CrisperWhisper2.0-large-ONNX

CrisperWhisper 2.0 Large — ONNX

ONNX export of nyralabs/CrisperWhisper2.0_large (verbatim, multilingual speech recognition with word-accurate timing), in the onnx-community / Optimum layout:

• onnx/encoder_model{_fp16,_q4}.onnx — mel encoder (input_features [1, 80, T] → last_hidden_state)
• onnx/decoder_model_merged{_fp16,_q4}.onnx — KV-cache merged decoder (use_cache_branch) with cross_attentions.{i} outputs for DTW word timestamps against the model's supervised alignment heads (generation_config.json)

Decoder prompt layout (the model was trained with mode tags before <|startoftranscript|>):

[verbatim_1][verbatim_2][verbatim_3][verbatim_4][verbatim_5] (<htx> hotwords <ehtx>)? <|startoftranscript|> <|lang|> <|transcribe|> <|notimestamps|>

Verbatim output keeps disfluencies as bracketed event tokens ([UM], [UH], [laughter], [cough], ...). Intended mode uses [intended_1..5] tags instead. See the upstream model card and nyrahealth/CrisperWhisper for details.

License — READ BEFORE USE

These weights are a Derivative Work of nyralabs/CrisperWhisper2.0_large and are distributed under the Nyra Health Non-Commercial Research License (see LICENSE.md in this repository — identical to the upstream license). In short:

• Non-commercial / research use only. No production or operational deployment, and no use from which commercial gain is derived.
• Share-alike. Any further derivative must remain under this same license and may not be re-released under a permissive or commercial license.
• By downloading or using these files you agree to be bound by the terms of LICENSE.md.
• Commercial licensing: contact [email protected].

Conversion: Optimum main_export (task automatic-speech-recognition-with-past, output_attentions=True), fp16 via onnxconverter-common (keep_io_types=True), q4 via onnxruntime MatMulNBitsQuantizer(bits=4, block_size=32, is_symmetric=True).