiv v2, 2025) |
| UA-Speech, fine-tuned Whisper-Turbo character error rate by intelligibility group | Very low 64.4%; low 44.8%; mid 40.1%; high 30.2% | de Groot et al. (TU Delft), Diffusion-Based Speech Enhancement for Dysarthric Speech 2025 |
Context: severe speakers remain the hardest group even with all of their data. In the Euphonia personalization study, 69% of severe speakers reached the 15% home automation target with their full recordings, against 99% of mild speakers (Tobin and Tomanek, 2021). The Speech Accessibility Project’s leader has said training on part of the dataset cut error rates for people with Parkinson’s by almost a factor of two (Speech Accessibility Project, data sharing announcement 2024). Most recent available data on per-speaker personalization: Shor et al. (2019) and Tobin and Tomanek (2021).
6. Benchmark Pitfalls: Prompt Overlap, Recording Artifacts, and Label Noise
Small corpora with repeated prompts make published accuracy look better than real use. TORGO contains only 957 unique texts among 16,394 text entries, and per-speaker prompt overlap for the eight dysarthric speakers ranges from 95.7% to 100% (Hui et al., 2024). When the authors rebuilt the split so that no prompt appears in both training and test, a wav2vec2 model’s WER for a mild speaker rose from 3.17% to 45.21%. A model can learn the sentence list instead of the speaker.
The same effect appears in the newest corpus. On prompts that training speakers had also read, the Speech Accessibility Project’s fine-tuned model scored 5.18% WER, and the paper reports shared and unshared prompts separately for that reason. Recording conditions are a second trap: Schu et al. (2022) found that most state-of-the-art dysarthria classifiers performed the same or better on the non-speech segments of UA-Speech and TORGO, meaning they may be detecting the recording setup rather than the speech.
| Metric | Value | Source |
|---|---|---|
| TORGO unique texts vs total text entries | 957 vs 16,394 | Hui et al., Enhancing AAC Software for Dysarthric Speakers, arXiv 2024 |
| TORGO per-speaker prompt overlap, eight dysarthric speakers | 95.7% to 100% | Hui et al., Enhancing AAC Software for Dysarthric Speakers, arXiv 2024 |
| wav2vec2 WER, severe speaker M04, original split vs no-overlap split | 64.50% vs 93.75% | Hui et al., Enhancing AAC Software for Dysarthric Speakers, arXiv 2024 |
| wav2vec2 WER, mild speaker M03, original split vs no-overlap split | 3.17% vs 45.21% | Hui et al., Enhancing AAC Software for Dysarthric Speakers, arXiv 2024 |
| Parkinson’s prompts shared with training speakers, model trained on LibriSpeech 960 hours vs fine-tuned on project data | 45.39% vs 5.18% WER | Speech Accessibility Project (Hasegawa-Johnson et al.), JSLHR 2024 |
| Euphonia transcript review, transcripts standardized per speaker vs utterances with actual word changes | 72% on average vs under 5% | Google (Jiang et al.), Interspeech 2024 |
| Euphonia automatic no-speech filter, share of flagged recordings that actually contained speech (false omission rate) | 47.7% | Google (Jiang et al.), Interspeech 2024 |
| USM fine-tuned on disordered speech, LibriSpeech Clean WER on typical speech | 2.4, unchanged at the tested data weights | Google (Tobin et al.), Towards a Single ASR Model That Generalizes to Disordered Speech 2024 |
Context: label noise in Euphonia was mostly cosmetic, since about 95% of corrections were punctuation, capitalization or other normalization; the larger risk was the automatic filter discarding valid but quiet or slow speech, which removes exactly the speakers a dataset most needs. The USM result answers a common worry: adding disordered speech did not degrade typical-speech accuracy at the weights the authors tested.
7. Prevalence Context: The Population These Datasets Serve
The corpora above cover a few thousand speakers; the population with speech differences numbers in the millions. In 2012, 7.7% of US children aged 3 to 17 had a communication or swallowing disorder in the past 12 months, and speech problems were the most common type at 5.0% (NCHS Data Brief No. 205). Only 55.2% of those children received an intervention service in the same year.
