MOSS Transcribe Diarize Author Benchmark - Podcast (YouTube subtitle references) - CER: leaderboard

MOSS Transcribe Diarize Author Benchmark - Podcast (YouTube subtitle references) - CER — MOSS Transcribe Diarize author evaluation, arXiv2601.01554v7 Table2 and exact own model-card Evaluation table. Podcast (YouTube subtitle references). CER disregards speaker identities; cpCER minimizes over speaker permutations, percentage errors lower better with no100 cap. Long-form single-pass model supports90minutes; evaluated generation settings are not stated and are not inferred. Scribev1 and GPT-4oTranscribeDiarize identities are explicitly named in paper section4.2, not generic ElevenLabs/GPT4o. AISHELL4 is test split with averaged far-field channels. Podcast references are YouTube subtitles; Movies references professional manual labels. AliMeeting split and channel are not specified. Preserve native source labels and settings. Unbound Doubao and MOSSPro cells, all negative Delta diagnostics and unbound Podcast peers remain retained in held_observations, not erased or invented as public model releases. Metric: CER (%). Source: https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-Diarize

Metric: CER (%). Source: huggingface.co. Saturation forecast: Rough model projection: around 2026. 6 models tracked. Updated 16 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1MOSS Transcribe Diarize Promoss_transcribe_diarize_author_benchmark run4.46
2MOSS-Transcribe-Diarize5.97
3Gemini 2.5 Pro7.38#269
4Doubaomoss_transcribe_diarize_author_benchmark run7.93
5Scribe v18.5
6VibeVoice-ASR27.94

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

Interactive version: theaggregate.ai/benchmark?slug=moss-transcribe-diarize-author-benchmark-podcast-youtube-subtitle-references-cer · How It Works · Data refreshed daily, snapshot 2026-11-16.