Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-2343
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-03T00:02:37.924278Z
status: published
lens: trace
sector: health
headline: Differential Accuracy of Semantic Versus Phonological Speech Error Transcription in Aphasia: Comparing Large Language Model (Whisper) to Clinician (Speech-Language Pathologist) Performance
dek: Purpose: Patient speech samples provide insights into cognitive and linguistic status, yet their utilization has been limited by labor-intensive manual transcription. Advances in automated speech recognition (ASR) with large language models (LLMs) potentially save time, but performance on impaired speech, such as due to poststroke aphasia, remains understudied.
Method: We transcribed 4,535 trials of the Philadelphia Naming Test (PNT) administered to 26 people with aphasia (PWA) using OpenAI's Whisper speech-to-t…
gain_title: Whisper transcribed a 175-trial aphasia naming test in about 2.3 minutes, versus about 180 minutes for a speech-language pathologist, and matched clinician transcripts on 92.22% of semantic errors.
problem_title: Whisper matched clinician transcripts on only 23.05% of trials with phonemic errors and 75.52% overall, below the over 95% match of trained humans, with more hallucinations in more severe aphasia.
trace_subject: Whisper transcription of speech errors in people with aphasia on the Philadelphia Naming Test
gain_reading: Whisper transcribed a 175-trial aphasia naming test in about 2.3 minutes, versus about 180 minutes for a speech-language pathologist, and matched clinician transcripts on 92.22% of semantic errors.
gain_evidence: transcriptions by SLP required 180 ± 40 min versus 2.3 ± 0.5 min by Whisper | 92.22% on semantic errors
problem_reading: Whisper matched clinician transcripts on only 23.05% of trials with phonemic errors and 75.52% overall, below the over 95% match of trained humans, with more hallucinations in more severe aphasia.
problem_evidence: 23.05% for trials with phonemic errors by PWA | Whisper achieved an overall match rate of 75.52%
quick_read: Researchers ran 4,535 Philadelphia Naming Test trials from 26 people with poststroke aphasia through OpenAI's Whisper and compared the output with transcripts by an experienced speech-language pathologist. The AI took a few minutes per test where manual work took hours.

The results suggest automation could ease a major bottleneck in aphasia clinical and research work, but only for some patients and error types. Whisper's weak handling of phonemic errors and its rising hallucination rate with severity leave open whether it is safe for broader clinical use.
limitation: Whisper reliably transcribed semantic but not phonological errors, so automation is best suited to fluent aphasia types with fewer phonemic distortions; the sample was only 26 people with aphasia.
tag: Dual reading
key_points: The study analysed 4,535 naming-test trials from 26 people with aphasia. | Manual transcription took 180 ± 40 minutes per 175-trial test versus 2.3 ± 0.5 minutes for Whisper. | Whisper's match rate was 88.83% on correct trials, 92.22% on semantic errors and 23.05% on phonemic errors. | Match and hallucination rates correlated with Western Aphasia Battery-Aphasia Quotient scores. | Low-frequency words produced more mismatches than medium- or high-frequency words.
rundown: Patient speech samples reveal cognitive and linguistic status, but labor-intensive manual transcription has limited their use. The study tested whether a large language model-based speech recognizer could handle impaired speech after stroke.

Whisper's output was compared with gold-standard transcription by an experienced speech-language pathologist. Agreement was analysed by error type, patient severity and word frequency.

The authors conclude that automation is best suited to fluent aphasia types with fewer phonemic distortions, and describe the work as a benchmark of LLM transcription for impaired speech.
sources:
- peer_reviewed | Journal of Speech, Language, and Hearing Research | https://doi.org/10.1044/2026_jslhr-25-00882 | 2026-11-02
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

e28aa915d24b8077af4dfeae1978688cea2341474023372eb652d888981c697e

previous

0000000000000000000000000000000000000000000000000000000000000000