In June 2025 I put together a practical guide on AI-assisted transcription for professors of music pedagogy at HfMT Köln — primarily a hands-on introduction to aTrain and noScribe. This post is the methodological companion to that guide: the stuff I could not fit into a workshop handout but that I think matters more than the installation instructions.

Literature and software-status cutoff: 25 October 2026.


The Seduction

AI transcription tools have reached a point where, for clean audio of a single speaker in a quiet room, the output is genuinely good. You load a 90-minute interview, click a button, wait roughly 20 minutes, and get a readable transcript with timestamps and speaker labels. With a first pass that quick, that is an order of magnitude faster than manual transcription: Dresing and Pehl plan five to ten times the interview length for typing with simple rules, proofreading included, which for 90 minutes is 7.5 to 15 hours [[1]]. aTrain’s developers give a fifth of the audio length with an NVIDIA graphics card [[2]]. On a processor alone, their three times the audio length for the most accurate model, and noScribe’s one to three hours per interview hour [[3]], make it roughly two to ten times faster. The appeal is obvious, especially if you are a qualitative researcher working with a backlog of interview recordings.

The two tools I have been evaluating — aTrain (developed at the University of Graz [[4]]) and noScribe (an independent open-source project by Kai Dröge [[3]]) — are designed for local transcription. That can avoid sending recordings to a transcription cloud service. It does not make a project automatically GDPR-compliant: lawful basis, information duties, access control, retention, computer security, and research governance remain the researcher’s responsibility. Dresing and Pehl advise clearing external providers with the university’s data-protection officer [[1]].

Both tools are built on OpenAI’s Whisper models, which are — despite the name — open-source and run offline: OpenAI released the code and the model weights under the MIT licence [[5]]. Both run them through faster-whisper, a reimplementation built for faster inference [[6]] [[2]] [[3]]. They differ in interface philosophy, feature depth, and what methodological commitments they make visible.

But the seduction is the problem. The speed and cleanliness of the output makes it easy to treat the transcript as a neutral record rather than as a construction. It is not. Every transcription is an act of interpretation; the methods literature describes transcription as “a process that is theoretical, selective, interpretive, and representational” (Davidson, as quoted by Wollin-Giering et al. [[7]]). An AI transcription is an act of interpretation performed by an algorithm that does not know what your research question is.


Why This Is a Grounded Theory Problem Specifically

In grounded theory — whether you follow the Strauss and Corbin tradition [[8]] or the constructivist reformulation by Charmaz [[9]] — the researcher is not a passive recorder of data. The analytical process begins with the first moment of contact with the material. Coding, memo-writing, constant comparison, and theoretical sampling all assume that you are working with data that you have genuinely engaged with and that reflects choices made with your research question in mind.

Transcription is the first of those choices. What counts as a pause? Do you mark hesitations and self-corrections? Do you capture overlapping speech? Do you note emphasis, speed changes, or trailing-off? The answers to these questions are not neutral. They are determined by what level of analysis you intend. A thematic analysis of interview content needs something different from a conversation analysis of turn-taking, which needs something different from a discourse analysis attending to hedges and disfluencies. GAT 2, the system of German conversation analysts and interactional linguists, says its minimal transcript suffices for content analysis of interviews but rarely for conversation-analytic work [[10]].

When you transcribe manually, you make these choices explicitly or implicitly, but you make them. When you delegate to an algorithm, the algorithm makes them — according to its training data and its default settings — and then presents you with output that looks authoritative.

The risk is not that AI transcription is inaccurate (though it sometimes is). The risk is that it is selectively accurate in ways you did not choose and that those choices shape what you subsequently see in the data.


What the Tools Actually Do

Under their interfaces the two tools contain the same stages (Fig. 1): voice-activity detection, Whisper decoding in 30-second windows, pyannote speaker turns with one speaker per decoded segment, and, in noScribe, pause marking [[2]] [[3]] [[6]] [[11]]. Each stage can drop material, and three can add it.

