GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
GoktugD
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
GoktugD
Model Tree
openai/whisper-small
Input Modalities
Output Modalities
Supported Functionality
Both checkpoints were decoded with the same Turkish transcription prompt, greedy generation, audio-duration filter and Turkish-aware text normalizer.
| Model | Parameters | WER ↓ | CER ↓ |
|---|---|---|---|
openai/whisper-small | 244M | 16.53% | 3.90% |
| DUSUNEN Dinle v1 | 244M | 14.41% | 3.49% |
Relative WER reduction: 12.83%. Raw, machine-readable
results are included as benchmark-baseline.json and benchmark-final.json.
python
import torchfrom transformers import pipelinedevice = 0 if torch.cuda.is_available() else -1transcribe = pipeline("automatic-speech-recognition",model="GoktugD/DUSUNEN-Dinle-244M-v1",device=device,chunk_length_s=30,)print(transcribe("turkce_konusma.wav")["text"])
For deterministic Turkish decoding with the low-level API, set language="tr"
and task="transcribe" in the generation configuration.
openai/whisper-small (Apache-2.0)google/fleurs, Turkish tr_tr configuration (CC BY 4.0)1e-5, linear decay, 10% warmuptrain_dinle_asr.pyDinle is intended for Turkish transcription, subtitle drafts, searchable audio, meeting notes and accessibility prototypes. Validate it on your own accents, microphones and acoustic conditions before production use.
model.safetensors SHA-256: 21cbca7eebae9c5f6f0df4b3d64cc3c414655c263207268041ed41e3b5d979ba
Developed and released by Göktuğ Düşünen.