Transcription for Qualitative Research: Methodology, Tools, and Ethical Considerations

Depends on your analytical framework. For thematic analysis (Braun & Clarke), grounded theory, and most narrative inquiry, intelligent verbatim (removing filler words but preserving all meaning-bearing speech) is standard. For conversation analysis (CA) and discourse analysis, Jeffersonian notation with full pauses, overlaps, and prosody markers is required. For interpretative phenomenological analysis (IPA), verbatim with pauses and emotional markers is typical because meaning-making in delivery matters. For oral history, strict verbatim per Oral History Association (OHA) standards is the norm. AI transcription produces a clean transcript that researchers can use directly for thematic and grounded theory work; for CA and discourse analysis, AI output serves as the starting point for manual Jefferson notation.

Yes, with proper disclosure and consent. Your IRB protocol and informed consent form must name the third-party transcription vendor (or describe the vendor category if vendors may change), disclose the data flow (where audio is processed, where transcripts are stored, retention timeline), and confirm the vendor has appropriate confidentiality terms. Cloud transcription violates protocols that promised 'no third-party sharing' without amendment. If your research involves Protected Health Information (PHI), you need a Business Associate Agreement (BAA) with the vendor — VexaScribe is not currently HIPAA-certified, so PHI research should use HIPAA-compliant alternatives. For EU researchers under GDPR, you need an Article 28 Data Processing Agreement (DPA) and ideally EU data residency. Most IRBs approve AI transcription with these safeguards in place.

Yes, with verification. AI transcription (Whisper Large-v3 and equivalents) achieves 92-97% word accuracy on clean recordings of native English speakers. However, Koenecke et al. (Proceedings of the National Academy of Sciences, 2020) documented significant racial disparities — average WER of 35% for Black speakers vs 19% for white speakers across major commercial ASR systems. For research on populations affected by ASR bias (including Black speakers, AAVE, heavily accented English, non-native English, regional dialects), accuracy gaps are substantial. Best practice: verify every transcript against the original audio before coding, document accuracy limitations in your methods section, and consider human transcription or verification for populations where AI bias is documented.

APA, Chicago, and MLA have no transcription-specific citation entry as of 2026. Current practice (per the APA Style Blog) treats AI tools like software. Example APA 7 reference: 'OpenAI. (2024). Whisper (large-v3) [Speech recognition model]. https://openai.com/whisper'. Sample methods-section disclosure: 'Audio recordings were transcribed using VexaScribe (Whisper Large-v3) and verified verbatim against original recordings by the first author. Transcription accuracy was estimated at approximately 95% on clean audio; all transcripts were corrected before coding. VexaScribe does not train models on user audio per their stated policy.' Field-wide citation standards for AI-assisted transcription are still emerging — expect this guidance to evolve.

Yes — DOCX with consistent speaker labels imports cleanly into all major CAQDAS tools. NVivo 14 (Lumivero) accepts DOCX, RTF, TXT, PDF; for timestamp sync, CSV/TSV formatted transcripts work. MAXQDA (VERBI) has the strongest auto-sync — it imports SRT and VTT natively with timestamps preserved. ATLAS.ti accepts DOCX, RTF, TXT, PDF and supports timestamped transcripts via SRT/VTT or RTF with timecodes. Dedoose imports DOCX, TXT, and spreadsheets. Quirkos imports DOCX, ODT, TXT, RTF, PDF, XLSX. Taguette (open-source) accepts PDF, DOCX, TXT, HTML, EPUB. VexaScribe exports DOCX, TXT, JSON, and SRT from every transcription — DOCX is the most portable for CAQDAS workflows; SRT/VTT enables timestamp sync in MAXQDA and ATLAS.ti specifically. The REFI-QDA standard allows project-level interchange between major CAQDAS tools.

Yes, with proper safeguards. EU researchers transcribing participant audio need: (1) a lawful basis for processing — typically informed consent for research purposes, (2) an Article 28 Data Processing Agreement (DPA) with the transcription vendor, (3) ideally EU data residency for the audio and transcripts, (4) documented cross-border transfer safeguards if data leaves the EU, (5) participant right to withdraw and request data deletion, (6) retention and deletion policy aligned with your ethics approval. VexaScribe stores data in AWS eu-west-2 (London) with TLS 1.2+ encryption in transit and encrypted at rest. For GDPR-strict research, also verify your institutional ethics committee's specific requirements — some EU universities require additional vendor assessments beyond standard DPAs.

Significantly, for research on populations affected by ASR demographic bias. Koenecke et al. (PNAS 2020) found average word error rate of 35% for Black speakers compared to 19% for white speakers across major commercial ASR systems including Amazon, Apple, Google, IBM, and Microsoft. The disparity is attributed to acoustic models trained on insufficient African American Vernacular English (AAVE) data. Graham & Roll (JASA Express Letters 2024) found similar accent-based disparities in Whisper across American vs British/Australian English, native vs non-native speakers, and read vs conversational speech. Implications: document the limitation in your methods section, verify every transcript manually, consider human transcription for affected populations, and report observed accuracy by participant demographic if accuracy is central to your findings.

Verbatim (also 'true verbatim' or 'strict verbatim') captures every utterance — filler words ('um', 'uh'), stutters, false starts, repetitions, non-verbal sounds, partial words. It preserves the full delivery of speech and is required for conversation analysis, IPA, and oral history. Intelligent verbatim (also 'clean verbatim') removes fillers, stutters, and false starts but preserves all meaning-bearing speech and full sentences. It's the standard for thematic analysis, grounded theory, and most applied qualitative research where content matters more than delivery. Clean read (also 'edited transcript') goes further — smoothing grammar and producing polished prose. Use it when summarizing content rather than analyzing speech. Naturalized vs denaturalized terms are used inconsistently across the literature (Oliver, Serovich & Mason 2005 and Bucholtz 2000 use them in reverse senses) — specify your framework when writing methods.

Yes, if your research involves Protected Health Information (PHI) and you're operating under HIPAA. The transcription vendor becomes a Business Associate and a Business Associate Agreement (BAA) is required before any PHI-containing audio leaves your secure environment. VexaScribe is not currently HIPAA-certified and does not sign BAAs — HIPAA-affected research should use a HIPAA-compliant vendor (Rev offers BAAs on Enterprise tier, Verbit offers BAAs, some institutional transcription services). For research involving health information that does not meet the HIPAA PHI threshold (e.g., wellness research without identifiable health data, public health research with anonymized recordings), the BAA requirement may not apply — consult your IRB and institutional compliance office for the specific determination on your project.

Yes, for specific scenarios. Use human transcription when: (1) your research focuses on populations affected by AI bias (Black speakers, AAVE, heavily accented English, regional dialects) and accuracy is central to findings, (2) you're doing conversation analysis with Jeffersonian notation that requires expert transcribers familiar with CA conventions, (3) your audio has heavy crosstalk or overlap (multi-person focus groups, family interviews) where AI diarization fails, (4) you need court-grade or publication-grade verbatim for legal, medical, or archival oral history work, (5) your IRB protocol or institutional policy requires human verification of all transcripts. The hybrid approach works well for many projects: AI for initial transcription of all recordings, human verification for critical interviews you'll quote directly. This typically costs 5-15% of pure human transcription at 95%+ accuracy on quoted passages.