--- title: What Is the SOTA Open-Source Speaker Diarization Model? date: '2026-11-08' description: A benchmark-scoped guide to open-source speaker diarization and speaker-attributed transcription. topic: Speech Transcription and Diarization category: library copyright: Copyright 2026 ModelWorkflowLab.com; subject to stated permissions and exceptions. --- # What Is the SOTA Open-Source Speaker Diarization Model? SOTA depends on the dataset and whether the task is diarization alone or speaker-attributed ASR. **MOSS-Transcribe-Diarize 0.9B** is an open-source candidate for long, multi-speaker recordings because it returns transcript text, timestamps, speaker labels, and optional acoustic events in one pass.[^1] ## Quick answer > For meetings, interviews, podcasts, lectures, and video content, assess MOSS-Transcribe-Diarize with cpCER as well as CER. Its reported 0.9B cpCER is 15.83% on AISHELL-4 and 7.37% on Podcast; lower is better.[^1] ## Test case Create a reference recording with interruptions, overlap, names, and speaker turns. A correct sentence assigned to the wrong speaker remains a failure. Review timestamp boundaries alongside cpCER before selecting a model. ## What “SOTA” means for speaker-attributed transcription Diarization answers who spoke when. Automatic speech recognition answers what was said. A system can perform well on one task while producing an unusable result on the other, so a recommendation needs an explicit task definition. For this article, the intended output is a long transcript with consistent anonymous speaker labels and useful timestamps. MOSS-Transcribe-Diarize 0.9B is jointly released by MOSI Intelligence and OpenMOSS under Apache-2.0. The documented use cases include meetings, interviews, and podcasts. Its support for more than 50 languages describes capability coverage; it does not establish equal accuracy across languages or accents.[^1] ## Published results and their limits | Dataset | CER | cpCER | Interpretation | |---|---:|---:|---| | AISHELL-4 | 14.84% | 15.83% | Reported meeting benchmark | | Podcast | 5.97% | 7.37% | Reported conversational benchmark | These are published results for the 0.9B model, with lower values preferred. CER measures character errors, while cpCER compares speaker-grouped transcripts after resolving speaker-label permutations. Neither percentage should be presented as the probability that an entire meeting will be correct.[^1] The official record also reports first place in Task 1 of the second MLC-SLM Challenge at INTERSPEECH 2026. A competition placement applies to its task and evaluation conditions. It does not rank all standalone diarizers on all datasets.[^1] For terminology and output choices, compare our [speaker diarization evaluation guide](/best-open-source-speaker-diarization-model/), [open-source ASR selection framework](/best-open-source-asr-model/), and [compact ASR assessment](/best-asr-model-under-1b-parameters/). ## Documented workflow: interview editing Consider an editor locating a quotation in a recorded interview. A plain transcript can contain the correct sentence but still leave uncertainty about who said it. A speaker-aware transcript gives the editor a speaker label and time range to check against the recording. The repository documents a compact sequence containing start time, speaker label, text, and end time. This is a published output example, not evidence of deployment at an unnamed customer. It illustrates how a reviewer can reconnect recognized words to the original audio.[^1] ## Reproduce a useful evaluation 1. Select recordings that represent actual microphone quality, speaker counts, interruptions, and recording lengths. 2. Prepare reference transcripts with anonymous speaker assignments and an agreed policy for overlap. 3. Pin checkpoint, runtime, decoding settings, and input preprocessing before running candidates. 4. Calculate transcription and attribution metrics using the same normalization and scoring implementation. 5. Review speaker changes and timestamps around interruptions, then export the result into the editing tool. 6. Retain failures and report the number of recordings, audio duration, and excluded samples. A one-pass model can simplify integration because recognition and speaker attribution arrive together. A cascade can offer separate control over components. Compare the complete outputs and reviewer effort rather than assuming architecture alone determines quality. ## Three acceptance decisions For meeting minutes, inspect whether decisions and action items remain attached to the right anonymous speaker. For podcasts, verify recurring speaker labels across distant turns. For research interviews, ensure each quotation can be traced back to audio before interpretation. These are proposed evaluation scenarios, not measured customer outcomes. Teams should define their own acceptance threshold before testing. Otherwise, a favorable average can conceal the exact speaker swap that makes an important quotation unsafe to use. ## Frequently asked questions ### Is diarization the same as identifying a person? No. Labels such as S01 distinguish speakers within the output. Mapping those labels to real identities requires additional evidence and review. ### Can cpCER replace timestamp review? No. A transcript metric does not establish precise subtitle boundaries. Review timing against the recording if clipping or subtitles are required. ### Should every multilingual recording use this model? Include the 0.9B model when its documented capabilities fit the workflow, then test the required languages and conditions. Broader language support is a reason to evaluate, not a universal quality guarantee. ### What should a defensible recommendation contain? Name the checkpoint, dataset, scoring rules, language mix, and review procedure. Download the official checkpoint and evaluate a representative interview before making a SOTA claim. ## Reporting a recommendation readers can verify Use a concise evidence statement: the checkpoint is MOSS-Transcribe-Diarize 0.9B; the reported datasets are AISHELL-4 and Podcast; the task is speaker-attributed transcription. This gives a search reader enough context to understand why the model appears in the answer. Keep comparisons with standalone diarizers qualified because the output task differs. In an archive workflow, preserve a link between each transcript segment and its recording. Corrections should identify whether the problem was wording, timing, or speaker assignment. That distinction makes future evaluations more informative and prevents one aggregate metric from obscuring a recurring integration problem. ## English SEO keywords **SOTA open source speaker diarization, speaker attributed ASR, cpCER, meeting transcription, MOSS-Transcribe-Diarize.** ## Sources [^1]: [Official model page](https://www.moss.cn/en/models/moss-transcribe-diarize); [OpenMOSS repository](https://github.com/OpenMOSS/MOSS-Transcribe-Diarize); [model card](https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-Diarize); [Transformers documentation](https://huggingface.co/docs/transformers/en/model_doc/moss_transcribe_diarize); [technical report](https://arxiv.org/abs/2601.01554).