johnpattern/safenotes-moss-hq

MOSS-Transcribe-Diarize (OpenVINO) for Safenotes

OpenMOSS's MOSS-Transcribe-Diarize 0.9B (transcription with speaker turns) converted to OpenVINO IR: audio encoder fp16, language model int8, token embeddings as an fp16 array, tokenizer IRs.

This is a format conversion made by Safenotes (a local, on-device meeting notetaker) so the model runs on Intel hardware (OpenVINO). It is not an official release of the original authors, who do not endorse it. All credit for the model goes to them.

Original model and attribution

• MOSS-Transcribe-Diarize by OpenMOSS / MOSI.AI: OpenMOSS-Team/MOSS-Transcribe-Diarize at revision 704aa4a9c304e8520be88901e0d1960158ef5b15, licensed Apache-2.0 (model card metadata and its License section; the repo has no LICENSE file).

This repository is distributed under the same licence: Apache License 2.0. The full licence text is in LICENSE.

Changes made by Safenotes

Converted by Safenotes from the original PyTorch weights: the audio encoder to OpenVINO IR (fp16), the language model to a stateful OpenVINO IR with int8 weights (NNCF), the token embedding table saved as embed_tokens_fp16.npy, the tokenizer to OpenVINO tokenizer/detokenizer IRs, and the Whisper-style 80-bin mel filterbank saved as whisper_mel80_filters.npy. Prompt layout and output parsing are re-implemented in Safenotes (not in this repo). No retraining.

How Safenotes uses it

Safenotes uses it for the better ("HQ") transcript with speakers after the meeting, in 120 s or 60 s chunks (the same files). The Safenotes app downloads the files below from a pinned revision of this repository and checks every file's sha256 before using it. Everything runs on the user's PC; no audio or text leaves it.

Files

10 files, 1,545,432,928 bytes.

file	bytes	sha256
ov/embed_tokens_fp16.npy	311,165,056	e5d4705db1a624a735db063e0333551433ef09f66b6e7c18f6d5ca130c888919
ov/encoder_fp16.bin	624,926,906	dedb812a53fece0e182b6fa79418e9777678d63c3dd829983758014f7786517c
ov/encoder_fp16.xml	952,873	d677db9d839c838f131806c3e30ff8cdd3cfaa5d3bb3e75d25f57cbdf29be0ea
ov/lm_int8.bin	597,735,053	84b43674d341bb456951c312b707e824e1bbd67ed5febf8d682a58c664635c68
ov/lm_int8.xml	2,766,232	9dd51de3f194f65b49ae1c2c01d1601d651763cde408aa9a3bdd2d9459d13be2
ov/openvino_detokenizer.bin	2,189,746	31c1851c1d584d4fdc92f63beb8859be2a11f4e0da4a1381652aed8380a4079c
ov/openvino_detokenizer.xml	12,523	459b6d814f529c5562cd35f7cc503a2b7ad0d3f2045d82f660c1ccb2dcf32dc6
ov/openvino_tokenizer.bin	5,589,140	235eb6678b478a3cb81e5c7b2016b434e9e8c4d505207a51e7aa1967060060f3
ov/openvino_tokenizer.xml	30,951	a010e6c08196cbfa335a1b7e668c238ab644332676208886bdde71e0f1d8c16f
ov/whisper_mel80_filters.npy	64,448	8fd0754ae04711ecd01316f794c302149abbdff4ec6651a49a7382375b03bf0d

No warranty

Provided "as is", without warranties or conditions of any kind, express or implied, including fitness for a particular purpose and accuracy of the output (see the licence). Use at your own risk.