How to Transcribe Interviews with AI: Verbatim Modes, Speaker Labels, and Keyterm Prompting

Manual interview transcription is time-consuming, especially for journalism and academic research where strict accuracy and direct quotes are required. AI-powered speech-to-text tools reduce this process to minutes, allowing professionals to focus on editing and analysis rather than typing.

Choosing Transcript Style

Transcription styles fall into two categories: verbatim and clean. Verbatim transcripts capture every word, including filler words like "um" and repetitions, which is essential for legal, journalistic, and academic standards. Clean transcripts remove fillers and repetitions while preserving original grammar, making them suitable for podcast show notes or blog posts.

ElevenCreative Speech to Text supports both modes via the Scribe v2 model. Verbatim is the default, but users can enable clean transcription in the Playground or by setting the no_verbatim parameter in the API.

Accurate Speaker Labeling

Correct speaker attribution is critical to avoid misquoting. Two methods are available:

• Diarization: Used when a single microphone records all speakers. The model automatically identifies and separates voices, but errors can occur with similar-sounding voices or cross-talk.
• Multichannel transcription: Requires a separate microphone for each speaker. Each audio channel is assigned a specific speaker label, ensuring accuracy even during interruptions or overlapping speech.

Capturing Context and Specifics

Non-speech events such as laughter, sighing, or whispering can alter the meaning of a statement. ElevenLabs tags these events by default using the tag_audio_events parameter, which can be disabled if not needed. This feature is particularly valuable for oral histories and qualitative research.

To handle proper nouns, company names, and industry jargon, keyterm prompting allows users to provide a list of terms before transcription begins. This prevents misspellings and confusion with common words. The ElevenLabs Scribe v2 model supports up to 1,000 keyterms per transcript, while the Scribe v2 Realtime model supports 50 keyterms.

Implementation Options

Users can transcribe interviews via two methods:

1. ElevenLabs Playground: A no-code interface where users upload audio, adjust settings (verbatim, multichannel, keyterms), and edit the final transcript. Exports are available in .txt, .docx, .pdf, .json, .html, .srt, and .vtt formats.
2. ElevenAPI: Integrates transcription directly into custom software, newsroom CMSs, or research tools for automated workflows.