cstr/distil-large-v3-GGUF

Distil Whisper Large v3 (ggml)

ggml conversion of distil-whisper/distil-large-v3 for use with CrispASR and whisper.cpp.

Model Details

• Architecture: Whisper encoder (32 layers, 1280-dim) + distilled decoder (2 layers only)
• Parameters: 756M (49% smaller than whisper-large-v3)
• Speed: 6.3x faster than whisper-large-v3, within 1% WER
• Language: English
• License: MIT

Usage

# Uses the standard whisper backend (auto-detected)
crispasr -m distil-large-v3-q5_0.bin -f audio.wav

Files

File	Size	JFK Result
distil-large-v3.bin	1.5 GB	perfect
distil-large-v3-q5_0.bin	513 MB	perfect

Why Distil Whisper?

• 6.3x faster than whisper-large-v3 (2 decoder layers vs 32)
• Within 1% WER on standard benchmarks
• Same encoder as whisper-large-v3 (32 layers, 1280-dim)
• Drop-in replacement — same ggml format, same CLI flags

Provenance and EU AI Act Art. 53 note

• Upstream model: distil-whisper/distil-large-v3 — published by distil-whisper.
• Upstream licence: mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.