Fixing NotebookLM Audio Overview When It Generates in the Wrong Language or Stalls Mid-Way

You feed NotebookLM a Japanese PDF, hit Generate, and the Audio Overview that comes back is two voices speaking fluent English. I have run into this exact problem more than once while analyzing user reviews for my own apps, and the first time I heard "Welcome to the deep dive..." after dropping in a stack of Japanese feedback, I almost closed the tab in frustration.

This page walks through what to check when the Audio Overview ignores the language you expected, stalls halfway through generation, or cuts off mid-playback. NotebookLM has become a regular part of how I digest user reviews and research, and the patterns below are the ones I have actually run into — not theoretical edge cases.

Start by sorting out the symptom

"Audio Overview is broken" usually means one of several different things. Identifying the exact symptom up front saves a lot of time.

The common patterns are roughly these:

The source is in your target language but the audio is fully in English
Two languages mix mid-conversation
The progress bar stays on "Generating..." for 30 minutes or more
Generation finishes but playback cuts off after a few minutes
A short "Generation failed" message appears and the job ends

Each of these has a different root cause, so spotting the right one is the first step.

Cause 1: the output language setting is still on English

The setting people overlook most often lives at the notebook level. NotebookLM does not always auto-detect from your sources — it looks at the Output Language dropdown and your Google account language. If those say English, the audio comes out in English even with non-English sources.

To check it:

Open the gear icon in the top-right of NotebookLM
Find the Output Language dropdown
Pick your target language and save
Refresh the page before regenerating the Audio Overview

I once had my Google account temporarily set to English for an AdMob console issue, and that setting silently propagated into NotebookLM. It is worth checking both your account-level language and the per-notebook setting whenever the output looks wrong.

Cause 2: the dominant language of your sources is English

Even with the setting correct, NotebookLM can be swayed by the actual content mix in your sources. Typical scenarios:

A few English research PDFs mixed in with a small amount of non-English material
Web articles pasted in that include English navigation, footers, or lang="en" blocks
OCR-scanned images where English menu text dominates the recognized output
Source documents full of English code comments

NotebookLM weighs the overall text ratio when deciding what language fits best. If the non-English text is a thin minority, the output language can flip to English regardless of your setting. The cleanest fix I have found is to split things into separate notebooks — one per language — rather than trying to wrangle a mixed corpus.

Cause 3: the customization prompt is written in English

The Audio Overview customization box lets you steer focus and tone in natural language. If you write that prompt in English, the model treats it as a language signal and the audio follows. Something like:

Focus on the technical implementation details and provide
concrete code examples for the iOS production environment.

It looks harmless, but it pulls the entire output toward English. Rewriting the same prompt in your target language, and explicitly stating the language at the end, fixes it:

Please focus on the technical implementation details and walk
through concrete code examples for an iOS production setting.
Generate roughly a 10-minute conversation in Japanese.

That last sentence with the language requirement matters more than it should. I now treat "generate in " as a required line in any customization prompt, and the rate of accidental English outputs in my workflow dropped to nearly zero.

Cause 4: the progress bar freezes or generation stalls

When the issue is not language but the generation process itself, the symptom is usually a progress bar stuck in place. The Audio Overview pipeline runs several stages in sequence — summarization, dialogue scripting, then voice synthesis — and any single stage hanging causes the whole job to stop visibly progressing.

Steps that have actually unstuck things for me:

Reload the page after 30 minutes of no movement; the job often retries automatically
Reduce the number of sources to 10 or fewer
Compress each source to under 100 pages
Disable browser extensions, especially ad blockers and translators
Retry in an incognito window

Translation extensions in particular cause problems. Chrome's built-in translation bar can rewrite text in the live DOM, which can confuse the generation request. If you see the translate banner, choose "Never translate this language" before retrying.

Cause 5: the audio cuts off mid-playback

Sometimes generation reports success but playback dies after a few minutes. The culprit is usually one of: a partial server-side write, a corrupted browser cache, or a CDN timeout on the audio file.

Work through it in this order:

Use the download button under the player to save the file locally
Play the downloaded file in a desktop audio player
If the local file plays all the way through, the issue is browser cache — clear it or retry in incognito
If the local file also cuts off, delete the Audio Overview inside the notebook and regenerate

The case I hit personally was a 20-minute Audio Overview where my network blipped during generation. The server marked the job complete but only a few minutes had actually been written. Once you confirm the file itself is short, regeneration is the fastest path.

A pre-flight checklist

Here is a short checklist I run through before triggering an Audio Overview now. It takes about thirty seconds and saves a lot of regeneration time.

Output Language matches the language you actually want to hear
The source set is dominated by the target language; English-only material lives in its own notebook
Customization prompt is in the target language and explicitly states the desired output language
Source count is 10 or fewer, with each source under 100 pages

For the latest official limits and capabilities, the NotebookLM Help Center is the source of truth — limits do shift, especially around the Plus and enterprise tiers.

If you have been losing time to the same issue, I hope this saves you a few rounds. I run into these as part of weekly review analysis for my apps and the occasional research stack for international art submissions, and the patterns above cover almost every miss I have seen so far.