drbaph/CrisperWhisper2.0-GGML

CrisperWhisper 2.0 GGML

Unofficial FP16 GGML conversions of the public CrisperWhisper 2.0 models for CrisperWhisper.cpp.

CrisperWhisper.cpp provides native C++ inference on Windows and Linux with CPU and NVIDIA CUDA backends. Python is not required at runtime.

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

Model files

Model	GGML file	Approximate size	Original model
Small	ggml-crisperwhisper-small-f16.bin	488 MB	nyralabs/CrisperWhisper2.0_small
Medium	ggml-crisperwhisper-medium-f16.bin	1.53 GB	nyralabs/CrisperWhisper2.0_medium
Turbo	ggml-crisperwhisper-turbo-f16.bin	1.62 GB	nyralabs/CrisperWhisper2.0_turbo
Large	ggml-crisperwhisper-large-f16.bin	3.09 GB	nyralabs/CrisperWhisper2.0_large

These are FP16 conversions, not integer-quantized models.

The gated CrisperWhisper2.0_large_pro checkpoint is not included. Access to Large Pro requires approval and is subject to its separate commercial model terms.

Native C++ runtime

Download the native runtime from:

• CrisperWhisper.cpp on GitHub
• CrisperWhisper.cpp v1.1.0 release

Available v1.1.0 packages:

• Windows x64 CUDA for RTX 30, RTX 40, and RTX 50 GPUs
• Windows x64 portable CPU
• Windows x64 AVX2 CPU
• Linux x64 portable CPU
• Linux x64 AVX2 CPU

Linux CUDA is available as a source build.

Model files are downloaded separately and are not duplicated inside the GitHub release archives.

Download

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

Replace small with medium, turbo, or large to download another model.

Windows usage

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

Optional supervised word timestamps and JSON:

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

Linux usage

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

Optional supervised word timestamps and JSON:

./crisper-whisper \
  --model ./models/ggml-crisperwhisper-small-f16.bin \
  --file ./audio.wav \
  --mode verbatim \
  --language en \
  --word-timestamps \
  --json

Audio input

The native CLI accepts WAV, MP3, FLAC, and Ogg Vorbis.

Audio is decoded, downmixed to mono, and resampled to 16 kHz automatically in memory. FFmpeg and a separate preprocessing script are not required.

Transcription modes

Verbatim transcription preserves spoken disfluencies:

crisper-whisper \
  -m models/ggml-crisperwhisper-small-f16.bin \
  -f audio.wav \
  --mode verbatim

Intended transcription produces a cleaner intended transcript:

crisper-whisper \
  -m models/ggml-crisperwhisper-small-f16.bin \
  -f audio.wav \
  --mode intended

Word timestamps

Supervised cross-attention word timestamps are optional:

crisper-whisper \
  -m models/ggml-crisperwhisper-small-f16.bin \
  -f audio.wav \
  --word-timestamps \
  --json

Without --word-timestamps, no alignment pass or additional timestamp model context is created.

Normal transcription keeps GGML Flash Attention enabled. Timestamp alignment uses an additional teacher-forced pass without Flash Attention because the fused Flash Attention path does not expose the decoder-to-audio attention matrix required for supervised alignment.

Long-form timestamp output includes seam ownership, duplicate-prefix removal, and monotonic timestamp correction to prevent duplicated words, overlapping word chunks, and invalid start > end timestamps at chunk boundaries.

Original project and paper

• Official CrisperWhisper repository
• Official Small model
• Official Medium model
• Official Turbo model
• Official Large model
• CrisperWhisper paper
• Native C++ implementation
• whisper.cpp

Limitations

• These files are conversions of the original checkpoints, not separately trained models.
• Small, medium, turbo, and large have different speed, memory, and accuracy characteristics.
• FP16 files require more storage and memory than integer-quantized models.
• Output can differ slightly from Python/Transformers because of backend and numerical implementation differences.
• Speculative draft-model decoding is not currently implemented.
• Large Pro is not included.

License

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

The Small, Medium, Turbo, and Large model weights in this repository remain governed by the Nyra Health Non-Commercial Research License.

Review that license before downloading, using, modifying, or redistributing the converted model files. These models are intended for permitted non-commercial research use.

The original models, architecture, research, and associated intellectual property belong to their respective authors and rights holders.