Figure 1. Stages both tools contain; noScribe computes speaker turns before decoding. Drops (−) and additions (+) from the Whisper paper and model card [[11]] [[5]], the tools’ code and documentation [[2]] [[3]], and [[7]] [[1]] [[10]].

aTrain

aTrain is the simpler of the two. Version 1.4.1 (January 2026) runs on Windows (Microsoft Store or installer), with beta builds for Apple Silicon Macs and a Debian package built for Ubuntu 24.04-based systems [[2]]. The interface has essentially one methodological decision point after you load your file: whether to activate speaker detection, optionally with the number of speakers. The other settings — model, language and, in an advanced panel, GPU use, compute type, temperature and a free-text initial prompt — set no transcription convention by default, though the initial prompt is the lever noScribe’s disfluency switch pulls [[6]]. Output formats are plain text with timestamps, SRT subtitle files, and — most useful for researchers — direct QDA exports for MAXQDA, ATLAS.ti, and NVivo with synchronised audio-timestamp links.

What aTrain does not do: it does not mark pauses. It does not detect disfluencies (the ähms, uhs, self-interruptions, false starts). It does not detect overlapping speech. (Release candidates for version 1.5, October 2026, add a verbatim model [[2]].) It produces clean, semantically coherent transcripts — which means it actively smooths what you gave it. If a speaker says “well — I mean — it was, I think it was more like — yeah, complicated”, aTrain will probably give you something closer to “I think it was complicated”. The hesitation structure disappears.

The smoothing is probably learned: Whisper was trained to reproduce the transcripts that came with 680,000 hours of internet audio, “without any significant standardization” [[11]]. Wollin-Giering et al., comparing Whisper with eight other tools, conclude that automatic transcription has “certain biases towards finishing sentences or deleting filler words” [[7]].

For a thematic interview study where you are interested in what people said about a topic, this is probably fine. For any analysis where how something was said is part of the data — pace, repair, emphasis, epistemic hedging — aTrain is erasing data you need.

noScribe

noScribe is more complex in almost every dimension. It runs on Windows, macOS and Linux; current versions need an Apple Silicon Mac, and the last Intel build is version 0.6 [[3]]. The interface exposes a meaningful number of configuration decisions (version 0.8, October 2026):

  • Mark Pause: off, or marked from a 1-, 2- or 3-second threshold with one dot per second — (.), (..), (...) — and written out from ten seconds, (10 seconds pause). Pauses are measured only in the gaps between the text segments the decoder returns, after trimming segment edges to voice-activity pauses, and rounded to whole seconds [[3]].
  • Speaker Detection: automatic count, fixed count, or disabled
  • Overlapping Speech: experimental detection, marked with //double slash//
  • Disfluencies: off or on — on aims to keep ähm, äh, self-corrections and false starts. The switch chooses the example sentence placed in front of the decoder: on, Äh, das ist, es ist, ähm, nicht so einfach (“Uh, that is, it is, um, not so simple”); off, the fluent Das ist nicht so einfach. The documentation calls it “more a ‘recommendation’ to the AI model, not a hard on/off switch” [[3]] [[6]].
  • Timestamps: by speaker turn or every 60 seconds

It also has an integrated editor (noScribeEdit) with synchronised audio playback: playback starts at the cursor, and the text selection follows the audio [[3]]. This is the single most useful feature for post-transcription review, and aTrain does not have anything equivalent.

The configuration complexity is not gratuitous. It reflects the fact that different methodological frameworks require different transcription conventions — but the conventions noScribe writes match content-oriented interview research, not conversation analysis. One dot per second and // around an overlapping interjection are the same symbols as in the optional extensions of Dresing and Pehl’s content-semantic transcription rules, in which (.), (..) and (...) stand for about one, two and three seconds [[1]]. In GAT 2 the same (.) is a micro-pause of up to about 0.2 seconds, and pauses up to one second have three further symbols [[10]]. Read with GAT eyes, a noScribe (.) is at least five times longer than the pause its symbol denotes, and the sub-second pauses GAT records do not appear at all. For GAT 2 transcripts, noScribe’s FAQ recommends exporting WebVTT and continuing in EXMARaLDA [[3]].

The disfluency option covers part of GAT 2’s lowest level, the Minimaltranskript (“minimal transcript”), which records filled pauses such as äh and ähm along with overlaps, pauses, audible breathing and laughter [[10]]; noScribe marks no breathing or laughter. The choices you make in the interface are methodological choices, not just technical preferences.

In an earlier version of this section I compared the disfluency option to a “GAT2-Light” transcription, a level GAT 2 does not have, and said the pause notation maps onto conversation-analytic conventions. That was wrong in exactly the way this post is about: the same symbol, read under another convention, records a different pause.


