edinburghnews.scotsman.com
Human transcribers more accurate than AI, study reveals
A comprehensive study comparing automated speech recognition to human transcription revealed that AI-based tools significantly lag in accuracy.
The study, conducted by Ditto Transcripts, tested eight AI transcription platforms - including Otter.ai, Rev.ai, TurboScribe, Notta, Amazon Transcribe, Sonix, Trint, and Microsoft Speech Recognition.
They found that AI transcription had an average performance of just 62 per cent against human-made transcripts, which rated 99 per cent.
According to the study, fourteen 15-minute audio recordings from legal, medical, and general business settings were submitted to each platform analysis.
The audios featured a numbers of speakers, from single-person monologues to multi-speaker board meetings and congressional testimonies.
The best-performing platform in this AI test was Sonix, with 73 per cent accuracy when adjusted for margin of error, while the lowest was Trint and Amazon Transcribe at just 61 per cent.
What mistakes was AI making?
It was revealed that AI struggled significantly with multiple speakers - even when speech was not overlapping - and was especially prone to errors when processing group speech, such as the Pledge of Allegiance.
Four out of eight platforms failed to transcribe the pledge altogether in one of the tests.
In other instances, single-word responses like “Aye” were routinely misheard or omitted.
Moreover, some platforms introduced entirely fabricated content when faced with unclear audio.
Notta and Amazon Transcribe offered similarly incorrect interpretations - and this tendency to "hallucinate" content underscores AI’s limitations in interpreting context and nuance.
Speaker labelling was another weak point, with several AI-generated transcripts misidentifying the number of speakers or reassigned dialogue inaccurately mid-sentence.
However, when challenged with the same audio bites, Ditto’s in-house human transcribers achieved a 99% accuracy rate, with an error rate of around 1 per cent.
The typical AI error rate hovered around 38%.
The consequences of inaccurate transcripts
Speaking on the shock result, a Ditto transcripts spokesperson, said: “Automated systems struggle with overlapping voices, complex dialogue, and contextual interpretation.
“These are common conditions in legal, medical, and law enforcement settings, where accuracy is non-negotiable.”
The study also documented user experience issues - with some software requiring certain formats to submit the transcription.
These transcription errors, if continued, would lead to serious outcomes.
For example, a misidentified voice in a wiretap led to a year-long wrongful imprisonment; a dosage error caused by outsourced transcription resulted in a $140 million lawsuit against a hospital; and a prisoner was released 33 years early due to a misread sentencing transcript.
According to the research, AI transcription may be useful for note-taking or casual use, but for professional settings, particularly those involving legal, medical, or regulatory documents, human transcription remains the more dependable choice.
The full report, including detailed methodology and comparisons, is available from Ditto Transcripts, a FINRA, HIPAA, and CJIS-compliant transcription provider based in Denver, Colorado.
How each AI platform performed after adjusting for margin of error:
Sonix: 72.61%
TurboScribe: 68.57%
Notta: 67.41%
Rev.ai: 64.86%
Otter.ai: 64.41%
Microsoft Speech-to-Text: 61.93%
Amazon Transcribe: 60.81%
Trint: 60.77%
