CrisperWhisper 2.0 GGML

Unofficial FP16 GGML conversions of the public CrisperWhisper 2.0 checkpoints for CrisperWhisper.cpp, a native Windows/Linux C++ runtime with CPU, NVIDIA CUDA, supervised word timestamps, and long-form stitching. Python is not required during inference.

This is an unofficial community conversion. It is not affiliated with, endorsed by, or maintained by Nyra Health or Nyra Labs.

Files and exact upstream revisions

These are FP16 conversions, not integer-quantized models. Each .bin has a matching .bin.metadata.json sidecar containing its immutable source revision, revision-specific supervised alignment heads, and generation suppression list. Keep both files together; CrisperWhisper.cpp v1.2.0 discovers the sidecar automatically.

The gated CrisperWhisper2.0_large_pro checkpoint is not included. It requires approval and its applicable commercial model license.

Download

hf download drbaph/CrisperWhisper2.0-GGML \
  ggml-crisperwhisper-turbo-f16.bin \
  ggml-crisperwhisper-turbo-f16.bin.metadata.json \
  --local-dir models

Replace turbo with small, medium, or large.

Native runtime

Download source and prebuilt Windows/Linux CPU, CUDA, and Linux ARM64 packages from:

• CrisperWhisper.cpp
• CrisperWhisper.cpp releases

The CUDA v1.2.0 packages contain native code for every CUDA 13 target from SM 86 through SM 121 and retain a CPU fallback. Models are not duplicated in the GitHub release archives.

Usage

Windows:

.\crisper-whisper.exe `
  --model .\models\ggml-crisperwhisper-turbo-f16.bin `
  --file .\audio.wav `
  --mode verbatim `
  --language en

Linux:

./crisper-whisper \
  --model ./models/ggml-crisperwhisper-turbo-f16.bin \
  --file ./audio.wav \
  --mode verbatim \
  --language en

Word timestamps and boundary-safe repeated-speech stitching:

./crisper-whisper \
  --model ./models/ggml-crisperwhisper-turbo-f16.bin \
  --file ./long-audio.wav \
  --longform-strategy chunked-lcs \
  --word-timestamps \
  --json

The CLI accepts WAV, MP3, FLAC, and Ogg Vorbis and automatically decodes, downmixes, and resamples input to mono 16 kHz in memory. No FFmpeg subprocess or preprocessing script is required.

Normal transcription keeps GGML Flash Attention enabled. The optional supervised timestamp path uses an additional teacher-forced no-Flash alignment pass because fused Flash Attention does not expose the required decoder/audio cross-attention matrix.

Validation

Every refreshed model passed native CPU transcription and supervised word-timestamp validation on the bundled JFK sample. Turbo and Large also passed native CUDA inference on an RTX 5090. The chunked-lcs 55-second, five-copy JFK regression returned 110 words across two chunks with zero invalid timestamps and zero adjacent overlaps.

Original project and paper

• Official CrisperWhisper repository
• CrisperWhisper paper
• Native C++ implementation
• whisper.cpp

Limitations

• These are conversions of upstream checkpoints, not separately trained models.
• Output can differ slightly from Python/Transformers because the numerical backend and decoding implementation differ.
• FP16 files use more storage and memory than integer-quantized models.
• Speculative draft-model decoding, forced alignment, and the complete upstream hallucination-recovery stack are not yet implemented in the C++ runtime.
• Large Pro is not included.

License

These conversions do not change or replace the original model license.

Small, Medium, Turbo, and Large weights, configuration, tokenizers, parameters, and model outputs remain governed by the Nyra Health Non-Commercial Research License. Review that license before downloading, using, modifying, or redistributing these converted files. The original models, architecture, research, and related intellectual property belong to their respective rights holders.