The Normalisation Problem

Both tools perform what I would call normalisation: they produce transcripts that read more fluently than the original speech. This is a feature from a usability standpoint and a methodological liability from a qualitative research standpoint. Wollin-Giering et al. found in 2023 that all available automatic tools “only offer transcripts on a level of detail between verbatim and gisted” and are therefore “not adequate for purposes of conversation analysis or similar approaches” [[7]].

Specific failure modes I observed in evaluation:

Compound word errors (more pronounced in noScribe for German): VR-Brille (“VR headset”) transcribed as Brille VR, proper nouns mangled, domain vocabulary rendered phonetically. In music research contexts this is particularly salient — instrument names, notation terms, composer names, and genre vocabulary are all potential failure points.

Speaker detection overcounting: both tools, when speaker detection is active, tend to identify more speakers than are present. A two-person interview with one hesitant speaker may generate three or four speaker labels. Manual correction is required. Both projects advise giving the number of speakers when you know it [[2]] [[3]].

Acoustic transcription: noScribe occasionally produces what my guide calls lautliche Transkriptionen — phonetic renderings rather than semantic ones. A speaker saying Beamer (data projector) may be transcribed as Bima. In a word-error count this is an error like any other, but an informative one: the output follows the sound of the word instead of resolving it into the word the speaker meant. noScribe’s documentation describes the same pattern for unusual names and slang, spelled “sometimes almost phonetically” [[3]]. For music researchers studying how non-specialist participants talk about technical equipment, this is interesting. For most interview research, it requires correction.

Pause and overlap reliability degrades with audio quality: both tools perform well on clean, close-mic mono recordings of single speakers in quiet rooms. Introduce a second speaker, ambient noise, variable recording distance, or a phone recording, and accuracy drops substantially. This matters specifically for music interview research, where the interview setting is often a rehearsal room or performance space rather than an acoustic booth.

The published numbers point the same way. Whisper large-v2 reached a word error rate of 4.5 per cent on the German part of FLEURS [[11]], Wikipedia sentences translated from English and read aloud [[12]]: a word accuracy of 95.5 per cent. On a five-minute clip of a German research interview with fast speech, frequent speaker changes and unfinished sentences, the same model’s word accuracy was “above 70%” [[7]]. Both tools now default to large-v3-turbo models [[2]] [[3]], so neither number need transfer. Dresing and Pehl name strong dialect, several people talking at once, and canteen or restaurant noise as settings where automatic recognition works badly or not at all [[1]].

Whisper can also insert text. Its developers list “complete hallucination where the model will output a transcript entirely unrelated to the actual audio” and getting “stuck in repeat loops” among its remaining errors [[11]]. In a 2023 test of OpenAI’s Whisper API on English speech, about 1 per cent of transcriptions contained invented phrases or sentences, more often for speakers with longer non-vocal stretches [[13]]. The study concerned aphasia, so the rate for hesitant interviewees is open. noScribe added voice-activity filtering in version 0.5 against hallucination and looping [[3]]; neither tool rules them out.


A Methodological Comparison, Not a Feature List

The useful comparison between aTrain and noScribe is not technical — it is about which methodological contexts each is suited to.

Research contextToolWhy
Thematic/content analysis, single speakeraTrainSpeed, simplicity, adequate accuracy, QDA export
Grounded theory with attention to epistemic hedgingnoScribe + disfluenciesAsks the model to keep the hesitation structure that carries methodological information
Conversation analysisNeither, or noScribe as starting pointCA requires phonetic detail neither tool reliably produces
Large corpus, initial open codingaTrainVolume and speed outweigh detail at early stages
Interpretative phenomenological analysis (IPA)noScribeThe pause and disfluency data is IPA-relevant
Teaching transcription as a research practiceBothSee below

The last row deserves its own section.


Using Both Tools to Teach About Transcription

The most pedagogically valuable use of these tools is probably not producing transcripts — it is using them to make the constructed nature of transcripts visible to students.

A simple exercise: take a three-minute excerpt of an interview recording. Have students transcribe it manually according to whatever convention the course uses. Then run the same excerpt through aTrain and noScribe with different settings. Compare the three or four resulting transcripts in a seminar discussion.

The differences that emerge are not about which transcript is “correct”. They are about what each transcript makes visible and what it hides. The aTrain transcript will be clean and readable. The manually-produced transcript will have annotation that the students chose based on what struck them as relevant. The noScribe transcript with disfluencies enabled will look noisy. All three are representations of the same three minutes of speech.

