You can deploy this model and start earning money today!

MolyProduction/
whisper-large-v3-turbo-german-ggml-q5_0

Name: whisper-large-v3-turbo-german-ggml-q5_0
Availability: InStock

cstr/whisper-large-v3-turbo-german-ggml

finetune

apache-2.0

Views

No views yet

whisper-large-v3-turbo-german-ggml-q5_0

GGML Q5_0 quantization of primeline/whisper-large-v3-turbo-german, ready for use with whisper.cpp.

The base model was obtained from cstr/whisper-large-v3-turbo-german-ggml (FP16) and quantized using the whisper-quantize tool from whisper.cpp.

Why this model?

The multilingual openai/whisper-large-v3-turbo achieves ~14% WER on German benchmarks. This German fine-tune brings that down to ~5% — a significant improvement for German-language applications. The Q5_0 quantization reduces the file size from 1.62 GB to 574 MB with minimal quality loss.

Benchmark

Evaluated with whisper.cpp on Google FLEURS de_de test split, CPU inference, language forced to de.

Samples	WER
50	5.46%
200	5.69%

Comparison with related models (different datasets, for rough orientation only):

Model	WER (DE)	Size
openai/whisper-large-v3-turbo (multilingual)	~14.2%	1.62 GB
primeline/whisper-large-v3-turbo-german FP16	~4.75%	1.62 GB
this model (Q5_0)	~5.7%	574 MB

The ~1% WER increase vs. FP16 is the expected trade-off for Q5_0 quantization on a large model.

Usage with whisper.cpp

./build/bin/whisper-cli -m ggml-large-v3-turbo-german-q5_0.bin -l de -f audio.wav

Quantization

Quantized using whisper.cpp v0.9.7 (ggml 30c5194c) on Windows x64:

whisper-quantize.exe ggml-model-fp16.bin ggml-large-v3-turbo-german-q5_0.bin q5_0

Model details

Base model: primeline/whisper-large-v3-turbo-german
Architecture: Whisper Large v3 Turbo (fine-tuned for German)
Format: GGML Q5_0
File size: 574 MB
Language: German (de)
License: Apache 2.0

MolyProduction/whisper-large-v3-turbo-german-ggml-q5_0