# My Scottish Accent vs 8 AI Models

The answer to the Scottish accent Short: Voxtral had the highest provisional score of the eight local speech-to-text models, 9.87 out of 10, and was also the slowest.

[HTML](https://lukejbyrne.com/library/guides/agmtkhem_es/)
Content language: en
Author: Luke Byrne
Video published: null
Guide published: 2026-11-07

[Watch the video](https://www.youtube.com/shorts/AGMtKhEM_es)
Video language: English

[Complete guide](https://lukejbyrne.com/library/guides/qpsxna4gtfq/index.md)

## Which one scored highest

Eight local speech-to-text models got the same two-minute recording of Luke’s Scottish accent on an M5 Pro Mac. Mistral’s Voxtral Mini 4B Realtime had the highest overall transcript score, 9.87 out of 10, ahead of NVIDIA’s Nemotron 3.5 Transcribe (8.97) and Canary-Qwen (8.93). It was also the slowest: about five minutes on its first pass over the file, against about five seconds for NVIDIA’s Parakeet.

The score is provisional, because the reference every model was marked against started from Voxtral’s own output. None of the eight spelled Ecclefechan or Milngavie correctly. The complete guide has the full score table, each model’s mistakes, the model links and how to test your own voice.

## The takeaway

Voxtral topped this provisional score table but was by far the slowest, so check the individual mistakes before you choose a model.

## Workflow resources

- [huggingface.co](https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602)
- [research.nvidia.com](https://research.nvidia.com/labs/nemotron/NVIDIA-Nemotron-3.5-Transcribe/)
- [huggingface.co](https://huggingface.co/nvidia/canary-qwen-2.5b)

## Sources and limitations

Short companion prepared 7 November 2026 from the approved 45.8-second English Short and the full guide’s saved test results. One recording, local Mac implementations and first-pass timings; the linked full guide covers the limits in detail.

[Original video](https://www.youtube.com/shorts/AGMtKhEM_es)