MOSS-Transcribe-Diarize — MAESTRO mirror

Verbatim mirror of OpenMOSS-Team/MOSS-Transcribe-Diarize (revision 704aa4a9c304e8520be88901e0d1960158ef5b15, 2026-09-02) packaged for the MAESTRO DAW's offline runtime.

MOSS-Transcribe-Diarize (OpenMOSS / MOSI.AI, 0.9B) is an end-to-end transcriber-diarizer: a Whisper-style audio encoder feeds a Qwen3 decoder that writes a recording out as [start][S01]text[end] turns — text, per-turn timestamps and anonymous speaker labels — with no separate diarization or clustering stage. 50+ languages, 128k context, up to ~90 minutes in a single pass, hotword prompting for names and jargon. Technical report: arXiv 2601.01554.

Layout

bf16/               everything MAESTRO downloads (manifest allow_patterns)
  model-00000-of-00001.safetensors   bf16, byte-identical to upstream
  model.safetensors.index.json
  config.json  generation_config.json
  preprocessor_config.json  processor_config.json  chat_template.jinja
  tokenizer_config.json  tokenizer.json  vocab.json  merges.txt
  added_tokens.json  special_tokens_map.json
  sha256.json       per-file digests of everything above
LICENSE             Apache-2.0 (text from the pinned upstream GitHub commit)
README.md           this file

What was changed

  • Nothing in the weights. The single shard is uploaded verbatim and sha256-matched against the upstream LFS metadata.
  • tokenizer_config.json: extra_special_tokens rewritten from the transformers-5.x LIST form to {} so transformers 4.57.x can load it. The three audio tokens (<|audio_start|>, <|audio_end|>, <|audio_pad|>) are also registered in tokenizer.json, so nothing is lost.
  • The upstream remote-code files (configuration_*.py, modeling_*.py, processing_*.py) are not shipped. MAESTRO vendors the GitHub package (OpenMOSS/MOSS-Transcribe-Diarize @ 61bc29cd4120be7b5d3b761b64cd5dff57263642) with two small patches for transformers 4.57.6 and never uses trust_remote_code. To use these weights with the upstream code, load the upstream repo instead.

License

Apache-2.0, as upstream. Copyright the OpenMOSS Team. See LICENSE.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for AEmotionStudio/moss-transcribe-diarize-models

Finetuned
(17)
this model