File size: 2,452 Bytes
3fc0baf 8fc24da 3fc0baf 8fc24da 3fc0baf 8fc24da 3fc0baf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 | ---
license: apache-2.0
language:
- de
library_name: transformers
pipeline_tag: automatic-speech-recognition
model-index:
- name: whisper-large-v3-turbo-german by Florian Zimmermeister @primeLine
results:
- task:
type: automatic-speech-recognition
name: Speech Recognition
dataset:
name: German ASR Data-Mix
type: flozi00/asr-german-mixed
metrics:
- type: wer
value: 2.628 %
name: Test WER
datasets:
- flozi00/asr-german-mixed
- flozi00/asr-german-mixed-evals
base_model:
- primeline/whisper-large-v3-german
---
# whisper-large-v3-turbo-german — GGUF
GGML conversions and quantisations of [`primeline/whisper-large-v3-turbo-german`](https://huggingface.co/primeline/whisper-large-v3-turbo-german) for use with **[CrispStrobe/CrispASR](https://github.com/CrispStrobe/CrispASR)** or any whisper.cpp-compatible tool.
## Available variants
| File | Quant | Size | Notes |
|---|---|---|---|
| `ggml-model.bin` | F16 | 1.6 GB | Original conversion, full precision |
| `ggml-model-q5_0.bin` | Q5_0 | 548 MB | Good quality/size tradeoff |
| `ggml-model-q4_k.bin` | Q4_K | 453 MB | Smallest, fastest on CPU |
All variants produce correct German transcription on test audio. Q4_K is recommended for CPU deployment.
## Model details
- **Architecture:** Whisper large-v3 encoder (32 layers) + turbo decoder (4 layers)
- **Parameters:** 809M
- **Languages:** German (primary), English
- **Base model:** [`primeline/whisper-large-v3-turbo-german`](https://huggingface.co/primeline/whisper-large-v3-turbo-german)
- **License:** MIT
The "turbo" variant uses only 4 decoder layers (vs 32 in large-v3), making it ~3x faster at inference with minimal quality loss for German.
## Usage with CrispASR
```bash
# Build CrispASR
git clone https://github.com/CrispStrobe/CrispASR && cd CrispASR
cmake -S . -B build && cmake --build build -j8
# Transcribe German audio
./build/bin/crispasr -m ggml-model-q4_k.bin -f german_audio.wav -l de
# With subtitles
./build/bin/crispasr -m ggml-model-q4_k.bin -f german_audio.wav -l de -osrt --split-on-punct
```
## Conversion
Converted from the original HuggingFace model using whisper.cpp's `convert-h5-to-ggml.py`, then quantised with `whisper-quantize`:
```bash
python models/convert-h5-to-ggml.py primeline/whisper-large-v3-turbo-german . models
whisper-quantize ggml-model.bin ggml-model-q5_0.bin q5_0
whisper-quantize ggml-model.bin ggml-model-q4_k.bin q4_k
```
|