Voice to Text for Researchers:
Transcribe Interviews, Dictate Field Notes

Why Manual Interview Transcription Is the Worst Part of Qualitative Research

Doing your own interview transcription manually is the single most miserable task in qualitative research. The standard ratio is four to six hours of transcription time per hour of recorded audio. A study with twenty 60-minute interviews comes out to 80 to 120 hours of transcription work. That is two to three full work weeks of nothing but typing what you already heard, before any actual analysis has started. Most researchers either outsource this to a paid service, hire an undergraduate, or burn out trying to do it themselves.

The outsourcing route brings its own problems. Services like Rev, Otter.ai, GoTranscript, and Trint upload your participant audio to a cloud server. The provider becomes a third-party data processor in your study, which means it goes into your IRB application, your consent form, your data management plan, and your university's vendor security review. For studies with sensitive populations (clinical, illegal behaviors, vulnerable groups, undocumented status, etc.), several institutions outright prohibit cloud transcription. For everyone else, it adds paperwork and risk.

The local-AI route changes the math. StarWhisper runs OpenAI Whisper on your Windows PC. A 60-minute interview transcribes in roughly five to fifteen minutes on a modern GPU, or 30 to 90 minutes on CPU only. Audio never leaves the laptop, so the IRB story is the same as if you were transcribing manually. The transcript quality is roughly 95 to 98 percent word accuracy on clean audio, which is comparable to a paid human transcriber. The cost is zero (free plan) or $10 per month (Pro plan), versus $1.50 to $2.00 per minute of audio at a human service.

Interview Transcription Workflow Step by Step

The standard workflow for a semi-structured interview study looks like this: conduct interviews on a phone, dedicated recorder, or Zoom/Teams with local recording enabled. Save the audio files into a project folder using your participant identifiers (P01.mp3, P02.mp3, and so on). Open StarWhisper, drag the files into the transcription queue, set the output format to SRT (for timestamped coding) or TXT (for clean reading), pick the language, and start the batch. Leave it overnight for large queues.

The next morning, you have a folder of transcripts ready to import into your coding software. Open each transcript, do a fast review pass against the audio for any obviously wrong segments (industry jargon, proper nouns, fast or accented speech), and fix the typos. This review pass typically takes 20 to 40 minutes per hour of audio, depending on audio quality. That is dramatically faster than the four to six hours of from-scratch transcription, even after the review time.

For studies where audio synchronization matters during coding (which is most studies that include voice tone or pause analysis), use SRT output and import both the SRT file and the original audio into your coding tool. Atlas.ti's audio-document linking, NVivo's synchronized media, and MAXQDA's audio-coding view all work directly with this format. The timestamps let you click any transcript line and jump to the matching audio segment, which is essential for catching tone or hesitation that the transcript flattens.

Field Notes and Memo-Writing by Voice

Field notes are the second-largest writing task in qualitative work. The traditional advice is to write them within a few hours of leaving the field, while detail is still fresh. In practice, this means dictating into a phone on the drive home and transcribing later, or sitting down at a keyboard while exhausted. Dictation into StarWhisper collapses these two steps. Open a Word document on the laptop, press the hotkey, talk for ten minutes about what just happened, save. The words are already in the document; no separate transcription step.

This works particularly well for the "observer notes" component that ethnographers and field researchers keep alongside the "descriptive notes" of what people did and said. Observer notes are where you record your reactions, hunches, methodological notes, and theoretical memos. These tend to come out better when spoken than when typed, because the looseness of speech invites the kind of half-formed observations that later become coding categories. Type-it-up culture in qualitative methods training has historically discouraged this, but the underlying instinct (capture the thought before it slips) maps perfectly to voice.

For follow-up workflows like literature review summarizing, dissertation chapter drafting, or transcribing your own seminar talks, the same dictation loop applies. Press hotkey, speak, edit. Lecture and seminar transcription is a different use case but uses the same audio-file path described above.

IRB and Consent Implications of Local-Only Transcription

This section is general guidance, not legal advice; confirm with your IRB. The short version: local-only transcription removes the third-party data processor from your study, which usually means a simpler consent form, fewer vendor security questions, and less paperwork in the IRB application.

Cloud transcription services typically require disclosure language in the consent form along the lines of "your interview audio will be transmitted to a third-party transcription service (vendor name) for processing. The vendor stores audio for up to N days under their privacy policy at (URL)." Some IRBs also require a signed business associate agreement or data processing agreement with the vendor before approving the protocol. For studies in clinical contexts (especially anything that touches HIPAA-protected information), this can stretch out IRB approval by weeks or block the study entirely.

With local transcription, the consent form language can be much simpler: "Your interview audio will be stored on the researcher's encrypted laptop and transcribed using locally-running software. Audio will not be transmitted to any third party. The recording will be deleted after transcription and verification, no later than X days post-interview." This matches how transcription was done before cloud services existed, and most IRBs are familiar and comfortable with it. The offline processing architecture page goes into more detail on what local-only actually means in StarWhisper's case.

Cloud Mode in StarWhisper does exist as an opt-in feature for non-sensitive audio. It transmits audio to the OpenAI Whisper API and falls under the same considerations as any other cloud service. For IRB-restricted data, keep Cloud Mode disabled.

Exporting to NVivo, MAXQDA, and Atlas.ti

All three major qualitative coding tools accept the output formats StarWhisper produces. The choice between TXT and SRT depends on whether you need timestamps during coding.

NVivo (versions 12 and later)

Open NVivo, go to Import > Documents, select the TXT file. The transcript becomes a document available for coding. For audio-synchronized coding, use NVivo's "Sources > Audios" import, attach the original audio file, then attach the transcript as a "synchronized transcript" using the timestamps from the SRT.

MAXQDA (versions 2020 and later)

Use Import > Texts > From File for plain TXT, or Import > Media > Audio File with attached transcript for the synchronized workflow. MAXQDA reads SRT timestamps natively.

Atlas.ti (versions 9 and later)

Add Documents > Text Document for TXT. For synchronized audio coding, add the audio as a primary document and link the transcript via the "Multimedia" panel using SRT timestamps. Atlas.ti's audio-coding mode lets you select a transcript segment and play the matching audio range directly.

Dedoose and Quirkos also accept TXT and SRT through their standard import dialogs. The interview transcription how-to covers the full pipeline including file naming conventions.

Cost Comparison vs Paid Transcription Services

For a study with 20 one-hour interviews, the dollar comparison runs: Rev AI at $300, Rev human at $1,800, Otter at $17/month for the duration of analysis, and StarWhisper at $0 to $10. The trade-off is that StarWhisper does not produce speaker labels automatically, which the cloud services do. For most semi-structured interviews where the interviewer is one person, this is a small loss; speaker turns are obvious from context.