Questions that come out of this reliably: Why did the student who transcribed manually mark that particular pause? What did the student not mark that the software did? What did the software produce that the student did not hear? What does the “cleaner” transcript lose?

This is the entry point to a genuinely grounded-theory-relevant conversation about data construction: the transcript is not the data. The transcript is a representation of the data made according to principles that should be theoretically motivated, and those principles should be stated explicitly in the methods section.


What These Tools Cannot Replace

The document I prepared for the HfMT professors ends with a sentence I want to quote directly from the German, because it is the methodological core of the whole thing:

Automatisierung ersetzt nicht das Nachdenken über Daten. Automation does not replace thinking about data.

More precisely: the algorithm makes decisions about what counts as a pause, what counts as language, whose voice counts as a separate speaker — without knowing what is scientifically relevant. It does not know that the half-second hesitation before a particular word is the most important moment in the interview. It does not know that the overlapping “mm-hm” is a data point for your analysis of how the interviewee manages discomfort. It does not know that the repeated self-correction in the middle of a sentence about teaching practice is where your emerging category is.

You have to know that. And you only know it if you have been in enough contact with the material to have developed theoretical sensitivity. Strauss and Corbin name several sources of it; one is the analytic process itself, the iterative relationship between data collection, coding, and theoretical development in grounded theory [[8]].

AI transcription tools save the hours of typing. They do not and cannot substitute for the analytical engagement that makes a grounded theory study produce knowledge rather than a theme list.

Use them. But use them knowing what they are doing.


Practical Summary

  • aTrain: one-click, local, good QDA integration, appropriate for thematic analysis. No disfluencies, no pauses, no overlap detection. Versions: Windows (Microsoft Store or installer); beta builds for Apple Silicon Macs and Debian; stable release 1.4.1 (January 2026). Check the project’s current release record before relying on a feature.
  • noScribe: more complex, highly configurable, disfluency and pause options, integrated audio-sync editor, appropriate for grounded theory and discourse-oriented work. More demanding to set up.
  • Neither tool is appropriate as a black-box solution for conversation analysis or prosodic research.
  • Both tools require manual post-processing. Estimate correction time at roughly 20–40% of the original interview length for clean recordings with a single speaker; more for multi-speaker or suboptimal audio. Published figures are higher: for the first seven interviews of a project that conducts expert interviews in German and English, Wollin-Giering et al. report an average of 3 hours 30 minutes of correction per interview hour with Whisper [[7]].
  • In teaching: the exercise of comparing manual, aTrain, and noScribe transcripts of the same excerpt is more pedagogically valuable than any of the transcripts individually.

Why grounded-theory techniques would have served an earlier project of mine better than design thinking: They Told Me Not to Use Design Thinking. They Were Right.


References

[1] Dresing, T., & Pehl, T. (2024). Praxisbuch Interview, Transkription & Analyse: Anleitungen und Regelsysteme für qualitativ Forschende (9th ed.). Eigenverlag, Marburg. https://www.audiotranskription.de/ressourcen/praxisbuch/

[2] aTrain developers, Business Analytics and Data Science-Center, University of Graz (2025–2026). aTrain [Computer software], versions 1.3.1 (May 2025) to 1.4.1 (January 2026), FAQ, and release notes for 1.5.0-rc4 and rc5 (October 2026). https://github.com/aTrainTranscription/aTrain

[3] Dröge, K. (2023–2026). noScribe [Computer software], versions 0.6 (February 2025) to 0.8 (October 2026), and documentation. https://github.com/kaixxx/noScribe; https://noscribe.de

[4] Haberl, A., Fleiß, J., Kowald, D., & Thalmann, S. (2024). Take the aTrain. Introducing an interface for the Accessible Transcription of Interviews. Journal of Behavioral and Experimental Finance, 41, 100891. DOI: 10.1016/j.jbef.2024.100891

[5] OpenAI (2022–2025). Whisper [Computer software, model weights and model card], latest release v20250625. https://github.com/openai/whisper

[6] Klein, G., and contributors (2023–2026). faster-whisper [Computer software]. SYSTRAN. https://github.com/SYSTRAN/faster-whisper

[7] Wollin-Giering, S., Hoffmann, M., Höfting, J., & Ventzke, C. (2024). Automatic transcription of English and German qualitative interviews. Forum Qualitative Sozialforschung / Forum: Qualitative Social Research, 25(1), Art. 8. DOI: 10.17169/fqs-25.1.4129

[8] Strauss, A., & Corbin, J. (1998). Basics of Qualitative Research: Techniques and Procedures for Developing Grounded Theory (2nd ed.). SAGE Publications.

