Whispered MLA Whisper, fp32 checkpoint archive

Full-precision (fp32) training checkpoints for the Whispered MLA Whisper release. These are the lossless master weights the published models were derived from.

Use the fp16 models for inference, not this repo

The inference-ready models are published separately as fp16 (easier to load, half the size):

• cv11 family, matched controls, and the 24-language line: the Whispered MLA collection
• study checkpoints: whispered-mla-experiments
• corpus: Whispered (CC0)

This archive exists for reproducibility, verification, and re-quantization from lossless weights. fp16 is a one-way downcast, so these fp32 originals cannot be reconstructed from the published models.

Contents

Each subfolder is one training run. MLA arms are named *_mla, their matched plain-architecture controls are named *_none.

• Released cv11 masters: rel_{tiny,base,small,medium,large}_{mla,none}
• Released 24-language masters: app24_{mla,none,base_mla,base_none,tiny_mla,tiny_none}
• Study checkpoints: slate, fair, first, mix*, mixmb*, pxc*, anchor_en
• Decode dumps and provenance: *.tgz, model_registry.json, beam_eval.*

Loading

These are raw fp32 training checkpoints. For normal use, load the packaged fp16 model repos linked above (the MLA ones use trust_remote_code=True with transformers==4.46.x). To load an MLA run directly from this archive you would add the MLA modeling files from the public code repository; the plain-architecture control runs (*_none) load with stock transformers as-is.

License and attribution

These weights are derivatives of OpenAI Whisper, which is released under the MIT License. The fine-tuning and MHA-to-MLA conversion contributions in this release are made available under Apache-2.0. Training used the CC0 Whispered corpus (derived from Common Voice 17), plus FLEURS (CC-BY-4.0) for Malay only, with FLEURS audio used for training and never redistributed.

As required by the MIT License, OpenAI's copyright and permission notice is retained:

MIT License

Copyright (c) 2022 OpenAI

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.