Among adults, dysarthria is a frequent feature of common neurological conditions rather than a rare disorder. ASHA’s review puts it at 44% to 88% of people with Parkinson’s disease and 22% to 58% of people with acute stroke (ASHA Practice Portal, Dysarthria in Adults), and those ranges, not the composition of a convenience sample, are what a representative dataset would need to reflect. Fluency disorders such as stuttering add another population, profiled in our stuttering demographics statistics.
| Metric | Value | Source |
|---|---|---|
| US children aged 3 to 17 with a communication or swallowing disorder, past 12 months; speech problems | 7.7%; 5.0% | NCHS Data Brief No. 205 (2015, 2012 NHIS data) |
| Prevalence by age, 3 to 6 years vs 11 to 17 years | 11.0% vs 4.9% | NCHS Data Brief No. 205 (2015, 2012 NHIS data) |
| Children with a disorder who received intervention services, past 12 months | 55.2% | NCHS Data Brief No. 205 (2015, 2012 NHIS data) |
| US adults reporting a voice problem in the past 12 months | 7.6% (about 17.9 million) | NIDCD, Quick Statistics About Voice, Speech, and Language |
| Americans who stutter; Americans living with aphasia | More than 3 million (about 1%); about 2 million, with nearly 180,000 new cases a year | NIDCD, Quick Statistics About Voice, Speech, and Language |
| Dysarthria rates, Parkinson’s disease; multiple sclerosis; acute stroke; acute to subacute traumatic brain injury | 44% to 88%; 40% to 51%; 22% to 58%; about 30% to 86% | ASHA Practice Portal, Dysarthria in Adults |
| Dysarthria rates, Huntington’s disease; neuromuscular disease such as ALS | 78% to 93%; 26% to 62% | ASHA Practice Portal, Dysarthria in Adults |
| US adults with hearing, speech or language difficulties who reported encountering stigma | 65% of 1,004 surveyed | ASHA-YouGov National Poll 2024 |
Most recent available data: the child prevalence figures come from the 2012 National Health Interview Survey, published in 2015, and NIDCD’s own summary (NIDCD Quick Statistics) cites the same survey. The stigma figure is from an online poll fielded February 21 to 26, 2024 (ASHA-YouGov, 2024).
Summary: Atypical and Dysarthric Speech AI Datasets by the Numbers
| Metric | Value | Source |
|---|---|---|
| Participants who had submitted recordings to the Speech Accessibility Project, end of September 2025 | About 2,000 | Speech Accessibility Project, homepage 2026 |
| Speech Accessibility Project data package, catalogue description | About 1,090 hours from about 1,130 participants | University of York YorVoice catalogue, Speech Accessibility Project record |
| Project Euphonia corpus, February 2025 | Over 1.5 million utterances from about 3,000 speakers | Google (Martin et al.), Frontiers in Language Sciences 2025 |
| Euphonia recordings with human-reviewed transcripts | 352,130 (29%) | Google (Jiang et al.), Interspeech 2024 |
| TORGO speakers and audio | 8 dysarthric and 7 control speakers; about 23 hours | Linguistic Data Consortium, LDC2012S02 (2012) |
| UA-Speech speakers with cerebral palsy, original vs currently distributed | 19 vs 16 | Kim et al., Interspeech 2008; University of Illinois |
| Speech Accessibility Project planned design | 2,000 people, 400 per etiology | ClinicalTrials.gov NCT05889260 |
| SAP-240430 training split | 369 speakers, 290.35 hours | Zheng et al., Interspeech 2025 Speech Accessibility Project Challenge |
| Euphonia 195-speaker subset, ALS vs Parkinson’s share | 39.2% vs 10.8% | Google (Tobin and Tomanek) 2021 |
| Euphonia non-English speakers, 2025 | 132 (Japanese 76, Spanish 38, French 14, Hindi 4) | Google (Martin et al.), Frontiers in Language Sciences 2025 |
| TORGO, typical speech vs severe dysarthria, eight AI systems | Below 5% vs above 51% WER | Alsayegh and Masood, IJIS 2026 |
| UA-Speech, zero-shot Whisper-Turbo character error rate, control vs dysarthric | 9.3% vs 165.2% | de Groot et al. (TU Delft) 2025 |
| USM WER on severe disordered speech, before vs after fine-tuning | 48.1 vs 31.3 | Google (Tobin et al.) 2024 |
| Personalized models, relative WER improvement for ALS speech | 62%, with 71% of the gain from 5 minutes of data | Google (Shor et al.), Interspeech 2019 |
| Parkinson’s prompts shared with training speakers, LibriSpeech 960-hour model vs project-fine-tuned model | 45.39% vs 5.18% WER | Speech Accessibility Project (Hasegawa-Johnson et al.), JSLHR 2024 |
| TORGO unique texts vs text entries | 957 vs 16,394 | Hui et al., arXiv 2024 |
| US children aged 3 to 17 with a communication or swallowing disorder (2012) | 7.7% | NCHS Data Brief No. 205 (2015) |
| Dysarthria among people with Parkinson’s disease | 44% to 88% | ASHA Practice Portal, Dysarthria in Adults |
Methodology and Sources
We ran 17 searches and read the primary dataset page, registry record, journal article or preprint behind every figure. Error rates are reported in each source’s own notation: WER for word error rate, character error rate (CER) where the source uses it (UA-Speech fine-tuning and the Mandarin corpus), with or without a percent sign as the source prints it. Figures older than three years are flagged in their sections as most recent available data.
