How AI Transcription Works
Modern AI speech-to-text systems use deep learning models trained on large corpora of spoken language to convert audio to text in real time or near-real time. Accuracy has improved dramatically over the past five years and continues to improve. Leading transcription engines now achieve word error rates below 5% on clear audio in standard accents and languages — comparable to human transcription for most hiring purposes.
Accuracy degrades with background noise, strong regional accents, technical jargon, or non-native speakers. Most enterprise hiring platforms apply noise reduction and post-processing to improve transcript quality, but these limitations remain relevant for candidate populations with diverse linguistic backgrounds or technical roles with specialist vocabulary.
Why Transcripts Matter for AI-Assisted Hiring
Transcript-based analysis is the most defensible approach to AI evaluation of video interview responses. By analysing the content of what candidates say — the quality of examples provided, the structure of reasoning, the relevance of experience to the competency being assessed — transcript-based systems evaluate the substance of responses without analysing physical characteristics.
This is directly relevant to EU AI Act compliance. Systems that analyse facial expressions, vocal tone, or physiological signals to infer emotional state or personality traits are prohibited in employment contexts under the Act. Transcript-based analysis, conducted on the content of responses, operates in a legally distinct and compliant space — provided the other high-risk AI requirements around transparency, human oversight, and audit trails are also met.
Transcripts as an Audit Trail
Under the EU AI Act's requirements for high-risk AI in employment decisions, employers must be able to demonstrate what evidence was used to generate any AI assessment and how that evidence was evaluated. A full transcript of each interview response, retained alongside the AI score, provides this evidential foundation. It enables the hiring team to review the basis for a score, candidates to access the record of their own responses, and regulators to audit decision-making processes after the fact.
Transcripts also have a practical benefit for hiring teams: reviewers can read a response in 30 seconds rather than watching three minutes of video, significantly reducing review time without losing the substance of what was said.
Accuracy and Fairness Considerations
Transcription accuracy is not uniform across all speakers. Research has documented higher word error rates for speakers with certain accents, non-native speakers, and speakers with speech differences. If AI scoring is applied to transcripts with differential accuracy, the downstream scores may reflect transcription quality rather than response quality — a source of unfairness that is difficult to detect without explicit audit.
Employers using AI-transcribed video interviews should ask their platform vendor for transcription accuracy data broken down by speaker demographics, and should ensure that human review is available for candidates who believe a transcript inaccurately captured their response.
How Palantrix uses transcript-based analysis
Every candidate response in Palantrix is automatically transcribed, and all AI scoring is performed on the transcript content — not on audio characteristics, facial expressions, or vocal tone. Hiring managers can read the full transcript of any response alongside the AI score, with the scored traits highlighted in context. Candidates can access their own transcripts through the candidate portal. The complete transcript record is retained as the evidential foundation for the EU AI Act audit trail, and is available for Subject Access Requests under GDPR.
How Team DNA Profiling works →Frequently Asked Questions
How accurate is AI transcription for interview responses?
For candidates speaking clearly in a standard accent with good audio quality, leading AI transcription engines achieve word error rates of 3–5% — accurate enough for meaningful content analysis. Accuracy is lower for non-native speakers, strong regional accents, technical vocabulary, and poor audio conditions. Enterprise platforms typically apply post-processing to improve accuracy, but employers should be aware of these limitations and have a process for handling transcription quality issues.
Can candidates review their own interview transcripts?
Under GDPR, candidates have the right to access personal data held about them, which includes interview transcripts. Employers using AI hiring platforms should ensure the platform can produce transcripts for Subject Access Requests. Some platforms — including Palantrix — provide candidates with direct access to their own transcripts through a candidate portal, which both meets the legal obligation and supports a more transparent candidate experience.
Does interview transcription raise GDPR concerns?
Transcripts are personal data under GDPR — they contain a verbatim record of what a candidate said. They must be stored securely, accessed only by those with a legitimate need, and retained only for as long as necessary for the purpose of the hiring process. The same retention guidelines that apply to video recordings (typically three to six months for unsuccessful candidates) apply to transcripts.
Why is transcript-based AI analysis preferred over multimodal analysis?
Transcript-based analysis evaluates the substance of what candidates say — the content of their examples, reasoning, and experience. Multimodal analysis additionally evaluates physical characteristics: facial expressions, vocal tone, eye contact. The EU AI Act prohibits emotion recognition in employment contexts. More fundamentally, the scientific validity of inferring competency or personality from facial and vocal cues is contested, while content-based analysis is more directly connected to what the interview is designed to assess.
Is manual transcription still used in hiring?
Very rarely in professional hiring contexts at scale. Manual transcription is time-consuming and expensive. AI transcription has reached a level of accuracy that makes it practical for most hiring applications, with human review of flagged passages where accuracy is uncertain. Some high-sensitivity or legally complex contexts — formal investigation interviews, regulated industry assessments — may still use human transcription for quality assurance, but this is not standard practice in mainstream recruitment.