GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
IonGrozea
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
IonGrozea
Model Tree
openai/whisper-medium
Input Modalities
Output Modalities
Supported Functionality
Evaluated on a held-out validation split of 25,348 samples, beam size 5:
| Metric | Value |
|---|---|
| WER | 6.63 % |
| WER (normalized: lowercase, punctuation stripped) | 6.3 % |
| CER | 2.31 % |
The validation set is a deliberately hard multi-source mix (read speech, parliamentary/spontaneous speech, command-domain utterances), so numbers are not directly comparable to single-corpus benchmarks like Common Voice test.
~185 K training samples drawn from a combined Romanian ASR corpus (~290 K samples / ~619 h total across splits), assembled from:
Audio pre-processed to 80-bin log-mel features (30 s windows, Whisper standard). Splits: train 185,175 / validation 25,348 / test 78,167.
Full fine-tune (no LoRA) on a single RTX 5070 Ti 16 GB, BF16 autocast, SDPA attention.
| Hyperparameter | Value |
|---|---|
| Parameters | 769 M |
| Effective batch size | 64 (4 × 16 grad accum) |
| Learning rate | 1e-5, cosine schedule |
| Warmup steps | 500 |
| Max epochs | 8 (early stopping, patience 5 on eval WER) |
| Optimizer | AdamW 8-bit (bitsandbytes) |
| Label smoothing | 0.0 |
| SpecAugment | 2 freq masks (≤27 bins), 2 time masks (≤100 frames) |
| Selection metric |
Labels use the canonical Whisper target sequence
(<|startoftranscript|><|ro|><|transcribe|><|notimestamps|> … <|endoftext|>)
with teacher-forcing shift, and decoder_start_token_id set to
<|startoftranscript|> so training and generation are exactly aligned.
python
from transformers import pipelineasr = pipeline("automatic-speech-recognition",model="IonGrozea/whisper-medium_ro-80mel",device="cuda",)print(asr("audio.wav", generate_kwargs={"language": "romanian", "task": "transcribe"}))
For low-latency deployment, the model converts cleanly to CTranslate2 / faster-whisper and whisper.cpp (GGUF).
<|ro|>.If you use this model, please also credit the underlying datasets (Common Voice, FLEURS, VoxPopuli, etc.) and OpenAI Whisper:
markdown
Radford et al., "Robust Speech Recognition via Large-Scale Weak Supervision", 2022.
| raw WER, best checkpoint reloaded at end |