Transcription Policy as a Latent Variable: Activating Controllable Verbatim ASR with Word-Level Timing
Paper • 2607.18934 • Published • 6
CrisperWhisper 2.0 — MAESTRO mirror
Mirror of nyra labs' CrisperWhisper 2.0 checkpoints (verbatim / intended speech recognition with ~30 ms word timings), laid out for MAESTRO's download-on-demand loader:
| path | upstream | notes |
|---|---|---|
small/ |
nyralabs/CrisperWhisper2.0_small | 0.2B params |
medium/ |
nyralabs/CrisperWhisper2.0_medium | 0.8B params |
turbo/ |
nyralabs/CrisperWhisper2.0_turbo | 0.8B params, 4 decoder layers — fastest |
large/ |
nyralabs/CrisperWhisper2.0_large | 2B params — best quality |
Weights are byte-identical to upstream (sha256-verified at mirror time, see
NOTICE). No converted (CTranslate2) artefacts are redistributed — MAESTRO
converts on the user's machine. Inference code:
nyrahealth/CrisperWhisper (MIT).
License
Non-commercial. The model weights and any outputs they generate are
licensed under the nyra health Non-Commercial Research License (Part B of
LICENSE.md, reproduced verbatim from upstream). By downloading you agree to
its terms. Commercial licenses are available from nyra health GmbH
(licensing@nyra-labs.com). This mirror adds no restrictions.
The CrisperWhisper 2.0 model weights are licensed under the nyra health Non-Commercial Research License, © 2026 nyra health GmbH, Vienna, Austria. All rights reserved. Commercial licenses are available from nyra health GmbH. The accompanying inference code is licensed separately under the MIT License.
Citation
See the upstream model cards and https://arxiv.org/abs/2607.18934.