Interview transcription

Why interviews are hard to transcribe responsibly

Interviews are usually the most sensitive audio a team records. A user research session may include a participant’s workplace details, a candidate interview includes salary expectations, and a journalist’s source may be relying on you to protect their identity. Consent forms often promise that recordings will only be handled by the research or hiring team.

Uploading those files to a cloud transcription service can quietly break that promise, or at least add a new data processor you have to disclose. Transcribing by hand keeps the data in-house but costs hours per interview. Local transcription removes that trade-off: the speech model runs on your computer, and the transcript never leaves it unless you export it.

Who uses local interview transcription

The workflow fits anyone who records one-to-one or small-group conversations and has to account for where the audio goes:

• UX and product researchers running customer interviews and usability tests
• Recruiters and hiring managers reviewing candidate conversations
• Journalists and podcasters working with sources and off-the-record material
• PhD students and academics whose ethics approval restricts data storage
• Market researchers and consultants synthesising stakeholder interviews

Worked example: a round of user research interviews

Imagine a product designer who runs eight 30-minute customer interviews over Google Meet and records each one. After each call, she imports the recording into Noteo in Local mode. ffmpeg converts it to the 16 kHz mono audio that whisper.cpp needs, and the transcript appears in her local library with timestamps next to every segment.

She then exports each interview as Markdown. The file includes a small front-matter block with the title and date, followed by timestamped lines, so it drops straight into an Obsidian vault or a research repository. There she tags quotes, clusters pain points across the eight sessions and links each insight back to the exact timestamp. The raw recordings and transcripts stayed on her laptop the whole time.

Local vs cloud for interviews

Both modes are available in Noteo. Pick per project, based on what your consent form and data policy allow:

• Local mode: audio and transcripts stay on your device, which makes it easier to stay within the terms participants agreed to.
• Local mode has no speaker diarization, so interviewer and interviewee are not separated automatically.
• Cloud mode adds speaker recognition, AI summaries and chat over the transcript, but the recording is uploaded for processing.
• Local speed depends on your hardware and the whisper model you choose. Cloud speed does not depend on your laptop.

Compared with the usual alternatives

Manual transcription is private and precise but slow. Human transcription services are accurate but send your files to outside transcribers. Cloud-only AI note takers are fast and label speakers, but uploading is built in. Local transcription with Noteo gives you a fast first draft without a new data processor. You still review it, but you review instead of typing from scratch.

Honest limitations

Local mode requires the Noteo desktop app plus ffmpeg, whisper.cpp and a model file that you install and select. Accuracy depends on the model and the recording: remote calls with poor microphones, overlapping speech and strong accents all produce more errors, so always check quotes against the audio before publishing or citing them. Because there is no local speaker separation, you will need to mark who is speaking yourself if your analysis depends on it.