How accurate it is

We ran Nanoru and five other systems on the same public recordings and scored each against the published transcript. These are our own measurements from November 2026.

Crediting the right person

The charts count, per 1,000 words a system transcribed correctly, how many it put under the wrong person.

On podcast interviews Nanoru put 2.1 words in 1,000 under the wrong person and ElevenLabs Scribe v3 put 1.2. On court hearings Nanoru put 1.9 and Scribe v3 put 8.7. The other four systems were behind Nanoru on both.

The other systems number their speakers instead of naming them. To compare like with like, every system’s voices, Nanoru’s included, were given the name of the person they matched best.

Podcast interviews: words credited to the wrong person, per 1,000
  • ElevenLabs Scribe v3 1.2
  • Nanoru 2.1
  • Whisper large-v3 with DiariZen 2.8
  • Whisper large-v3 with pyannote 3.1 6.0
  • ElevenLabs Scribe v2 8.3
  • NVIDIA Nemotron-3.5-Transcribe 34.1
Court hearings: words credited to the wrong person, per 1,000
  • Nanoru 1.9
  • ElevenLabs Scribe v3 8.7
  • Whisper large-v3 with DiariZen 11.0
  • Whisper large-v3 with pyannote 3.1 23.1
  • ElevenLabs Scribe v2 26.9
  • NVIDIA Nemotron-3.5-Transcribe 66.4

The bars show words credited to the wrong person, per 1,000. Shorter is better.

On recordings nothing was tuned on

Nanoru’s settings were chosen on the recordings above, and it ran there with voiceprints from other episodes of the same shows. Both can flatter it, so we also ran it on 7.9 more hours with no voiceprints and no list of who was in each show.

On these Nanoru put 2.6 words in 1,000 under the wrong person on court hearings and 7.4 on podcasts, ahead of both open-source systems. The other three systems have not been run on these recordings yet.

Podcast interviews: words credited to the wrong person, per 1,000
  • Nanoru 7.4
  • Whisper large-v3 with DiariZen 15.1
  • Whisper large-v3 with pyannote 3.1 21.2
Court hearings: words credited to the wrong person, per 1,000
  • Nanoru 2.6
  • Whisper large-v3 with DiariZen 14.1
  • Whisper large-v3 with pyannote 3.1 21.4

The bars show words credited to the wrong person, per 1,000. Shorter is better.

How we measured

The recordings
9.8 hours in eight recordings: five episodes of NASA’s Houston We Have a Podcast and three US Supreme Court oral arguments. Both publish transcripts that name each speaker. Neither body endorses Nanoru. Eight recordings is a small sample.
The scoring
Every system’s words are matched to the published transcript by their text alone, so no system’s own timings decide the result. Numbers, contractions and filler words are written the same way on both sides.
The other systems
Each ran with its default settings and was told nothing about the recording. Nanoru was given the list of people in each show. Nemotron was run in ten-minute pieces. DiariZen’s published model is licensed for non-commercial use.
What is missing
We have not yet measured panel shows where people talk over each other, broadcast news, radio, sports commentary or live streams.

Test it on your own recordings

Send 20 episodes or 20 hours of recordings. We return them with your presenters named, at no charge.

Ask for a free trial