Blind test
Vozonda, NotebookLM, Open Notebook: you decide, blind.
Two short blind tests on the same sources, default settings on every side: Vozonda against Google's NotebookLM and against the open-source Open Notebook. The script test judges the writing, the voice test judges the voices. Every vote is published; results update every five minutes.
Test 1 · script
Script test: read and pick
What Vozonda really does is write the conversation. Here you read the same part of three episodes about the same source, as text in turns, without voices or names. Which one explains the topic best?
The test needs JavaScript.
Test 2 · voices
Voice test: listen and pick
This one mostly judges the voices. In Vozonda the voice engine is swappable (Kokoro on the CPU, Qwen3-TTS on a GPU, Voxtral in the cloud, more to come) and keeps improving; these clips use the default local Qwen3-TTS voices. We redo this test whenever a better engine becomes the default.
The blind test needs JavaScript.
Method and fine print (both tests)
- Script test: about 150 words from 30 % into each episode, from the start of a turn to a sentence end, shown as turns without speaker names. Turns are found from the audio alone (the two voices told apart by their sound), the same way for every side. Every side is transcribed the same way from its audio (Whisper medium, then punctuation restored by the local model with a word-for-word check), so fillers and punctuation do not give a side away.
- Preference test with three options and a "no difference" option, a multi-choice form of the pairwise tests used to evaluate speech synthesis. A, B and C are shuffled for each visitor; the order is kept on your device, so a reload does not change it. After your vote the page shows which was which.
- 60 seconds from 30 % into each episode, so the intros do not give it away; loudness normalised to −16 LUFS (EBU R 128); same format and bit rate.
- Default settings on every side; Vozonda used local Qwen3-TTS voices and an unedited script. Nothing was regenerated or picked.
- Open Notebook (lfnovo/open-notebook, Docker image v1-latest, October 2026): its default two-host "tech_discussion" profile, the same local model as Vozonda (Qwen3.6-35B) with the model's thinking turned off, as Vozonda runs it. Its first run failed with thinking on (the model returned empty JSON); after that every episode was used as it came. Its scripts are voiced with Vozonda's own voices, so the voice test compares writing and pacing, not voice engines.
- One vote per source, test and visitor. No cookies; duplicates are caught with a salted hash, no IP address is stored with the votes, server logs are deleted after 14 days.
- Raw data: results.json · votes.csv, updated every five minutes.
Results
Results so far
Live from the published votes. Each column counts who was picked; percentages appear once a source has 20 votes with a preference.
Script test
| Source | Vozonda | NotebookLM | Open Notebook | No difference | Votes |
|---|---|---|---|---|---|
| Loading… | |||||
Voice test
| Source | Vozonda | NotebookLM | Open Notebook | No difference | Votes |
|---|---|---|---|---|---|
| Loading… | |||||
Raw data: results.json · votes.csv, updated every five minutes.