NotebookLM Audio Transcription: From Recording to Useful Notes

Gemini Notebook, the current name for NotebookLM since July 2026, accepts audio files as sources and transcribes them automatically, making the spoken content available for questions and citations. The feature is useful for lecture recordings, meeting notes, interviews and podcasts. Tools like fenbs can keep the verified findings, source links and follow-up tasks from each audio session in one place, available to any AI app you use next.

For other AI tools that handle audio, see /blog/notebooklm-alternatives. For Claude's audio capabilities, see /blog/claude-projects-vs-notebooklm.

How Gemini Notebook handles audio files

When you upload an audio file, Gemini Notebook transcribes the speech and adds the resulting text as a source. It does not store the original audio after processing; the notebook works from the transcript text. This means accuracy depends on speech clarity, background noise, accent and technical vocabulary in the recording.

Once the transcript is loaded, you can ask questions about the audio content just as you would with a PDF or a web page. You can ask for a summary, extract action items, compare the audio against other sources, or find where a specific topic was discussed.

What formats and sizes are accepted

Gemini Notebook accepts common audio formats including MP3 and WAV files. Check the current supported formats and file size limits on the Gemini Notebook help page before uploading, as Google updates these limits; third-party articles often carry outdated numbers. If your recording is in a format that is not supported, convert it to MP3 first using a free audio converter.

Long recordings take longer to process and may hit size limits. If you have a long interview or lecture, consider splitting it into logical sections before uploading. Named sections also make it easier to ask the notebook about a specific part of the conversation.

How to upload audio and start working with it

In the notebook's source panel, choose the option to add a file and select your audio recording. After processing, open the source to review the transcript before asking questions. This preview step is important: if the transcript has significant errors, the notebook's answers about the audio will carry those errors through.

• Check the transcript for accuracy before using the source: look for misheard words, run-on sentences and missed speaker turns.
• Technical terms, proper names and non-standard accents are the most error-prone parts of an auto-generated transcript.
• If the transcript has critical errors, edit them in the source before asking questions that depend on accuracy.
• For meeting recordings, ask the notebook to identify action items and decisions after confirming the transcript loaded correctly.

What the transcript looks like and how accurate it is

The transcript is plain text without speaker labels by default. Gemini Notebook may add some structure, but do not expect clean diarization (labeling who said what) automatically. If you need speaker-labeled transcripts, transcribe the audio with a dedicated transcription service first and then upload the resulting text file as a source.

Accuracy varies widely. Clear, single-speaker recordings in standard English with minimal background noise tend to transcribe well. Multi-speaker conversations, phone-quality audio, heavy accents and domain-specific technical terminology all increase the error rate. For high-stakes uses—legal proceedings, medical notes, verbatim quotes—verify the transcript against the original audio before relying on it.

What transcription does and does not give you

Transcription converts speech to searchable text, which is useful. But it does not capture everything: tone, emphasis, pauses and non-verbal cues are lost. A transcript of a negotiation, a lecture or a clinical interview lacks the context that makes the spoken exchange meaningful. Use the transcript as a starting point for analysis, not as a complete record. For high-stakes material, listen to the recording alongside the transcript rather than relying on the text alone.

What to do when audio processing fails

If an audio upload fails or the source loads as empty, check the format and file size first. Convert to MP3 if the format is unusual. If the file is large, trim it to a shorter section and try again. If a short, clear MP3 also fails, try logging out and back in and adding the file again—session state issues can cause intermittent upload failures.

If the source loads but the transcript is too poor to be useful, transcribe the audio externally with a dedicated tool and add the cleaned transcript as a text source. This bypasses the built-in transcription and gives you a usable source even when the audio quality is poor.

How to do this with fenbs

fenbs does not store audio files or transcripts. It keeps the work that comes out of your audio research—the tasks, decisions, verified findings and next steps—available across sessions.

1. Create a free account and open or add the project the recording belongs to.
2. After processing audio in Gemini Notebook, add a reference to the recording and the key findings to the project's Links and listings.
3. When the audio confirms a decision or action item, record it as a task or a Decision in the project so it survives beyond the notebook session.
4. If you switch to a different AI tool, connect it from Integrations so it reads the project notes, including findings from the audio, before continuing.
5. Save progress at the end of each session: note what was confirmed, what is still unclear and where to start next time.