Transcribe research interviews for free – for your thesis or dissertation

Updated 9 October 2026

TranscriptBalance transcribes interviews for your thesis or dissertation for free, with no minute limit, speaker detection, timestamps and export to Word. Your recording stays on your device: Whisper runs right in your browser, and nothing is uploaded.

Why transcribe interviews locally?

Research interviews often contain personal information: names, opinions, sometimes health or workplace details. Many transcription services upload your recording to their servers to do the work. TranscriptBalance transcribes in your browser, with no account. Neither the recording nor the transcript reaches a server, ours included, so no service processes your interviews on your behalf.

What does leave your device: the first time, your browser downloads the models from Hugging Face and the runtime from jsDelivr. These providers see your IP address, but no content. The app also counts anonymously which features are used, never content, and you can switch that off. Details are in the privacy policy and under transcription without upload.

Whether this is enough for your project depends on your university’s rules, your ethics approval, your data management plan and the consent form your participants signed. Check them or ask your supervisor or ethics committee. This is not legal advice.

From recording to transcript, step by step

1. Record: with your phone, a voice recorder or by recording an online interview. Common audio and video files such as MP3, M4A, WAV or MP4 work.
2. Drop the file: drag one or more recordings into TranscriptBalance.
3. Set it up: choose a model, let the app detect the language or set it yourself, and switch on “Speakers”. The app works out the number of speakers, or you set 2 to 6.
4. Transcribe: keep the tab open. With more than 10 minutes of audio, the app asks whether it may send you a notification when everything is done.
5. Name the speakers: rename “Speaker A” and “Speaker B” to, say, “Interviewer” and “P1”, so no real names head the paragraphs.
6. Export: as DOCX, TXT, SRT, VTT, Markdown, JSON or CSV, with or without timestamps and speakers. Several interviews go into one document or into a ZIP file, one file each.

Which model suits your interview?

• Flash (Whisper tiny): runs anywhere, but struggles with accents.
• Fast (Whisper base): fast and solid, weaker on jargon.
• Balanced (Whisper small): very good quality, slower without a graphics card.
• Precise (Whisper large-v3-turbo): the best accuracy. Needs a graphics card with WebGPU and has the largest download.

The app recommends the model that suits your device. Each model is downloaded once, roughly 45 to 760 MB depending on the model and device, and then stays in your browser. For the transcript you’ll quote in your thesis, use the most accurate model your device can handle.

Tips for an accurate transcript

• Sound quality matters most: a quiet room, the microphone close to both people, no background music. Do a short test recording first.
• Set the language: each file is transcribed in one language. If you know it, set it instead of letting the app detect it.
• Proofread: listen to the interview again while you read. Check names, technical terms, numbers and every passage you quote. Replace names that come up in the conversation with pseudonyms as you go.
• Transcription conventions: Whisper usually produces a cleaned-up text. Filler words like “um”, stutters and pauses are often missing. If your conventions require them, add them by hand.
• Check speaker changes: when two people talk at once or sound alike, speaker detection can assign passages to the wrong person.

Long recordings and many interviews

There’s no limit on minutes or files. The app works through several recordings one after another and shows an estimated time remaining. How long it takes depends on your device and the model: with a graphics card via WebGPU it’s much faster than on the processor alone.

Keep the tab open; the app keeps your screen awake meanwhile. If the work is interrupted, for example because the tab was closed, the text so far stays saved in your browser. Drop the same file again and the app carries on after the last finished section. Very long files need a lot of memory, which can get tight on phones and low-end laptops.

If you like, an AI on your device writes an “Interview” summary of 1 to 5 pages: with Chrome AI, or with Gemma, which runs locally in the browser, needs WebGPU and is a one-time download of about 0.8 GB. In file mode nothing is sent to Gemini or any other online service. The summary is AI-generated; your analysis should rest on the transcript.

TranscriptBalance, aTrain or noScribe?

aTrain, from the University of Graz, and noScribe are well-known free programs that also transcribe locally with Whisper, detect speakers and are built for research interviews. The differences:

• Installation: aTrain and noScribe are programs you install. TranscriptBalance runs in the browser, including on computers where you’re not allowed to install software.
• Correcting: noScribe has its own editor where you correct the text while listening. TranscriptBalance has no editor; you proofread the exported Word document.
• Analysis: aTrain offers exports for programs such as MAXQDA, ATLAS.ti and NVivo. TranscriptBalance gives you DOCX, TXT and other standard formats.
• Computing power: an installed program can use your computer’s full power. In the browser, the most accurate model needs a graphics card with WebGPU.
• Beyond files: TranscriptBalance also has a live mode for lectures and meetings.

Which tool fits depends on your computer and the way you work. For a wider view, see the comparison of free transcription tools.