How good is Taleward?
Every tool of this kind makes mistakes, ours too. We tell you how many, how we measure them and how you correct them.
How we measure
- We use public recordings of real sessions, for example a session of The Dark Eye of a good two and a half hours, a Call of Cthulhu session of a good two and a half hours or a Shadowrun session of seven hours. All of them can be watched freely online.
- Before we read a chapter, we write a list of the events that must be in it and sort them into “crucial” and “important”.
- Then we have the same session processed several times and count: how many events are correctly in the chapter? Where is something wrong? A run that goes completely wrong counts too.
- This is not a one-off test. We keep adding more sessions, in other languages too, and measure every improvement against them. This page grows along.
What comes out
| Processing | Session of 2 to 3 hours | Very long evening, 7 hours |
|---|---|---|
| Cloud, Mistral in the EU | a little over 80 % | 75 to 85 % |
| Local, graphics card with 8 GB | just under 70 % | not yet measured with the current version |
| Local, graphics card with 16 GB | being measured | being measured |
| AMD and Intel graphics cards | not supported yet | not supported yet |
Wrong spots: usually zero to three per chapter, often where people at the table spoke unclearly. That is exactly what the note is for, which you use to correct them.
For very long evenings we recommend the cloud or a graphics card with more memory until then.
The number does not tell everything. How well a chapter reads, how long it is and whether the mood is right, we do not measure in percent. That is why we ask the pilot tables.
You have a different graphics card and want to help? Get in touch on Discord. We measure together with you, check the values and add them here.
How you correct
A note in one sentence
You write what is wrong, and Taleward reworks only the paragraphs concerned. In our measurements, one to three rounds were enough, without new mistakes.
Uncertain names
Where a name was recognised with uncertainty, you see it and choose the right one. As long as the recording is still there, you can also have the session transcribed again.
The name help learns along
Names from the chapter plan and the bible already help the next time.
Rewrite the chapter
When a chapter is completely off.
The transcript stays untouched
Corrections are suggestions, you decide.
What we do not know yet
Whether other tools are better or worse, we have not yet measured with the same recording. Until then, we claim nothing about it.
This page is updated as soon as new measurements are available. Last updated: October 2026.
Rather try it yourself? We are looking for pilot tables.