GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Superkon3000
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Superkon3000
Model Tree
openai/whisper-medium
Input Modalities
Output Modalities
Supported Functionality
openai/whisper-medium (769M), German decoding, VAD, repetition mitigationq_proj, k_proj, v_proj, out_proj, fc1, fc2StiAna (452 synthetic consultations), partly voice-cloned from consenting
actors' voices. Scripts and records are public under CC BY 4.0 in the
conversim repository; the audio is not,
as the speakers did not consent to its distribution.On StiReal, 28 held-out human-recorded consultations (2.5h audio, 22.5k reference words):
| Metric | Baseline | Adapted |
|---|---|---|
| nWER | 4.75 | 2.84 |
| Terminology accuracy | 88.75 | 97.04 |
| SeMaScore | 95.92 | 97.56 |
The gain is significant under a paired bootstrap over dialogues (p<0.001).
python
from transformers import WhisperForConditionalGeneration, WhisperProcessorfrom peft import PeftModel base = WhisperForConditionalGeneration.from_pretrained("openai/whisper-medium")model = PeftModel.from_pretrained(base, "Superkon3000/whisper-medium-sti-anamnese-lora")processor = WhisperProcessor.from_pretrained("openai/whisper-medium")Merge with model.merge_and_unload(), or convert to CTranslate2 for faster-whisper.
Trained on synthetic and acted speech, not real consultations. Spontaneous clinical speech with overlap, hesitation and background noise is untested. One German clinical domain, 28 evaluation dialogues.
MIT, following the base model.