GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
opencouncil
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
opencouncil
Model Tree
openai/whisper-large-v3
Input Modalities
Output Modalities
Supported Functionality
Evaluated against a held-out set of human-corrected council utterances. val_corr
is the hard set (utterances a human actually corrected); val_reg is a
regression check on general speech (to confirm the model does not forget everyday
Greek). Lower is better; deltas are vs. the un-adapted whisper-large-v3 baseline
scored the same way.
| Set | n | WER | WER (norm) | CER |
|---|---|---|---|---|
| val_corr (corrections) | 3,157 | 37.74 (−17.9) | 29.35 (−17.6) | 17.46 (−15.5) |
| val_reg (general, regression) | 4,722 | 10.46 (−14.3) | 4.87 (−10.4) | 3.43 (−10.9) |
WER (norm) lowercases, strips accents, folds final sigma, and drops
punctuation. The adapter improves the hard corrections set and general speech
— no regression.
Greek-language automatic speech recognition for municipal-council recordings and
similar formal Greek speech. Designed as a drop-in ASR provider for the
OpenCouncil transcription pipeline (schemalabz/opencouncil-tasks). It performs
transcription only — speaker diarization is done downstream (pyannote), so
this model assigns no speakers.
python
import torchfrom transformers import WhisperForConditionalGeneration, WhisperProcessor, pipelinefrom peft import PeftModelbase = WhisperForConditionalGeneration.from_pretrained("openai/whisper-large-v3", torch_dtype=torch.float16)model = PeftModel.from_pretrained(base, "opencouncil/whisper-large-v3-el-council-lora").merge_and_unload()model.generation_config.language = "greek"model.generation_config.task = "transcribe"proc = WhisperProcessor.from_pretrained("openai/whisper-large-v3", language="greek", task="transcribe")asr = pipeline("automatic-speech-recognition", model=model,tokenizer=proc.tokenizer, feature_extractor=proc.feature_extractor,chunk_length_s=30, torch_dtype=torch.float16, device="cuda")print(asr("council_clip.wav", generate_kwargs={"language": "greek", "task": "transcribe"}))
A harness that emits transcripts in the exact OpenCouncil Transcript JSON
schema (drop-in for the Scribe transcriber) lives in the project repo at
eval/oc_inference_harness.py.
Human-corrected transcripts of Greek municipal-council meetings, combined with an un-edited backbone of the same meetings (~28.6h total). The dataset is private and under a GDPR/DPO legal hold and is not released. Only this model artifact is public.
q_proj/v_proj carries a low but
nonzero risk of memorizing rare tokens (e.g. names) from the training audio.
Do not treat the model as a source of personal data; use it to transcribe, not
to recall training content.Fine-tuned as part of OpenCouncil's GSoC 2026 Greek-ASR work. Base model:
openai/whisper-large-v3 (Apache-2.0). Adapter weights released under Apache-2.0.