/

← All benchmarks

TED-LIUM 3

audio-stt · prepared-speech↗ Quelle

49Evidence

saturated

What it measures

Word error rate on TED talk recordings: prepared, clearly articulated speech.

What it does not measure

Spontane Gespraeche, Ueberlappungen, Telefonqualitaet, Fachvokabular.

Die einfachste der Langform-Aufgaben. Aktuelle Modelle liegen unter 3 % — als Unterscheidungsmerkmal kaum noch tauglich.

How the score is composed

Currency100

Discrimination5

Headroom2

Contamination safety55

Gaming resistance100

Methodological rigor80

Reproducibility60

Details

Benchmaxxing risk: low
Difficulty: low
Public test set: yes
Human-verified: yes
LLM-judged: no
Letzter Messwert: 2026-09-28
Skala: 0–100 % WER (niedriger besser)

Best measured models

1elevenlabs-scribe-v22.12 % WER
2transcribe-03Cohere2.23 % WER
3speechmatics-enhanced2.26 % WER
4assemblyai-universal-3-pro2.32 % WER
5parakeet-tdt-v3NVIDIA2.77 % WER
6revai-machine2.92 % WER
7parakeet-tdt-v2NVIDIA3.07 % WER
8whisper-v3OpenAI3.15 % WER
9parakeet-ctcNVIDIA3.23 % WER
10whisper-v3OpenAI3.23 % WER