|
|
--- |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- de |
|
|
base_model: |
|
|
- primeline/whisper-tiny-german |
|
|
pipeline_tag: automatic-speech-recognition |
|
|
tags: |
|
|
- whisper |
|
|
- faster-whisper |
|
|
- ctranslate2 |
|
|
- asr |
|
|
- german |
|
|
- audio |
|
|
- transcription |
|
|
- speech-to-text |
|
|
--- |
|
|
|
|
|
# Whisper Tiny German - CTranslate2 |
|
|
|
|
|
This model is a CTranslate2-optimized version of [primeline/whisper-tiny-german](https://huggingface.co/primeline/whisper-tiny-german), converted for efficient inference with the [faster-whisper](https://github.com/SYSTRAN/faster-whisper) library. |
|
|
|
|
|
--- |
|
|
|
|
|
## Model Details |
|
|
|
|
|
- **Base Model**: primeline/whisper-tiny-german |
|
|
- **Format**: CTranslate2 with FP16 quantization |
|
|
- **Language**: German (de) |
|
|
- **Model Size**: 72 MB |
|
|
- **Architecture**: Whisper Tiny (39M parameters) |
|
|
|
|
|
--- |
|
|
|
|
|
## Performance |
|
|
|
|
|
Benchmarked on NVIDIA RTX 5090: |
|
|
- **Processing Speed**: 46x real-time on average |
|
|
- **Average Latency**: 0.108s per audio segment |
|
|
- **WER on FLEURS (de_de)**: 29.15% |
|
|
- **CER on FLEURS (de_de)**: 10.11% |
|
|
- **Reliability**: 99% of samples processed without issues (9/862 problematic cases) |
|
|
- **Stability**: No hallucination patterns detected |
|
|
|
|
|
--- |
|
|
|
|
|
## Installation |
|
|
```bash |
|
|
pip install faster-whisper |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## Acknowledgments |
|
|
- Original Whisper model by OpenAI |
|
|
- German fine-tuning by primeline |
|
|
- CTranslate2 optimization framework by SYSTRAN |