[9] Charmaz, K. (2014). Constructing Grounded Theory (2nd ed.). SAGE Publications.

[10] Selting, M., Auer, P., Barth-Weingarten, D., et al. (2009). Gesprächsanalytisches Transkriptionssystem 2 (GAT 2). Gesprächsforschung – Online-Zeitschrift zur verbalen Interaktion, 10, 353–402. PDF. English version: Couper-Kuhlen, E., & Barth-Weingarten, D. (trans.) (2011). A system for transcribing talk-in-interaction: GAT 2. Gesprächsforschung, 12, 1–51. PDF

[11] Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., & Sutskever, I. (2022). Robust speech recognition via large-scale weak supervision. arXiv preprint arXiv:2212.04356. https://arxiv.org/abs/2212.04356

[12] Conneau, A., Ma, M., Khanuja, S., Zhang, Y., Axelrod, V., Dalmia, S., Riesa, J., Rivera, C., & Bapna, A. (2022). FLEURS: Few-shot learning evaluation of universal representations of speech. arXiv preprint arXiv:2205.12446. https://arxiv.org/abs/2205.12446

[13] Koenecke, A., Choi, A. S. G., Mei, K. X., Schellmann, H., & Sloane, M. (2024). Careless Whisper: Speech-to-text hallucination harms. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT ‘24), pp. 1672–1681. DOI: 10.1145/3630106.3658996


Changelog

  • 2026-01-20: Updated the aTrain reference to the published form: Haberl, A., Fleiß, J., Kowald, D., & Thalmann, S. (2024), “Take the aTrain. Introducing an interface for the accessible transcription of interviews.”
  • 2026-07-11: Corrected the privacy boundary: local processing can reduce external transfer, but does not by itself establish GDPR compliance; removed stale version numbers.
  • 2026-10-03: Corrected the noScribe author in the references from “Kailscheuer, K.” to Kai Dröge (“Dröge, K.”, per the project README); restored the published capitalisation of “Accessible Transcription of Interviews” in the aTrain title; removed “GDPR-compliant” from the aTrain summary bullet, which contradicted the post’s own point that local processing does not by itself establish compliance.
  • 2026-10-25: Corrected two transcription-convention claims against the GAT 2 paper (Selting et al. 2009) and Dresing and Pehl’s rules, and said so in the body: “GAT2-Light” is not a GAT 2 level (GAT 2 defines minimal, basic and fine transcripts), and noScribe’s pause notation does not map onto conversation-analytic conventions; its symbols match Dresing and Pehl’s, where (.) is about one second, while GAT 2’s (.) is a micro-pause of up to 0.2 seconds. Added that noScribe measures pauses only between decoded segments and recommends EXMARaLDA for GAT 2 work. Corrected the disfluency option from “captures” to a choice between two example sentences given to the model, which its documentation calls a “recommendation”. Updated the software status from 11 July to 25 October 2026: aTrain also asks for model, language and an optional speaker count and has advanced settings, including an initial prompt (1.4.1), with beta builds for Apple Silicon and Debian, and a verbatim model in the 1.5 release candidates; noScribe’s last Intel-Mac build is 0.6, so “Apple Silicon and Intel” no longer holds, and it added voice-activity filtering against hallucination in 0.5. Replaced “not an error in the conventional sense” for Bima with an error in any word-error count, and “Interpretive” with “Interpretative phenomenological analysis”; narrowed the Strauss and Corbin sentence on theoretical sensitivity. Repaired the garbled summary. Added sourced context: a manual-transcription baseline and hardware-bound run times for the 20-minute example; Dresing and Pehl’s data-protection advice; Davidson on transcription as interpretation; GAT 2 on what the minimal transcript suffices for; the shared stages (new Fig. 1); the speaker-count remedy; Wollin-Giering et al. (2024) on verbatim versus gisted transcripts, word error rates and correction times; and Whisper hallucination (Radford et al.; Koenecke et al. 2024). Converted the references to a numbered list; updated Dresing and Pehl from the 8th (2018) to the 9th edition (2024), whose symbols are quoted, and the aTrain and noScribe software references to the versions checked.