- Speech Accessibility Project homepage (University of Illinois Beckman Institute), 2026
- Speech Accessibility Project, Conduct Research Through the Project
- Speech Accessibility Project, Speech Accessibility Project Data Leads to Recognition Improvements on Microsoft Azure, 2025
- Speech Accessibility Project, Now Sharing Recordings and Data, 2024
- University of York YorVoice catalogue, Speech Accessibility Project record
- ClinicalTrials.gov, Speech Accessibility Project, NCT05889260
- Hasegawa-Johnson et al., Community-Supported Shared Infrastructure in Support of Speech Accessibility, JSLHR 2024
- Zheng et al., The Interspeech 2025 Speech Accessibility Project Challenge
- Google, Project Euphonia publications and data access page
- Jiang et al. (Google), Learnings from Curating a Trustworthy, Well-Annotated, and Useful Dataset of Disordered English Speech, Interspeech 2024
- Martin et al. (Google), Project Euphonia: Advancing Inclusive Speech Recognition through Expanded Data Collection and Evaluation, Frontiers in Language Sciences 2025
- Tobin and Tomanek (Google), Personalized Automatic Speech Recognition Trained on Small Disordered Speech Datasets, 2021
- Tobin, Tomanek and Venugopalan (Google), Towards a Single ASR Model That Generalizes to Disordered Speech, 2024
- Shor et al. (Google), Personalizing ASR for Dysarthric and Accented Speech with Limited Data, Interspeech 2019 and the Google Research blog post on Euphonia personalization, 2019
- Linguistic Data Consortium, TORGO Database of Dysarthric Articulation, LDC2012S02
- Kim et al., Dysarthric Speech Database for Universal Access Research, Interspeech 2008 and the University of Illinois UA-Speech distribution page
- Turrisi et al., EasyCall Corpus: A Dysarthric Speech Dataset, Interspeech 2021
- Wang et al., CDSD: Chinese Dysarthria Speech Database, arXiv v2 2025
- Alsayegh and Masood, Benchmarking Commercial Speech Recognition and Multimodal Large Language Models on Dysarthric Speech, International Journal of Intelligent Systems 2026
- de Groot et al. (TU Delft), Objective and Subjective Evaluation of Diffusion-Based Speech Enhancement for Dysarthric Speech, 2025
- Hui, Zhang and Mohan, Enhancing AAC Software for Dysarthric Speakers in e-Health Settings: An Evaluation Using TORGO, 2024
- Schu, Janbakhshi and Kodrasi, On Using the UA-Speech and TORGO Databases to Validate Automatic Dysarthric Speech Classification Approaches, 2022
- OpenSLR, LibriSpeech ASR corpus
- Black, Vahratian and Hoffman, Communication Disorders and Use of Intervention Services Among Children Aged 3-17 Years, NCHS Data Brief No. 205, 2015
- NIDCD, Quick Statistics About Voice, Speech, and Language
- ASHA Practice Portal, Dysarthria in Adults
- ASHA-YouGov National Poll: Most Adults With Communication Disorders Report Facing Stigmas, 2024
- Data watch: the Speech Accessibility Project’s size is described three ways. Its research page gives 1,500 hours from about 999 participants, the University of York catalogue record gives about 1,090 hours from about 1,130 participants, and the homepage counts about 2,000 people who had submitted recordings by September 2025 (submitted recordings are not the same as released, transcribed data); we report all three rather than pick one. The ClinicalTrials.gov record lists only Parkinson’s disease and related parkinsonian conditions, so its 485 actual enrollment should not be read as the project total. The JSLHR paper gives the first package’s training set as 190 speakers in its Method section and 191 in its Results. TORGO’s journal abstract describes seven individuals with speech impairments, while the LDC release lists eight dysarthric speakers; we use the LDC figure. The USM paper’s abstract describes about 1,000 hours of disordered speech and less than 1% of training data, while its training table lists 1,158 hours. The CDSD paper describes Part B as 8 speakers in one section and 7 in another after one speaker was removed for annotation errors. In Tobin and Tomanek, the text says 90% of mild speakers reach the target with 50 utterances while the table shows 93%. Microsoft’s 18% to 60% figure is described as accuracy gains without a stated metric or test set. The two Euphonia corpus figures (2024 and 2025) reflect growth, not a conflict. Child prevalence: NIDCD reports 7.2% for voice, speech or language disorders alone, against NCHS’s 7.7% which also counts swallowing.
Last updated: October 24, 2026. We update this roundup quarterly, and the next refresh is expected when the Speech Accessibility Project publishes its next data package and train/test split release.