SpeechEditBench - Compositional (Two Edits): leaderboard

SpeechEditBench (2026) tests speech language models on editing a recording from a natural-language instruction in English and Mandarin, scored by automatic checks and an audio judge. This board covers instructions that request two different edits at once, both of which must succeed. The score is joint success in percent: the edit must succeed and the spoken words must be preserved; higher is better.

Metric: Joint success (%): the requested edit is applied (task-specific automatic check) and the linguistic content is preserved (Whisper large-v3 or Paraformer transcript within 10% WER or CER of the expected text); source speech plus a natural-language instruction, English and Mandarin samples, each model through its released editing or conversation interface; two-edit instructions without a speaker component (content, emotion, prosody and acoustic pairs, 200 samples), every requested edit must succeed and the content must be preserved; higher is better. Source: arxiv.org. Saturation forecast: Around January 2028. 8 models tracked.

Top models

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=speecheditbench-compositional-two-edits · How It Works · Data refreshed daily, snapshot 2026-11-13.