shrisha/CrisperWhisper2.0_large-ggml

CrisperWhisper 2.0 Large - GGML / GGUF

This repository contains GGML/GGUF converted weights for nyralabs/CrisperWhisper2.0_large, optimized for use with whisper.cpp.

Model Summary

• Original Model: nyralabs/CrisperWhisper2.0_large by Nyra Labs
• Architecture: Whisper Large (32 layers, 20 heads, 1280 d_model, 80 mel bins)
• Special Features: High-precision verbatim speech recognition with filler token detection (<vtx>, <ctx>, etc.) and word-level timestamps.
• Conversion Tools: whisper.cpp conversion script (convert-h5-to-ggml.py) with updated multi-token added vocabulary.

Available Files & Checksums

File	Format	Quantization	Size	SHA-256 Checksum
ggml-crisperwhisper2.0-large-q8_0.bin	GGML / GGUF	Q8_0 (8-bit)	~1.54 GiB	f84b2aeced1d89d5dc71a622bfa0077546e849e76a8b181786cfd05de30c1f9a
ggml-model.bin	GGML / GGUF	F16 (16-bit)	~3.09 GiB	9f6843ea92487c371c64f638976663f79062bf73e428eedd8f28efc69bbdf627

Note (2026-08-03): The initial conversion appended the 31 added tokens at the end of the vocab, which broke whisper.cpp's positional special-token arithmetic (Russian LID, translate, French/Spanish recognition). Both files were rebuilt with the corrected token layout via rebuild_ggml.py (see walkthrough.md). The checksums above are for the corrected files.

Usage with whisper.cpp

Command Line Interface

# Clone whisper.cpp
git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp
make

# Run transcription
./build/bin/whisper-cli -m /path/to/ggml-crisperwhisper2.0-large-q8_0.bin -f /path/to/audio.wav -l en

Credits & License

• Original weights by Nyra Labs.
• License follows upstream base model (MIT).